diff --git a/.dockerignore b/.dockerignore
index 4efb037..bc12e04 100644
--- a/.dockerignore
+++ b/.dockerignore
@@ -33,6 +33,10 @@ tmp/
 output/
 models/
 video/
+migrate/
+*.tar.gz
+*.tar
+*.zip
 *.mp4
 *.wav
 *.docx
@@ -60,3 +64,9 @@ server.log
 .env.local
 *.key
 *.secret
+
+# OpenSpec / Git
+openspec/
+.opencode/
+openspec/changes/
+openspec/specs/
diff --git a/.env.example b/.env.example
new file mode 100644
index 0000000..2d1cd48
--- /dev/null
+++ b/.env.example
@@ -0,0 +1,18 @@
+# WhisperX + Native Python RAG transcription service
+# Скопируйте в .env: `cp .env.example .env` (или `make init`)
+
+# HuggingFace (для pyannote-диаризации)
+# Регистрация: https://huggingface.co/settings/tokens
+HF_TOKEN=hf_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
+
+# OpenCode / DeepSeek API (LLM для классификации и чат-ответов)
+OPENCODE_API_KEY=sk-xxxxxxxxxxxxxxxxxxxxxxxx
+OPENCODE_URL=https://opencode.ai/zen/v1
+
+# JWT-секрет для аутентификации
+# Сгенерируйте: openssl rand -hex 32
+JWT_SECRET=change-me-in-production-32-bytes-random-hex
+
+# Пароль админа по умолчанию (при первом бутстрапе)
+# Смените сразу после первого входа!
+AUTH_ADMIN_PASSWORD=admin123
diff --git a/.gitignore b/.gitignore
index 087ef9a..103a050 100644
--- a/.gitignore
+++ b/.gitignore
@@ -60,8 +60,8 @@ processed/
 .DS_Store
 Thumbs.db
 
-# Secrets (раскомментируйте для публичных репозиториев)
-# .env
+# Secrets (даже в приватном Gitea лучше держать .env вне репо)
+.env
 .env.local
 *.key
 *.secret
@@ -76,3 +76,18 @@ server.log
 
 # Auth database
 data/
+
+# Migration archives (do not commit — see scripts/migrate_lightrag_to_qmd.py)
+migrate/
+*.tar.gz
+*.tar
+
+# OpenSpec local state (changes/specs are committed; archive metadata is fine)
+.opencode/
+
+# Build artifacts
+_docker_out/
+_*.log
+_build*.log
+_smoke_*.py
+_debug*.py
diff --git a/AGENTS.md b/AGENTS.md
index f008426..8c01da7 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -6,3 +6,46 @@
 - **Push to remote**: Once the local commit(s) are ready, push them to the remote repository. Use `git push -u origin main` if the upstream branch is not yet tracked; otherwise use `git push`.
 - **No uncommitted changes left behind**: Before finishing a task, ensure all intended changes are committed and pushed to avoid losing work.
 - **No empty commits**: Avoid creating empty or placeholder commits.
+
+## Native RAG Engine
+
+The project uses a **native Python RAG engine** (no external daemons, no Node.js):
+hybrid BM25 (SQLite FTS5) + vector (sqlite-vec with numpy fallback) + LLM rerank
+through OpenCode.
+
+### Layout
+
+- `src/rag/engine/` — the engine itself:
+  - `db.py` — `Database` (SQLite + sqlite-vec + FTS5 schema, fallback detection).
+  - `chunker.py` — markdown-aware recursive splitter (~900 chars, 15% overlap).
+  - `embeddings.py` — singleton sentence-transformers model (lazy load).
+  - `bm25.py` — FTS5 BM25 with `rank_bm25` fallback.
+  - `vector.py` — sqlite-vec with numpy cosine fallback.
+  - `hybrid.py` — RRF fusion (k=60).
+  - `rerank.py` — LLM rerank through OpenCode.
+  - `engine.py` — public facade: `index_file`, `index_text`, `search`, `vsearch`, `query`, `get`, `status`, `warmup`.
+- `src/rag/qmd/` — compatibility layer preserving the old `qmd_*` API:
+  `qmd_query`, `qmd_chat`, `qmd_chat_stream`, `qmd_index_meeting`, `qmd_index_document`.
+  `main.py` / `queue.py` / `ingest_worker.py` use these.
+- `src/ingest/stub_writer.py` — `.md` stubs for binary files (videos, archives).
+
+### Conventions
+
+- Коллекция = `processed/<org>/qmd_collections/<project_slug>/` (или `_global/`) — внутри лежит `index.sqlite`.
+- Перед изменением `src/rag/engine/` — прочитай `openspec/changes/native-rag-engine/design.md`.
+- При добавлении нового retrieval-режима — обнови `LEGACY_MODE_MAP` в `src/rag/qmd/query.py`.
+- При добавлении нового LLM-вызова — обнови `CHAT_MODES` в `src/rag/qmd/query.py`.
+
+### Tests
+
+- Все новые модули `src/rag/engine/` обязаны иметь unit-тест в `tests/test_native_engine.py`.
+- Реальные данные: 3–5 `.md` файлов в `tempfile.TemporaryDirectory()`.
+- Запуск: `python -m pytest tests/ -q` (46 passed на момент написания).
+- E2E: `tests/test_native_engine_e2e.py` — ingest → search → chat-stream с подменой OpenCode.
+
+### Fallback-стратегии
+
+- FTS5 недоступен → `rank_bm25` in-memory.
+- sqlite-vec недоступен → numpy cosine in-memory.
+- Embedding-модель не загрузилась → BM25-only режим.
+
diff --git a/Dockerfile b/Dockerfile
index dd7d141..bc0dfaf 100644
--- a/Dockerfile
+++ b/Dockerfile
@@ -1,6 +1,6 @@
 FROM python:3.11-slim-bookworm
 
-# Установка системных зависимостей
+# Системные зависимости (без Node.js — native RAG engine работает in-process)
 RUN apt-get update && apt-get install -y --no-install-recommends \
     ffmpeg \
     build-essential \
@@ -11,21 +11,16 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
     tesseract-ocr-eng \
     && rm -rf /var/lib/apt/lists/*
 
-# Рабочая директория
 WORKDIR /app
 
-# Копируем зависимости
+# Python-зависимости (включают native RAG: sentence-transformers, rank_bm25, sqlite-vec)
 COPY requirements.txt pip.conf ./
 RUN pip install --no-cache-dir --timeout 300 -r requirements.txt
 
-# Копируем код проекта
 COPY . .
 
-# Создаём директории для данных
-RUN mkdir -p uploads processed tmp
+RUN mkdir -p uploads processed tmp data
 
-# Открываем порт
 EXPOSE 8000
 
-# Запускаем напрямую (модели скачаются при первом запуске, не при сборке!)
 CMD ["uvicorn", "backend.main:app", "--host", "0.0.0.0", "--port", "8000"]
diff --git a/Dockerfile.rag b/Dockerfile.rag
index 94e4dcf..687bfdc 100644
--- a/Dockerfile.rag
+++ b/Dockerfile.rag
@@ -1,15 +1,26 @@
-# Быстрое обновление: добавляет RAG-зависимости к уже собранному образу с Whisper/PyTorch.
-FROM transcription-transcription:latest
+FROM python:3.11-slim-bookworm
 
-COPY pip.conf /etc/pip.conf
-RUN pip install --no-cache-dir --timeout 300 \
-    lightrag-hku>=1.4.0 \
-    openai>=1.0.0 \
-    python-dotenv>=1.0.0 \
-    sentence-transformers>=3.0.0 \
-    bcrypt>=4.0.0 \
-    "python-jose[cryptography]" \
-    pymupdf>=1.24.0 \
-    openpyxl>=3.1.0 \
-    Pillow>=10.0.0 \
-    pytesseract>=0.3.10
+# Native RAG engine — никаких внешних сервисов не требуется
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    ffmpeg \
+    build-essential \
+    libsndfile1 \
+    curl \
+    ca-certificates \
+    tesseract-ocr \
+    tesseract-ocr-rus \
+    tesseract-ocr-eng \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+
+COPY requirements.txt pip.conf ./
+RUN pip install --no-cache-dir --timeout 300 -r requirements.txt
+
+COPY . .
+
+RUN mkdir -p uploads processed tmp data
+
+EXPOSE 8000
+
+CMD ["uvicorn", "backend.main:app", "--host", "0.0.0.0", "--port", "8000"]
diff --git a/Makefile b/Makefile
new file mode 100644
index 0000000..30572d7
--- /dev/null
+++ b/Makefile
@@ -0,0 +1,50 @@
+SHELL := /bin/bash
+.DEFAULT_GOAL := help
+
+# Detect whether `docker compose` v2 or `docker-compose` v1 is available
+COMPOSE := $(shell command -v docker-compose 2>/dev/null || echo "docker compose")
+
+.PHONY: help init pull-models up down restart logs status clean test deploy
+
+help:        ## Показать эту справку
+	@awk 'BEGIN {FS = ":.*?## "} /^[a-zA-Z_-]+:.*?## / {printf "  \033[36m%-15s\033[0m %s\n", $$1, $$2}' $(MAKEFILE_LIST)
+
+init:        ## Создать .env из .env.example (если его нет)
+	@if [ ! -f .env ]; then \
+		cp .env.example .env && \
+		echo "Created .env from .env.example"; \
+		echo "==> Edit .env with real secrets, then run: make deploy"; \
+	else \
+		echo ".env already exists, skipping"; \
+	fi
+
+pull-models: ## Загрузить модели (rsync + fallback на download_models.py)
+	@bash scripts/pull_models.sh
+
+up:          ## Запустить контейнеры (build + -d)
+	$(COMPOSE) up --build -d
+	@$(COMPOSE) logs --tail=30 transcription || true
+
+down:        ## Остановить контейнеры
+	$(COMPOSE) down
+
+restart:     ## Перезапустить transcription
+	$(COMPOSE) restart transcription
+
+logs:        ## Логи transcription (follow)
+	$(COMPOSE) logs -f transcription
+
+status:      ## Статус контейнеров
+	$(COMPOSE) ps
+	@echo ""
+	@cmd /c "curl -s http://localhost:8000/api/health" 2>/dev/null || \
+	  curl -s http://localhost:8000/api/health || \
+	  echo "(service not responding on :8000)"
+
+clean:       ## Удалить контейнеры + volumes (ОСТОРОЖНО: стирает uploads/processed/data)
+	$(COMPOSE) down -v
+
+test:        ## Прогнать pytest
+	pytest tests/ -q
+
+deploy: pull-models up   ## Полный деплой: модели + запуск
diff --git a/README.md b/README.md
index fce62ed..7c7f8e1 100644
--- a/README.md
+++ b/README.md
@@ -7,6 +7,8 @@
 - **WhisperX** — ASR + alignment + диаризация (всё-в-одном)
 - **python-docx** — генерация `.docx`
 - **PyYAML** — конфигурация
+- **Native Python RAG engine** — гибридный поиск BM25 (FTS5) + vector (sqlite-vec) + LLM-реранкер
+- **OpenCode / DeepSeek** — LLM для классификации и чат-ответов
 
 ## Установка
 
@@ -315,6 +317,98 @@ docker compose down -v
 
 Данные сохраняются между перезапусками контейнера.
 
+## 🧠 База знаний / RAG (Native Python Engine)
+
+База знаний — **in-process Python-движок**: гибридный поиск BM25 (SQLite FTS5) + vector (sqlite-vec с numpy fallback) + LLM-реранкер через OpenCode. Хранение: один `index.sqlite` на коллекцию. **Внешних сервисов не требуется.**
+
+### Архитектура
+
+```
+аудио/видео → WhisperX → extracted.md, summary.md
+документы (PDF/DOCX/XLSX/...) → extracted.md
+бинарники (.mp4, .zip) → stub_writer → *.md с YAML frontmatter
+                                                 ↓
+                              Native Python RAG engine (in-process)
+                              ├─ chunker (markdown-aware, 900 chars, 15% overlap)
+                              ├─ embeddings (sentence-transformers, 384 dim, мультиязычный)
+                              ├─ FTS5 BM25 + sqlite-vec cosine
+                              ├─ RRF fusion (k=60)
+                              └─ LLM rerank (OpenCode/DeepSeek, опционально)
+                                                 ↓
+                                  OpenCode / DeepSeek chat-completions
+                                                 ↓
+                                  WebSocket → rag_context → rag_chunk* → rag_response
+```
+
+### Установка зависимостей
+
+Native engine использует только Python-пакеты (никаких npm/node):
+
+```bash
+pip install -r requirements.txt
+# Скачает ~50 MB модели при первом запуске (sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2)
+```
+
+### Конфигурация
+
+Секция `rag:` в `config.yaml`:
+
+```yaml
+rag:
+  enabled: true
+  auto_index: true
+  qmd_collection_root: ./processed        # корень коллекций
+  qmd_use_rerank: true                     # LLM-реранкер через OpenCode
+  embed_model: sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
+  chat_model: deepseek-v4-flash-free
+  chat_max_tokens: 8192
+```
+
+ENV-переменные: `QMD_COLLECTION_ROOT`, `RAG_EMBED_MODEL`, `OPENCODE_API_KEY`, `OPENCODE_URL`.
+
+### Где хранятся индексы
+
+```
+processed/
+  <org_slug>/
+    qmd_collections/                # ← native engine (in-process, без демонов)
+      <project_slug>/
+        index.sqlite                # FTS5 + sqlite-vec + chunks
+      _global/
+        index.sqlite                # cross-project поиск
+    meetings/<folder>/              # .docx, .md, .json для совещаний
+    documents/<doc_id>/             # .pdf, extracted.md, metadata.json
+    lightrag_caches/                # legacy: для миграции
+```
+
+### Поддерживаемые форматы и stub'ы
+
+Движок индексирует `.md`/`.txt` нативно. Для бинарных форматов
+(видео, не-OCR PDF, архивы) `src/ingest/stub_writer.py` создаёт
+`<filename>.md` со ссылкой на оригинал. Пользователь видит stub в результатах
+поиска и кликает на ссылку — открывается оригинал.
+
+### Fallback-стратегии
+
+- **FTS5 недоступен** в системной сборке Python → `rank_bm25` in-memory.
+- **sqlite-vec недоступен** → numpy cosine in-memory.
+- **Embedding-модель не загрузилась** → BM25-only режим, `qmd: degraded` в healthcheck.
+
+### Legacy-миграция с LightRAG
+
+```bash
+# 1. Снапшот
+tar -czf ../processed-pre-qmd.tar.gz processed
+
+# 2. dry-run
+python scripts/migrate_lightrag_to_qmd.py --org merakom --dry-run
+
+# 3. реальная миграция
+python scripts/migrate_lightrag_to_qmd.py --org merakom
+```
+
+Скрипт идемпотентен: повторный запуск безопасен. См. `scripts/README.md`.
+
 ## 🌐 Веб-интерфейс
 
 Проект включает веб-сервис с минималистичным фронтендом для удобной работы через браузер.
diff --git a/backend/ingest_worker.py b/backend/ingest_worker.py
index 893bea8..b8b615d 100644
--- a/backend/ingest_worker.py
+++ b/backend/ingest_worker.py
@@ -7,12 +7,13 @@ from datetime import datetime
 from pathlib import Path
 from typing import Any, Dict
 
-from backend.paths import org_documents_dir, org_rag_index_dir, write_folder_project_meta
+from backend.paths import org_documents_dir, org_qmd_root, write_folder_project_meta
 from src.config import load_config, resolve_opencode_credentials
 from src.ingest.classify import classify_document
 from src.ingest.formatter import format_global_index_document, format_index_document
-from src.ingest.router import extract_document
-from src.rag.indexer import index_meeting
+from src.ingest.router import extract_document, is_extractable
+from src.ingest.stub_writer import write_stub
+from src.rag.qmd.indexer import qmd_index_document
 
 
 async def process_document_ingest(job: Dict[str, Any], tasks: dict, send_progress):
@@ -31,39 +32,60 @@ async def process_document_ingest(job: Dict[str, Any], tasks: dict, send_progres
         ingest_cfg = config.get("ingest", {})
         pdf_ocr = ingest_cfg.get("pdf_ocr", True)
 
-        doc = await asyncio.to_thread(
-            extract_document,
-            file_path,
-            project_slug,
-            doc_type,
-            None,
-            pdf_ocr,
-        )
-
-        if not doc.full_text.strip():
-            raise ValueError("Не удалось извлечь текст из документа")
-
         documents_dir = org_documents_dir(org_slug)
-        output_dir = documents_dir / doc.document_id
+        output_dir = documents_dir / f"doc_{datetime.now().strftime('%Y%m%d_%H%M%S')}_{task_id[-8:]}"
         await asyncio.to_thread(output_dir.mkdir, parents=True, exist_ok=True)
 
         original_dest = output_dir / file_path.name
         await asyncio.to_thread(shutil.copy2, file_path, original_dest)
-        await asyncio.to_thread(
-            (output_dir / "extracted.md").write_text,
-            doc.full_text,
-            encoding="utf-8",
-        )
         await asyncio.to_thread(write_folder_project_meta, output_dir, project_slug)
 
+        if is_extractable(file_path.name):
+            doc = await asyncio.to_thread(
+                extract_document,
+                file_path,
+                project_slug,
+                doc_type,
+                None,
+                pdf_ocr,
+            )
+            if not doc.full_text.strip():
+                raise ValueError("Не удалось извлечь текст из документа")
+            await asyncio.to_thread(
+                (output_dir / "extracted.md").write_text,
+                doc.full_text,
+                encoding="utf-8",
+            )
+        else:
+            try:
+                stub = await asyncio.to_thread(
+                    write_stub, file_path, project_slug
+                )
+                print(f"[Ingest] {task_id}: создан stub {stub.name}")
+            except (FileNotFoundError, OSError) as exc:
+                print(f"[Ingest] {task_id}: stub_writer failed: {exc}")
+            doc = type("StubDoc", (), {})()
+            doc.full_text = ""
+            doc.document_id = output_dir.name
+            doc.filename = file_path.name
+            doc.doc_type = doc_type
+            doc.metadata = {}
+            doc.to_metadata_dict = lambda: {
+                "document_id": doc.document_id,
+                "filename": doc.filename,
+                "doc_type": doc.doc_type,
+                "project": project_slug,
+                "stub": True,
+            }
+
         tasks[task_id].update({"status": "postprocessing", "message": "Анализ документа...", "progress": 40})
         await send_progress(task_id, 40, "Анализ документа...", "postprocessing")
 
-        metadata = doc.to_metadata_dict()
+        metadata = doc.to_metadata_dict() if callable(getattr(doc, "to_metadata_dict", None)) else {}
         rag_cfg = config.get("rag", {})
         api_key, base_url = resolve_opencode_credentials(config)
 
-        if api_key and ingest_cfg.get("auto_classify", True):
+        if api_key and ingest_cfg.get("auto_classify", True) and doc.full_text:
             metadata = await classify_document(
                 text=doc.full_text,
                 project=project_slug,
@@ -82,34 +104,37 @@ async def process_document_ingest(job: Dict[str, Any], tasks: dict, send_progres
             encoding="utf-8",
         )
 
-        doc_text = format_index_document(doc, metadata)
-        index_path = output_dir / "index.txt"
-        await asyncio.to_thread(index_path.write_text, doc_text, encoding="utf-8")
+        if doc.full_text:
+            doc_text = format_index_document(doc, metadata)
+            index_path = output_dir / "index.txt"
+            await asyncio.to_thread(index_path.write_text, doc_text, encoding="utf-8")
+        else:
+            doc_text = ""
+            index_path = None
 
         result_data = {
-            "document_id": doc.document_id,
+            "document_id": getattr(doc, "document_id", output_dir.name),
             "dir": str(output_dir),
             "rel_dir": str(output_dir.relative_to(documents_dir)),
-            "extracted": str(output_dir / "extracted.md"),
-            "index": str(index_path),
+            "extracted": str(output_dir / "extracted.md") if (output_dir / "extracted.md").exists() else None,
+            "index": str(index_path) if index_path else None,
             "project": project_slug,
             "doc_type": metadata.get("doc_type", doc_type),
             "kind": "document",
         }
 
         if rag_cfg.get("enabled", False) and rag_cfg.get("auto_index", True):
-            tasks[task_id].update({"message": "Индексация в RAG...", "progress": 75})
-            await send_progress(task_id, 75, "Индексация в RAG...", "postprocessing")
-            global_doc_text = format_global_index_document(doc_text, metadata)
-            await index_meeting(
-                doc_text=doc_text,
-                global_doc_text=global_doc_text,
-                project_name=project_slug,
-                working_dir_base=org_rag_index_dir(org_slug),
-                model=rag_cfg.get("index_model", "mimo-v2.5-free"),
-                api_key=api_key,
-                base_url=base_url,
-            )
+            tasks[task_id].update({"message": "Индексация в qmd...", "progress": 75})
+            await send_progress(task_id, 75, "Индексация в qmd...", "postprocessing")
+            try:
+                await qmd_index_document(
+                    org_slug=org_slug,
+                    project_slug=project_slug,
+                    document_dir=output_dir,
+                    extracted_md=output_dir / "extracted.md",
+                )
+            except Exception as idx_exc:
+                print(f"[Ingest] {task_id}: qmd index failed: {idx_exc}")
 
         from backend.queue import _cleanup_upload
         await asyncio.to_thread(_cleanup_upload, file_path)
diff --git a/backend/main.py b/backend/main.py
index bae11a4..98bcac6 100644
--- a/backend/main.py
+++ b/backend/main.py
@@ -16,7 +16,7 @@ from backend.auth.models import UserContext
 from backend.auth.routes import admin_router, router as auth_router
 from backend.auth import database as auth_db
 from backend.auth.service import ensure_project_access, list_accessible_projects
-from backend.paths import org_documents_dir, org_meetings_dir, org_rag_index_dir, resolve_document_path, resolve_meeting_path
+from backend.paths import org_documents_dir, org_meetings_dir, org_qmd_root, org_rag_index_dir, resolve_document_path, resolve_meeting_path
 from backend.queue import (
     delete_folder,
     get_all_tasks,
@@ -35,8 +35,10 @@ from backend.queue import (
 sys.path.insert(0, str(Path(__file__).parent.parent))
 from src.config import load_config, resolve_opencode_credentials
 from src.rag.formatter import format_global_document
-from src.rag.indexer import get_project_names, index_meeting
-from src.rag.query import rag_chat
+from src.rag.qmd import collections as qmd_collections
+from src.rag.qmd.cache import get_search_cache
+from src.rag.qmd.indexer import qmd_index_document, qmd_index_meeting
+from src.rag.qmd.query import map_retrieval_mode, qmd_chat, qmd_chat_stream
 
 STATIC_DIR = Path(__file__).parent / "static"
 
@@ -76,6 +78,7 @@ set_progress_callback(manager.broadcast)
 
 @asynccontextmanager
 async def lifespan(app: FastAPI):
+    from src.rag.qmd.lifespan import qmd_lifespan
     config = load_config()
     auth_db.init_db(config)
     auth_db.bootstrap_from_config(config)
@@ -92,9 +95,11 @@ async def lifespan(app: FastAPI):
         postprocess_workers=postprocess_workers,
         ingest_workers=ingest_workers,
     )
-    yield
+    async with qmd_lifespan(app):
+        yield
     print("🛑 Остановка рабочих процессов...")
     stop_workers()
+    get_search_cache().clear()
 
 
 app = FastAPI(title="Transcription Service", version="2.0.0", lifespan=lifespan)
@@ -111,15 +116,51 @@ app.include_router(auth_router)
 app.include_router(admin_router)
 
 
-def _org_index_dir(user: UserContext) -> Path:
-    return org_rag_index_dir(user.org_slug)
+def _org_qmd_root(user: UserContext) -> Path:
+    return org_qmd_root(user.org_slug)
 
 
 async def _list_rag_project_slugs(user: UserContext) -> List[str]:
-    projects = await get_project_names(_org_index_dir(user))
+    root = _org_qmd_root(user)
+    if not root.exists():
+        return []
+    projects = sorted(
+        p.name for p in root.iterdir()
+        if p.is_dir() and p.name != "_global"
+    )
     return user.filter_projects(projects)
 
 
+def _validate_rag_access(user: UserContext, project_name: Optional[str]) -> None:
+    if project_name:
+        ensure_project_access(user, project_name)
+    elif not user.can_global_search():
+        raise HTTPException(status_code=403, detail="Глобальный поиск доступен только администратору")
+
+
+def _normalize_rag_modes(chat_mode: str, retrieval_mode: str) -> tuple[str, str]:
+    if chat_mode not in ("hybrid", "compare", "timeline"):
+        chat_mode = "hybrid"
+    if retrieval_mode not in ("naive", "local", "global", "hybrid"):
+        retrieval_mode = "hybrid"
+    return chat_mode, retrieval_mode
+
+
+def _rag_params(user: UserContext, chat_mode: str, retrieval_mode: str) -> dict:
+    config = load_config()
+    rag_cfg = config.get("rag", {})
+    api_key, base_url = resolve_opencode_credentials(config)
+    return {
+        "org_slug": user.org_slug,
+        "api_key": api_key,
+        "chat_model": rag_cfg.get("chat_model", "deepseek-v4-flash-free"),
+        "max_tokens": int(rag_cfg.get("chat_max_tokens", 8192)),
+        "mode": retrieval_mode,
+        "chat_mode": chat_mode,
+        "use_rerank": bool(rag_cfg.get("qmd_use_rerank", True)),
+    }
+
+
 async def _rag_chat_for_user(
     user: UserContext,
     question: str,
@@ -128,30 +169,14 @@ async def _rag_chat_for_user(
     chat_mode: str = "hybrid",
     retrieval_mode: str = "hybrid",
 ):
-    if project_name:
-        ensure_project_access(user, project_name)
-    elif not user.can_global_search():
-        raise HTTPException(status_code=403, detail="Глобальный поиск доступен только администратору")
-
-    if chat_mode not in ("hybrid", "compare", "timeline"):
-        chat_mode = "hybrid"
-    if retrieval_mode not in ("naive", "local", "global", "hybrid"):
-        retrieval_mode = "hybrid"
-
-    config = load_config()
-    rag_cfg = config.get("rag", {})
-    api_key, base_url = resolve_opencode_credentials(config)
-    return await rag_chat(
+    _validate_rag_access(user, project_name)
+    chat_mode, retrieval_mode = _normalize_rag_modes(chat_mode, retrieval_mode)
+    params = _rag_params(user, chat_mode, retrieval_mode)
+    return await qmd_chat(
         question=question,
-        working_dir_base=_org_index_dir(user),
         history=history,
-        api_key=api_key,
-        project_name=project_name,
-        base_url=base_url,
-        chat_model=rag_cfg.get("chat_model", "deepseek-v4-flash-free"),
-        mode=retrieval_mode,
-        chat_mode=chat_mode,
-        index_model=rag_cfg.get("index_model", "mimo-v2.5-free"),
+        project_slug=project_name,
+        **params,
     )
 
 
@@ -329,6 +354,18 @@ async def api_download(path: str, user: UserContext = Depends(get_current_user))
 async def api_delete_folder(folder_name: str, user: UserContext = Depends(get_current_user)):
     try:
         delete_folder(user, folder_name)
+        from backend.queue import _folder_project_slug
+        if folder_name.startswith("documents/"):
+            base_dir = org_documents_dir(user.org_slug)
+        else:
+            base_dir = org_meetings_dir(user.org_slug)
+        folder_name_only = folder_name.split("/")[-1]
+        project = _folder_project_slug(folder_name_only, base_dir)
+        if project:
+            try:
+                await qmd_collections.embed(user.org_slug, project)
+            except Exception as idx_exc:
+                print(f"[delete_folder] qmd reindex skipped: {idx_exc}")
         return {"deleted": folder_name}
     except PermissionError as e:
         raise HTTPException(status_code=403, detail=str(e)) from e
@@ -395,13 +432,10 @@ async def api_rag_index_folder(folder_name: str, user: UserContext = Depends(get
         if folder_name.startswith("documents/"):
             folder_path = resolve_document_path(user.org_slug, folder_name[len("documents/"):])
             base_dir = org_documents_dir(user.org_slug)
-            index_files = list(folder_path.glob("index.txt"))
-            txt_files = index_files
         else:
             rel = folder_name[len("meetings/"):] if folder_name.startswith("meetings/") else folder_name
             folder_path = resolve_meeting_path(user.org_slug, rel)
             base_dir = org_meetings_dir(user.org_slug)
-            txt_files = list(folder_path.glob("*.txt"))
 
         if not folder_path.exists():
             return {"error": "Folder not found"}
@@ -411,24 +445,10 @@ async def api_rag_index_folder(folder_name: str, user: UserContext = Depends(get
             return {"error": "Project metadata not found"}
         ensure_project_access(user, project)
 
-        if not txt_files:
-            return {"error": "No index.txt or .txt protocol found in folder"}
-
-        doc_text = txt_files[0].read_text(encoding="utf-8")
-        config = load_config()
-        rag_cfg = config.get("rag", {})
-        api_key, base_url = resolve_opencode_credentials(config)
-        metadata = {"project": project, "section": "Общие вопросы", "topic": "Переиндексация"}
-        global_doc_text = format_global_document(doc_text, metadata)
-
-        await index_meeting(
-            doc_text=doc_text,
-            global_doc_text=global_doc_text,
-            project_name=project,
-            working_dir_base=_org_index_dir(user),
-            model=rag_cfg.get("index_model", "mimo-v2.5-free"),
-            api_key=api_key,
-            base_url=base_url,
+        await qmd_index_document(
+            org_slug=user.org_slug,
+            project_slug=project,
+            document_dir=folder_path,
         )
         return {"indexed": folder_name, "project": project}
     except PermissionError as e:
@@ -442,20 +462,43 @@ async def _handle_rag_query_ws(websocket: WebSocket, msg: dict, user: UserContex
         project = msg.get("project")
         if msg.get("action") == "rag_query_global":
             project = None
-        result = await _rag_chat_for_user(
-            user,
-            msg.get("question", ""),
-            msg.get("history", []),
-            project,
-            chat_mode=msg.get("chat_mode", msg.get("mode", "hybrid")),
-            retrieval_mode=msg.get("retrieval_mode", "hybrid"),
-        )
-        await websocket.send_json({
-            "type": "rag_response",
-            "answer": result["answer"],
-            "context": result["context"],
-            "project": result["project"],
-        })
+
+        chat_mode = msg.get("chat_mode", msg.get("mode", "hybrid"))
+        retrieval_mode = msg.get("retrieval_mode", "hybrid")
+        _validate_rag_access(user, project)
+        chat_mode, retrieval_mode = _normalize_rag_modes(chat_mode, retrieval_mode)
+        params = _rag_params(user, chat_mode, retrieval_mode)
+
+        async for event in qmd_chat_stream(
+            question=msg.get("question", ""),
+            history=msg.get("history", []),
+            project_slug=project,
+            **params,
+        ):
+            event_type = event.get("type")
+            if event_type == "context":
+                await websocket.send_json({
+                    "type": "rag_context",
+                    "context": event["context"],
+                    "project": event["project"],
+                })
+            elif event_type == "chunk":
+                await websocket.send_json({
+                    "type": "rag_chunk",
+                    "content": event["content"],
+                })
+            elif event_type == "done":
+                await websocket.send_json({
+                    "type": "rag_response",
+                    "answer": event["answer"],
+                    "context": event["context"],
+                    "project": event["project"],
+                })
+            elif event_type == "rag_error":
+                await websocket.send_json({
+                    "type": "rag_error",
+                    "error": event.get("error", "unknown error"),
+                })
     except HTTPException as e:
         await websocket.send_json({"type": "rag_error", "error": e.detail})
     except PermissionError as e:
@@ -466,7 +509,11 @@ async def _handle_rag_query_ws(websocket: WebSocket, msg: dict, user: UserContex
 
 @app.get("/api/health")
 async def health():
-    return {"status": "ok"}
+    return {
+        "status": "ok",
+        "qmd": getattr(app.state, "qmd_status", "down"),
+        "search_cache": get_search_cache().stats(),
+    }
 
 
 app.mount("/static", StaticFiles(directory=str(STATIC_DIR)), name="static")
diff --git a/backend/paths.py b/backend/paths.py
index 7fb4c40..7ee5d86 100644
--- a/backend/paths.py
+++ b/backend/paths.py
@@ -1,6 +1,7 @@
 """Org-scoped filesystem paths."""
 
 import json
+import os
 from datetime import datetime
 from pathlib import Path
 
@@ -8,6 +9,7 @@ DATA_ROOT = Path("data")
 UPLOAD_ROOT = Path("uploads")
 PROCESSED_ROOT = Path("processed")
 RAG_CACHE_DIRNAME = "lightrag_caches"
+QMD_COLLECTIONS_DIRNAME = "qmd_collections"
 MEETINGS_DIRNAME = "meetings"
 DOCUMENTS_DIRNAME = "documents"
 
@@ -25,11 +27,21 @@ def org_meetings_dir(org_slug: str) -> Path:
 
 
 def org_rag_index_dir(org_slug: str) -> Path:
+    """Legacy: путь к lightrag_caches/<org>/ (deprecated, kept for migration)."""
     path = PROCESSED_ROOT / org_slug / RAG_CACHE_DIRNAME
     path.mkdir(parents=True, exist_ok=True)
     return path
 
 
+def org_qmd_root(org_slug: str) -> Path:
+    """Корень qmd-коллекций организации: ``processed/<org>/qmd_collections/``."""
+    env_root = os.environ.get("QMD_COLLECTION_ROOT")
+    base = Path(env_root) if env_root else PROCESSED_ROOT
+    path = base / org_slug / QMD_COLLECTIONS_DIRNAME
+    path.mkdir(parents=True, exist_ok=True)
+    return path
+
+
 def org_documents_dir(org_slug: str) -> Path:
     path = PROCESSED_ROOT / org_slug / DOCUMENTS_DIRNAME
     path.mkdir(parents=True, exist_ok=True)
diff --git a/backend/queue.py b/backend/queue.py
index 255dc9d..6b11a92 100644
--- a/backend/queue.py
+++ b/backend/queue.py
@@ -13,7 +13,7 @@ sys.path.insert(0, str(Path(__file__).parent.parent))
 
 from backend.auth.models import UserContext
 from backend.auth.service import ensure_project_access
-from backend.paths import org_documents_dir, org_meetings_dir, org_rag_index_dir, org_upload_dir, resolve_document_path, resolve_meeting_path, write_folder_project_meta
+from backend.paths import org_documents_dir, org_meetings_dir, org_qmd_root, org_upload_dir, resolve_document_path, resolve_meeting_path, write_folder_project_meta
 from src.ingest.router import is_audio_file, is_document_file
 from src.audio_utils import prepare_audio_input
 from src.config import load_config, resolve_opencode_credentials
@@ -25,8 +25,8 @@ from src.rag.formatter import (
     format_meeting_document,
     format_summary_markdown,
 )
-from src.rag.indexer import index_meeting
 from src.rag.parser import classify_meeting, generate_meeting_brief
+from src.rag.qmd.indexer import qmd_index_meeting
 
 tasks: Dict[str, Dict[str, Any]] = {}
 _progress_callback: Optional[Callable] = None
@@ -264,6 +264,8 @@ async def process_postprocessing(job: Dict[str, Any]):
                         model=summary_model,
                         chunk_size=summary_chunk_size,
                     )
+                    if not (brief or "").strip():
+                        print(f"[Postprocess Warning] {task_id}: пустое краткое содержание summary")
                     summary_md = format_summary_markdown(metadata, brief, display_name)
                     await asyncio.to_thread(summary_path.write_text, summary_md, encoding="utf-8")
                     result_data["summary"] = str(summary_path)
@@ -276,17 +278,19 @@ async def process_postprocessing(job: Dict[str, Any]):
                 result_data["rel_dir"] = str(output_dir.relative_to(meetings_dir))
 
                 if rag_cfg.get("enabled", False) and rag_cfg.get("auto_index", True):
-                    await _send_progress(task_id, 92, "Индексация в базу знаний...", "postprocessing")
-                    index_dir = org_rag_index_dir(org_slug)
-                    global_doc_text = format_global_document(doc_text, metadata)
-                    await index_meeting(
-                        doc_text=doc_text,
-                        global_doc_text=global_doc_text,
-                        project_name=project,
-                        working_dir_base=index_dir,
-                        model=index_model,
-                        api_key=api_key,
-                        base_url=base_url,
+                    await _send_progress(task_id, 92, "Индексация в qmd...", "postprocessing")
+                    qmd_root = org_qmd_root(org_slug)
+                    context_description = (
+                        f"Протоколы совещаний и заметки проекта '{project}', "
+                        f"раздел: {metadata.get('section', 'Общие вопросы')}"
+                    )
+                    await qmd_index_meeting(
+                        org_slug=org_slug,
+                        project_slug=project,
+                        body_path=txt_path,
+                        summary_path=summary_path,
+                        txt_path=txt_path,
+                        context_description=context_description,
                     )
             except Exception as e:
                 print(f"[Postprocess Warning] {task_id}: {e}")
diff --git a/backend/static/app.js b/backend/static/app.js
index b116b40..85b3dea 100644
--- a/backend/static/app.js
+++ b/backend/static/app.js
@@ -12,6 +12,8 @@ class TranscriptionApp {
         this.chatProjects = [];
         this.uploadProjects = [];
         this.isChatThinking = false;
+        this.streamingBubble = null;
+        this.streamingText = '';
         this.user = null;
         this.init();
     }
@@ -210,6 +212,10 @@ class TranscriptionApp {
             this.updateQueueSummary(data.queue);
         } else if (data.type === 'file_tree') {
             this.renderFileTree(data.tree);
+        } else if (data.type === 'rag_context') {
+            this.handleChatContext(data);
+        } else if (data.type === 'rag_chunk') {
+            this.handleChatChunk(data);
         } else if (data.type === 'rag_response') {
             this.handleChatResponse(data);
         } else if (data.type === 'rag_error') {
@@ -686,6 +692,8 @@ class TranscriptionApp {
         }
 
         input.value = '';
+        this.streamingBubble = null;
+        this.streamingText = '';
         this.addChatBubble('user', question);
         this.setChatThinking(true);
 
@@ -700,14 +708,58 @@ class TranscriptionApp {
         });
     }
 
+    handleChatContext(data) {
+        const hasContext = data.context && data.context.trim().length > 0;
+        const label = hasContext ? 'Формирую ответ...' : 'Контекст не найден, отвечаю...';
+        this.setChatThinking(true, label);
+    }
+
+    handleChatChunk(data) {
+        const chunk = data.content || '';
+        if (!chunk) return;
+
+        if (!this.streamingBubble) {
+            this.setChatThinking(false);
+            const container = document.getElementById('chatMessages');
+            const bubble = document.createElement('div');
+            bubble.className = 'chat-bubble bot streaming';
+            const body = document.createElement('div');
+            body.className = 'bubble-body md-content';
+            bubble.appendChild(body);
+            container.appendChild(bubble);
+            container.scrollTop = container.scrollHeight;
+            this.streamingBubble = bubble;
+            this.streamingText = '';
+        }
+
+        this.streamingText += chunk;
+        const body = this.streamingBubble.querySelector('.bubble-body');
+        if (body) {
+            body.textContent = this.streamingText;
+        }
+        const container = document.getElementById('chatMessages');
+        container.scrollTop = container.scrollHeight;
+    }
+
     handleChatResponse(data) {
         this.setChatThinking(false);
         const answer = data.answer || 'Нет ответа';
         const project = data.project ? `Проект: ${data.project}` : 'Все проекты';
         const html = `<div class="bubble-body md-content">${this.renderMarkdown(answer)}</div>`;
-        this.addChatBubble('bot', html, { isHtml: true, meta: project });
 
-        // Сохраняем в историю
+        if (this.streamingBubble) {
+            this.streamingBubble.innerHTML = html;
+            this.streamingBubble.classList.remove('streaming');
+            const meta = document.createElement('div');
+            meta.className = 'bubble-sources';
+            meta.textContent = project;
+            this.streamingBubble.appendChild(meta);
+            this.streamingBubble = null;
+            this.streamingText = '';
+        } else {
+            this.addChatBubble('bot', html, { isHtml: true, meta: project });
+        }
+
         const lastUserMsg = this.getLastUserMessage();
         if (lastUserMsg) {
             this.chatHistory.push({ question: lastUserMsg, answer: answer });
@@ -717,6 +769,8 @@ class TranscriptionApp {
 
     handleChatError(error) {
         this.setChatThinking(false);
+        this.streamingBubble = null;
+        this.streamingText = '';
         this.addChatBubble('bot', `Ошибка: ${this.escapeHtml(error || 'Неизвестная ошибка')}`);
     }
 
@@ -729,15 +783,17 @@ class TranscriptionApp {
         return null;
     }
 
-    setChatThinking(thinking) {
+    setChatThinking(thinking, message = 'Думаю...') {
         this.isChatThinking = thinking;
         const container = document.getElementById('chatMessages');
         const existing = container.querySelector('.chat-thinking');
         if (thinking) {
-            if (!existing) {
+            if (existing) {
+                existing.textContent = message;
+            } else {
                 const div = document.createElement('div');
                 div.className = 'chat-thinking';
-                div.textContent = 'Думаю...';
+                div.textContent = message;
                 container.appendChild(div);
                 container.scrollTop = container.scrollHeight;
             }
diff --git a/config.yaml b/config.yaml
index 43b3a7a..cf25ec7 100644
--- a/config.yaml
+++ b/config.yaml
@@ -42,7 +42,8 @@ output:
   speaker_label_style: name  # name | id | none
   paragraph_pause_sec: 2.0   # новый абзац, если пауза > N секунд
 
-# Настройки RAG (LightRAG для протоколов совещаний)
+# Настройки RAG (native Python engine: BM25/FTS5 + sqlite-vec + LLM rerank)
+# Движок работает in-process; внешние сервисы не требуются.
 rag:
   enabled: true
   auto_index: true
@@ -54,17 +55,26 @@ rag:
     - Общие вопросы
     - Согласование
     - Контроль качества
-  index_backend: opencode
-  index_model: mimo-v2.5-free
+
+  # Native RAG engine (in-process, без демонов)
+  qmd_collection_root: ./processed        # ENV override: QMD_COLLECTION_ROOT
+  qmd_use_rerank: true                     # LLM-реранкер (2-3 сек warm)
+  embed_model: sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2  # ENV: RAG_EMBED_MODEL
+
+  # OpenCode LLM (чат-генерация ответов поверх найденного контекста)
   chat_backend: opencode
   chat_model: deepseek-v4-flash-free
+  chat_max_tokens: 8192
+  opencode_api_key: "sk-4jJBUMS7WJyBOtZZAexsSy6aT4NKOYp2gA19WLlaux8jHMw0HvyCl1V45Jf8SONz"  # Или через env: OPENCODE_API_KEY
+  opencode_url: "https://opencode.ai/zen/v1"
+
+  # Классификация и summary (metadata layer, поверх движка)
   summary_auto: true
   summary_model: deepseek-v4-flash-free
   summary_chunk_size: 10000
   classify_chunk_size: 7000
-  opencode_api_key: "sk-4jJBUMS7WJyBOtZZAexsSy6aT4NKOYp2gA19WLlaux8jHMw0HvyCl1V45Jf8SONz"  # Или через env: OPENCODE_API_KEY
-  opencode_url: "https://opencode.ai/zen/v1"
-  project_index_dir: ./processed/lightrag_caches
+  classify_model: mimo-v2.5-free
+  index_model: mimo-v2.5-free  # legacy alias, используется в legacy-классификации
 
 # Очередь обработки
 queue:
diff --git a/deploy/install_docker.sh b/deploy/install_docker.sh
new file mode 100644
index 0000000..e86fe7f
--- /dev/null
+++ b/deploy/install_docker.sh
@@ -0,0 +1,34 @@
+#!/bin/bash
+set -euo pipefail
+
+SUDO_PASS="${1:-}"
+
+if [ -z "$SUDO_PASS" ]; then
+  echo "Usage: $0 <sudo_password>"
+  exit 1
+fi
+
+run_sudo() {
+  echo "$SUDO_PASS" | sudo -S "$@"
+}
+
+export DEBIAN_FRONTEND=noninteractive
+
+run_sudo apt-get update -y
+run_sudo apt-get install -y ca-certificates curl gnupg lsb-release git
+
+if ! command -v docker >/dev/null 2>&1; then
+  curl -fsSL https://get.docker.com -o /tmp/get-docker.sh
+  run_sudo sh /tmp/get-docker.sh
+  rm -f /tmp/get-docker.sh
+fi
+
+run_sudo usermod -aG docker art || true
+
+if ! docker compose version >/dev/null 2>&1; then
+  run_sudo apt-get install -y docker-compose-plugin || true
+fi
+
+docker --version
+docker compose version || docker-compose --version || true
+echo "Docker installation complete"
diff --git a/deploy/setup_server.sh b/deploy/setup_server.sh
new file mode 100644
index 0000000..743de22
--- /dev/null
+++ b/deploy/setup_server.sh
@@ -0,0 +1,51 @@
+#!/bin/bash
+set -euo pipefail
+
+APP_DIR="/home/art/transcription"
+MIGRATE_DIR="$APP_DIR/migrate"
+SUDO_PASS="${1:-}"
+
+cd "$APP_DIR"
+
+mkdir -p data uploads processed models tmp
+
+if [ -f "$MIGRATE_DIR/data.tar.gz" ]; then
+  tar xzf "$MIGRATE_DIR/data.tar.gz" -C data
+fi
+
+if [ -f "$MIGRATE_DIR/processed.tar.gz" ]; then
+  tar xzf "$MIGRATE_DIR/processed.tar.gz" -C processed
+fi
+
+if [ -f "$MIGRATE_DIR/uploads.tar.gz" ]; then
+  echo "Extracting uploads (may take several minutes)..."
+  tar xzf "$MIGRATE_DIR/uploads.tar.gz" -C uploads
+fi
+
+if [ -f "$MIGRATE_DIR/models_huggingface.tar.gz" ]; then
+  echo "Extracting models (may take several minutes)..."
+  mkdir -p models/huggingface
+  tar xzf "$MIGRATE_DIR/models_huggingface.tar.gz" -C models/huggingface
+fi
+
+if [ -f "$MIGRATE_DIR/nltk_data.tar.gz" ]; then
+  mkdir -p models/nltk_data
+  tar xzf "$MIGRATE_DIR/nltk_data.tar.gz" -C models/nltk_data
+fi
+
+docker_cmd() {
+  if docker info >/dev/null 2>&1; then
+    docker "$@"
+  else
+    echo "$SUDO_PASS" | sudo -S docker "$@"
+  fi
+}
+
+echo "Building base image..."
+docker_cmd build -f Dockerfile -t transcription-transcription:latest .
+
+echo "Starting service..."
+docker_cmd compose -f docker-compose.yml -f docker-compose.prod.yml up -d --build
+
+sleep 15
+curl -sf http://localhost:8000/api/health && echo " HEALTH OK" || echo " HEALTH CHECK FAILED (service may still be starting)"
diff --git a/docker-compose.prod.yml b/docker-compose.prod.yml
new file mode 100644
index 0000000..ec14254
--- /dev/null
+++ b/docker-compose.prod.yml
@@ -0,0 +1,16 @@
+# Production overrides for Linux server deployment (bind-mount data dirs)
+services:
+  transcription:
+    volumes:
+      - ./uploads:/app/uploads
+      - ./processed:/app/processed
+      - ./tmp:/app/tmp
+      - ./config.yaml:/app/config.yaml:ro
+      - ./backend:/app/backend:ro
+      - ./src:/app/src:ro
+      - ./scripts:/app/scripts:ro
+      - ./models/huggingface:/root/.cache/huggingface
+      - ./models/nltk_data:/root/nltk_data:ro
+      - ./data:/app/data
+    healthcheck:
+      test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8000/api/health')"]
diff --git a/docker-compose.yml b/docker-compose.yml
index 0fce2dc..ae980d1 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -14,6 +14,8 @@ services:
       - HF_TOKEN=${HF_TOKEN}
       - OPENCODE_API_KEY=${OPENCODE_API_KEY}
       - OPENCODE_URL=${OPENCODE_URL:-https://opencode.ai/zen/v1}
+      - QMD_COLLECTION_ROOT=${QMD_COLLECTION_ROOT:-/app/processed}
+      - RAG_EMBED_MODEL=${RAG_EMBED_MODEL:-sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2}
       - HF_HOME=/root/.cache/huggingface
       - NLTK_DATA=/root/nltk_data
       - JWT_SECRET=${JWT_SECRET:-change-me-in-production}
diff --git a/requirements.txt b/requirements.txt
index 655eb71..313ad8d 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -10,12 +10,14 @@ whisperx
 bcrypt>=4.0.0
 python-jose[cryptography]
 
-# RAG / LightRAG
-lightrag-hku>=1.4.0
+# RAG (native Python engine: BM25/FTS5 + sqlite-vec + LLM rerank)
 openai>=1.0.0
 python-dotenv>=1.0.0
+httpx>=0.27.0
 sentence-transformers>=3.0.0
 numpy>=1.24.0
+rank_bm25>=0.2.2
+sqlite-vec>=0.1.0
 
 # Document ingestion
 pymupdf>=1.24.0
diff --git a/scripts/README.md b/scripts/README.md
new file mode 100644
index 0000000..44c99f8
--- /dev/null
+++ b/scripts/README.md
@@ -0,0 +1,31 @@
+# scripts/
+
+Утилиты для миграции и обслуживания.
+
+## migrate_lightrag_to_qmd.py
+
+Переиндексирует существующие ``.md`` (протоколы совещаний, извлечённые из
+документов) в qmd-коллекции. Используется при переезде с LightRAG.
+
+```bash
+# dry-run: посмотреть, что будет сделано
+python scripts/migrate_lightrag_to_qmd.py --org merakom --dry-run
+
+# реальная миграция
+python scripts/migrate_lightrag_to_qmd.py --org merakom
+```
+
+**Снапшот перед запуском в проде (обязательно):**
+
+```bash
+cd transcription
+tar -czf ../processed-pre-qmd.tar.gz processed
+```
+
+Скрипт идемпотентен: повторный запуск не дублирует данные (qmd content-hash
++ маркер ``.migrated_to_qmd`` в каждой обработанной папке).
+
+## Другие скрипты
+
+Существующие утилиты (`backfill_summaries.py`, `download_models.py` и т.д.)
+работают без изменений.
diff --git a/scripts/backfill_summaries.py b/scripts/backfill_summaries.py
new file mode 100644
index 0000000..4e3bb3f
--- /dev/null
+++ b/scripts/backfill_summaries.py
@@ -0,0 +1,121 @@
+"""Перегенерация пустых или отсутствующих summary для уже обработанных совещаний."""
+
+import argparse
+import asyncio
+import json
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+from backend.paths import PROCESSED_ROOT, MEETINGS_DIRNAME
+from src.config import load_config, resolve_opencode_credentials
+from src.rag.formatter import build_meeting_text_only, format_summary_markdown
+from src.rag.parser import classify_meeting, generate_meeting_brief
+
+
+def _summary_body(summary_path: Path) -> str:
+    if not summary_path.exists():
+        return ""
+    text = summary_path.read_text(encoding="utf-8")
+    if "## Суть" not in text:
+        return ""
+    return text.split("## Суть", 1)[-1].strip()
+
+
+def _find_meeting_jobs(org_slug: str) -> list[dict]:
+    meetings_dir = PROCESSED_ROOT / org_slug / MEETINGS_DIRNAME
+    if not meetings_dir.exists():
+        return []
+
+    jobs = []
+    for folder in sorted(meetings_dir.iterdir()):
+        if not folder.is_dir():
+            continue
+        segments_files = list(folder.glob("*_segments.json"))
+        if not segments_files:
+            continue
+        segments_path = segments_files[0]
+        stem = segments_path.name[: -len("_segments.json")]
+        summary_path = folder / f"{stem}_summary.md"
+        jobs.append({
+            "folder": folder,
+            "stem": stem,
+            "segments_path": segments_path,
+            "summary_path": summary_path,
+            "display_name": f"{stem}.webm",
+        })
+    return jobs
+
+
+async def _regenerate_job(job: dict, config: dict, dry_run: bool) -> bool:
+    body = _summary_body(job["summary_path"])
+    if body:
+        return False
+
+    segments = json.loads(job["segments_path"].read_text(encoding="utf-8"))
+    meeting_text = build_meeting_text_only(segments)
+    rag_cfg = config.get("rag", {})
+    api_key, base_url = resolve_opencode_credentials(config)
+
+    meta_path = job["folder"] / ".project.json"
+    project = "unknown"
+    if meta_path.exists():
+        try:
+            project = json.loads(meta_path.read_text(encoding="utf-8")).get("project_slug", project)
+        except Exception:
+            pass
+
+    if not api_key:
+        print(f"[skip] {job['folder'].name}: нет API-ключа")
+        return False
+
+    if dry_run:
+        print(f"[dry-run] {job['folder'].name}")
+        return True
+
+    sections = rag_cfg.get("sections", ["Общие вопросы"])
+    metadata = await classify_meeting(
+        text=meeting_text,
+        project=project,
+        sections=sections,
+        api_key=api_key,
+        base_url=base_url,
+        model=rag_cfg.get("index_model", "mimo-v2.5-free"),
+        chunk_size=int(rag_cfg.get("classify_chunk_size", 7000)),
+    )
+    brief = await generate_meeting_brief(
+        text=meeting_text,
+        metadata=metadata,
+        api_key=api_key,
+        base_url=base_url,
+        model=rag_cfg.get("summary_model", "deepseek-v4-flash-free"),
+        chunk_size=int(rag_cfg.get("summary_chunk_size", 10000)),
+    )
+    if not (brief or "").strip():
+        print(f"[warn] {job['folder'].name}: LLM вернул пустой brief")
+        return False
+
+    summary_md = format_summary_markdown(metadata, brief, job["display_name"])
+    job["summary_path"].write_text(summary_md, encoding="utf-8")
+    print(f"[ok] {job['folder'].name} ({len(brief)} символов)")
+    return True
+
+
+async def main():
+    parser = argparse.ArgumentParser(description="Backfill пустых summary совещаний")
+    parser.add_argument("--org", default="merakom", help="org_slug")
+    parser.add_argument("--dry-run", action="store_true", help="Только показать, что будет обновлено")
+    args = parser.parse_args()
+
+    config = load_config()
+    jobs = _find_meeting_jobs(args.org)
+    updated = 0
+    for job in jobs:
+        if await _regenerate_job(job, config, args.dry_run):
+            updated += 1
+    print(f"Готово: {updated} из {len(jobs)} совещаний")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/scripts/deploy.sh b/scripts/deploy.sh
new file mode 100644
index 0000000..0f5d36a
--- /dev/null
+++ b/scripts/deploy.sh
@@ -0,0 +1,67 @@
+#!/usr/bin/env bash
+# scripts/deploy.sh — деплой кода на удалённый сервер через rsync + ssh.
+#
+# Использование:
+#   ./scripts/deploy.sh user@server /opt/transcription
+#   SERVER=user@server REMOTE_DIR=/opt/transcription ./scripts/deploy.sh
+#
+# Что делает:
+#   1) rsync кода (исключая models, processed, uploads, data, tmp, .git, migrate)
+#   2) Копирует .env (на текущем этапе — целиком с ключами; позже — Vault)
+#   3) По SSH запускает на сервере: make pull-models && make up
+#
+# Переменные окружения:
+#   SERVER          — user@host (обязательно)
+#   REMOTE_DIR      — путь на сервере (по умолчанию /opt/transcription)
+#   WITH_ENV        — если "1", копирует локальный .env на сервер (по умолчанию 1)
+#   MODELS_SOURCE   — путь к моделям (если пусто — на сервере попробует fallback)
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PROJECT_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
+
+SERVER="${1:-${SERVER:-}}"
+REMOTE_DIR="${2:-${REMOTE_DIR:-/opt/transcription}}"
+WITH_ENV="${WITH_ENV:-1}"
+
+if [ -z "$SERVER" ]; then
+    echo "Usage: $0 user@server [/opt/transcription]" >&2
+    echo "Or:   SERVER=user@server REMOTE_DIR=/path $0" >&2
+    exit 1
+fi
+
+echo "[deploy] Target: $SERVER:$REMOTE_DIR"
+
+# 1. rsync кода
+echo "[deploy] Syncing code..."
+rsync -avz --progress \
+    --exclude='.git' \
+    --exclude='models/' \
+    --exclude='processed/' \
+    --exclude='uploads/' \
+    --exclude='tmp/' \
+    --exclude='data/' \
+    --exclude='migrate/' \
+    --exclude='__pycache__/' \
+    --exclude='.pytest_cache/' \
+    --exclude='.env' \
+    --exclude='*.log' \
+    "$PROJECT_ROOT/" "$SERVER:$REMOTE_DIR/"
+
+# 2. .env (если включено)
+if [ "$WITH_ENV" = "1" ] && [ -f "$PROJECT_ROOT/.env" ]; then
+    echo "[deploy] Copying .env (with secrets)..."
+    scp "$PROJECT_ROOT/.env" "$SERVER:$REMOTE_DIR/.env"
+else
+    echo "[deploy] .env not copied (WITH_ENV=$WITH_ENV). Run 'make init' on server."
+fi
+
+# 3. Remote deploy
+echo "[deploy] Triggering remote deploy..."
+ssh "$SERVER" "cd $REMOTE_DIR && \
+    export MODELS_SOURCE='${MODELS_SOURCE:-}' && \
+    make pull-models && \
+    make up"
+
+echo "[deploy] Done. Check logs: ssh $SERVER 'cd $REMOTE_DIR && make logs'"
diff --git a/scripts/migrate_lightrag_to_qmd.py b/scripts/migrate_lightrag_to_qmd.py
new file mode 100644
index 0000000..ed96f1a
--- /dev/null
+++ b/scripts/migrate_lightrag_to_qmd.py
@@ -0,0 +1,182 @@
+"""migrate_lightrag_to_qmd.py — перенос индексов из LightRAG в qmd.
+
+Идемпотентный скрипт: переиндексирует существующие .md в qmd-коллекции.
+Повторный запуск безопасен (qmd content-hash проверка + перезапись stub).
+
+Использование::
+
+    python scripts/migrate_lightrag_to_qmd.py [--org merakom] [--dry-run]
+
+Что делает:
+1. Находит все ``processed/<org>/meetings/<folder>/<stem>.md`` и ``<stem>_summary.md``.
+2. Находит все ``processed/<org>/documents/<doc_id>/extracted.md``.
+3. По ``.project.json`` определяет проект и вызывает
+   ``src.rag.qmd.indexer.qmd_index_meeting`` или ``qmd_index_document``.
+4. После успешной миграции помечает папку через ``.migrated_to_qmd``.
+
+Снапшот::
+
+    Перед запуском в проде: ``tar -czf processed-pre-qmd.tar.gz processed``
+"""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import json
+import sys
+from pathlib import Path
+from typing import Iterable, Optional
+
+
+ROOT = Path(__file__).resolve().parent.parent
+sys.path.insert(0, str(ROOT))
+
+from src.config import load_config  # noqa: E402
+from src.rag.qmd.indexer import qmd_index_document, qmd_index_meeting  # noqa: E402
+
+MIGRATION_MARKER = ".migrated_to_qmd"
+
+
+def _iter_meetings(org_dir: Path) -> Iterable[Path]:
+    meetings = org_dir / "meetings"
+    if not meetings.exists():
+        return
+    for folder in sorted(meetings.iterdir()):
+        if not folder.is_dir():
+            continue
+        if (folder / MIGRATION_MARKER).exists():
+            continue
+        yield folder
+
+
+def _iter_documents(org_dir: Path) -> Iterable[Path]:
+    documents = org_dir / "documents"
+    if not documents.exists():
+        return
+    for folder in sorted(documents.iterdir()):
+        if not folder.is_dir():
+            continue
+        if (folder / MIGRATION_MARKER).exists():
+            continue
+        yield folder
+
+
+def _read_project(folder: Path, fallback: Path) -> Optional[str]:
+    meta_path = folder / ".project.json"
+    if meta_path.exists():
+        try:
+            return json.loads(meta_path.read_text(encoding="utf-8")).get("project_slug")
+        except json.JSONDecodeError:
+            pass
+    return None
+
+
+def _write_marker(folder: Path) -> None:
+    (folder / MIGRATION_MARKER).write_text(
+        json.dumps(
+            {"migrated_at": _now_iso(), "engine": "qmd"},
+            ensure_ascii=False,
+        ),
+        encoding="utf-8",
+    )
+
+
+def _now_iso() -> str:
+    from datetime import datetime
+    return datetime.now().isoformat()
+
+
+async def migrate_org(org_slug: str, dry_run: bool = False) -> dict:
+    config = load_config()
+    rag_cfg = config.get("rag", {})
+    if not (rag_cfg.get("enabled", True) and rag_cfg.get("auto_index", True)):
+        print(f"[migrate] RAG disabled in config — nothing to do for org='{org_slug}'")
+        return {"meetings": 0, "documents": 0}
+
+    org_dir = ROOT / "processed" / org_slug
+    if not org_dir.exists():
+        print(f"[migrate] No processed/{org_slug} — skipping")
+        return {"meetings": 0, "documents": 0}
+
+    meetings_count = 0
+    documents_count = 0
+    errors: list[str] = []
+
+    for folder in _iter_meetings(org_dir):
+        project = _read_project(folder, org_dir / "meetings")
+        if not project:
+            print(f"[migrate] skip meeting folder {folder.name}: no project")
+            continue
+        body_path = next(iter(sorted(folder.glob("*.txt"))), None)
+        if body_path is None:
+            md_files = sorted(folder.glob("*.md"))
+            body_path = md_files[0] if md_files else None
+        if body_path is None:
+            print(f"[migrate] skip meeting folder {folder.name}: no .md/.txt")
+            continue
+        summary_path = next(iter(sorted(folder.glob("*_summary.md"))), None)
+        print(f"[migrate] meeting: org={org_slug} project={project} folder={folder.name}")
+        if dry_run:
+            continue
+        try:
+            await qmd_index_meeting(
+                org_slug=org_slug,
+                project_slug=project,
+                body_path=body_path,
+                summary_path=summary_path,
+                txt_path=body_path,
+            )
+            _write_marker(folder)
+            meetings_count += 1
+        except Exception as exc:
+            errors.append(f"meeting {folder.name}: {exc}")
+
+    for folder in _iter_documents(org_dir):
+        project = _read_project(folder, org_dir / "documents")
+        if not project:
+            print(f"[migrate] skip document folder {folder.name}: no project")
+            continue
+        extracted = folder / "extracted.md"
+        print(f"[migrate] document: org={org_slug} project={project} folder={folder.name}")
+        if dry_run:
+            continue
+        try:
+            await qmd_index_document(
+                org_slug=org_slug,
+                project_slug=project,
+                document_dir=folder,
+                extracted_md=extracted,
+            )
+            _write_marker(folder)
+            documents_count += 1
+        except Exception as exc:
+            errors.append(f"document {folder.name}: {exc}")
+
+    if errors:
+        print(f"[migrate] {len(errors)} errors:")
+        for err in errors:
+            print(f"  - {err}")
+
+    return {"meetings": meetings_count, "documents": documents_count, "errors": len(errors)}
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description="Migrate LightRAG caches to qmd collections.")
+    parser.add_argument("--org", help="Org slug (default: bootstrap org from config)")
+    parser.add_argument("--dry-run", action="store_true", help="Show what would be done")
+    args = parser.parse_args()
+
+    org_slug = args.org
+    if not org_slug:
+        config = load_config()
+        org_slug = config.get("auth", {}).get("bootstrap", {}).get("org_slug", "merakom")
+
+    print(f"[migrate] target org: {org_slug}  dry-run: {args.dry_run}")
+    result = asyncio.run(migrate_org(org_slug, dry_run=args.dry_run))
+    print(f"[migrate] done: {result}")
+    return 0 if not result.get("errors") else 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/scripts/pull_models.sh b/scripts/pull_models.sh
new file mode 100644
index 0000000..91af614
--- /dev/null
+++ b/scripts/pull_models.sh
@@ -0,0 +1,55 @@
+#!/usr/bin/env bash
+# scripts/pull_models.sh — загрузка моделей (rsync из источника + fallback на download).
+#
+# Использование:
+#   MODELS_SOURCE=user@host:/opt/transcription/models/ bash scripts/pull_models.sh
+#   bash scripts/pull_models.sh            # если MODELS_SOURCE не задан — попробует скачать из интернета
+#
+# Переменные окружения:
+#   MODELS_SOURCE   — rsync-источник (user@host:path) — приоритетный способ
+#   MODELS_DIR      — локальная папка (по умолчанию ./models)
+#   FORCE_DOWNLOAD  — если "1", игнорирует кэш и качает заново
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+PROJECT_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
+MODELS_DIR="${MODELS_DIR:-$PROJECT_ROOT/models}"
+MARKER="$MODELS_DIR/huggingface/hub/models--Systran--faster-whisper-large-v3"
+
+# Если модели уже есть — выходим
+if [ -d "$MARKER" ] && [ "${FORCE_DOWNLOAD:-0}" != "1" ]; then
+    echo "[pull_models] Models already present at $MODELS_DIR — skipping"
+    exit 0
+fi
+
+# 1. Приоритет: rsync
+if [ -n "${MODELS_SOURCE:-}" ]; then
+    echo "[pull_models] Pulling models from $MODELS_SOURCE ..."
+    mkdir -p "$MODELS_DIR/huggingface"
+    rsync -avz --progress "$MODELS_SOURCE/huggingface/" "$MODELS_DIR/huggingface/"
+    rsync -avz --progress "$MODELS_SOURCE/nltk_data/" "$MODELS_DIR/nltk_data/" 2>/dev/null || true
+    echo "[pull_models] Done."
+    exit 0
+fi
+
+# 2. Fallback: скачать из интернета (если rsync не задан)
+echo "[pull_models] MODELS_SOURCE not set, falling back to download_models.py ..."
+cd "$PROJECT_ROOT"
+
+# Проверяем, что python доступен
+if ! command -v python >/dev/null 2>&1; then
+    echo "[pull_models] ERROR: python not found in PATH" >&2
+    exit 1
+fi
+
+# Устанавливаем минимальный набор для download
+pip install --quiet --no-cache-dir openai-whisper 2>/dev/null || true
+pip install --quiet --no-cache-dir "whisperx @ git+https://github.com/m-bain/whisperx.git" 2>/dev/null || true
+
+python scripts/download_models.py || {
+    echo "[pull_models] ERROR: download_models.py failed." >&2
+    echo "Hint: set MODELS_SOURCE=user@host:/opt/transcription/models/ and retry." >&2
+    exit 1
+}
+echo "[pull_models] Downloaded to $MODELS_DIR"
diff --git a/scripts/smoke_rag.py b/scripts/smoke_rag.py
new file mode 100644
index 0000000..0820ffe
--- /dev/null
+++ b/scripts/smoke_rag.py
@@ -0,0 +1,67 @@
+"""Ручной smoke-test движка через CLI: ingest → query → exit."""
+import asyncio
+import sys
+import tempfile
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parent.parent
+sys.path.insert(0, str(ROOT))
+
+from src.rag.engine import get_or_create_engine  # noqa: E402
+
+
+SAMPLES = {
+    "plan.md": (
+        "# План 3-го этажа\n\n"
+        "План 3-го этажа жилого дома. Оси: А, Б, В, Г. Размеры между осями А и Б: 5400 мм.\n"
+        "Квартиры: 301, 302, 303.\n"
+    ),
+    "auth.md": (
+        "# Авторизация\n\n"
+        "Авторизация работает через JWT-токены с TTL 24 часа.\n"
+        "Refresh-токен живёт 30 дней.\n"
+    ),
+    "schedule.md": (
+        "# График работ\n\n"
+        "Строительство начинается 1 июня 2026. Срок сдачи — 30 ноября 2027.\n"
+    ),
+}
+
+
+async def main() -> int:
+    with tempfile.TemporaryDirectory() as tmp:
+        coll = Path(tmp) / "demo" / "qmd_collections" / "demo"
+        coll.mkdir(parents=True)
+        eng = get_or_create_engine(coll)
+        eng.warmup()
+
+        # Ingest
+        for name, text in SAMPLES.items():
+            result = eng.index_text(text, source_path=name)
+            print(f"  + {name}: chunks={result.chunks_indexed} vectors={result.vectors_indexed} skipped={result.skipped}")
+
+        # Status
+        print("\nStatus:", eng.status())
+
+        # Search BM25
+        print('\n--- BM25 "авторизация JWT" ---')
+        for h in eng.search("авторизация JWT"):
+            print(f"  {h.score:.3f} {h.file_path}: {h.snippet(60)}")
+
+        # Search vector
+        print('\n--- Vector "как устроен вход в систему" ---')
+        for h in eng.vsearch("как устроен вход в систему"):
+            print(f"  {h.score:.3f} {h.file_path}: {h.snippet(60)}")
+
+        # Hybrid query
+        print('\n--- Hybrid "когда сдача объекта" ---')
+        for h in eng.query("когда сдача объекта", limit=3, use_rerank=False):
+            print(f"  {h.score:.3f} {h.source_annotation()}: {h.snippet(60)}")
+
+        eng.close()
+        print("\nAll good [OK]")
+        return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(asyncio.run(main()))
diff --git a/src/config.py b/src/config.py
index 8432f36..e4d04b6 100644
--- a/src/config.py
+++ b/src/config.py
@@ -2,7 +2,7 @@
 
 import os
 from pathlib import Path
-from typing import Any, Dict
+from typing import Any, Dict, Optional, Tuple
 
 import yaml
 
@@ -43,3 +43,25 @@ def resolve_opencode_credentials(config: Dict[str, Any] | None = None) -> tuple[
         "OPENCODE_URL", "https://opencode.ai/zen/v1"
     )
     return api_key, base_url
+
+
+def resolve_qmd_config(config: Dict[str, Any] | None = None) -> Dict[str, Any]:
+    """Возвращает словарь настроек native RAG engine (config → env → defaults).
+
+    Ключи: ``collection_root``, ``use_rerank``, ``embed_model``,
+    ``chat_model``, ``chat_max_tokens``, ``enabled``, ``auto_index``.
+    """
+    if config is None:
+        config = load_config()
+    rag = config.get("rag", {})
+    return {
+        "collection_root": os.environ.get("QMD_COLLECTION_ROOT") or rag.get("qmd_collection_root", "./processed"),
+        "use_rerank": bool(rag.get("qmd_use_rerank", True)),
+        "embed_model": os.environ.get("RAG_EMBED_MODEL") or rag.get(
+            "embed_model", "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
+        ),
+        "chat_model": rag.get("chat_model", "deepseek-v4-flash-free"),
+        "chat_max_tokens": int(rag.get("chat_max_tokens", 8192)),
+        "enabled": bool(rag.get("enabled", True)),
+        "auto_index": bool(rag.get("auto_index", True)),
+    }
diff --git a/src/ingest/router.py b/src/ingest/router.py
index de49b59..e898803 100644
--- a/src/ingest/router.py
+++ b/src/ingest/router.py
@@ -10,6 +10,19 @@ from src.ingest.extractors.xlsx_extractor import extract_csv, extract_xlsx
 from src.ingest.models import AUDIO_EXTENSIONS, SUPPORTED_DOCUMENT_EXTENSIONS, NormalizedDocument
 
 
+_EXTRACTABLE_SUFFIXES = {
+    ".txt",
+    ".md",
+    ".markdown",
+    ".docx",
+    ".doc",
+    ".pdf",
+    ".xlsx",
+    ".xls",
+    ".csv",
+}
+
+
 def is_audio_file(filename: str) -> bool:
     return Path(filename).suffix.lower() in AUDIO_EXTENSIONS
 
@@ -18,6 +31,11 @@ def is_document_file(filename: str) -> bool:
     return Path(filename).suffix.lower() in SUPPORTED_DOCUMENT_EXTENSIONS
 
 
+def is_extractable(filename: str) -> bool:
+    """True если файл поддерживается extract_document (конвертируется в .md)."""
+    return Path(filename).suffix.lower() in _EXTRACTABLE_SUFFIXES
+
+
 def resolve_upload_kind(filename: str) -> str:
     """Return 'document', 'audio', or raise ValueError."""
     if is_document_file(filename):
diff --git a/src/ingest/stub_writer.py b/src/ingest/stub_writer.py
new file mode 100644
index 0000000..9522920
--- /dev/null
+++ b/src/ingest/stub_writer.py
@@ -0,0 +1,84 @@
+"""Stub-генератор для бинарных файлов, которые не конвертируются в текст.
+
+Создаёт ``<file>.md`` с YAML frontmatter (source/size/mime/kind/project) рядом
+с оригиналом — qmd индексирует stub, и пользователь получает ссылку на
+оригинал в результатах поиска.
+"""
+
+from __future__ import annotations
+
+import mimetypes
+from datetime import datetime
+from pathlib import Path
+from typing import Optional
+
+
+STUB_BODY = "Бинарный файл; для просмотра откройте оригинал.\n"
+
+
+def _detect_mime(file_path: Path) -> str:
+    mime, _ = mimetypes.guess_type(file_path.name)
+    return mime or "application/octet-stream"
+
+
+def _infer_kind(file_path: Path) -> str:
+    """Грубая эвристика по расширению: video/audio/image/archive/other."""
+    suffix = file_path.suffix.lower().lstrip(".")
+    video = {"mp4", "webm", "avi", "mkv", "mov"}
+    audio = {"wav", "mp3", "m4a", "ogg", "flac"}
+    image = {"png", "jpg", "jpeg", "gif", "bmp", "tiff", "webp"}
+    archive = {"zip", "tar", "gz", "bz2", "7z", "rar"}
+    if suffix in video:
+        return "video"
+    if suffix in audio:
+        return "audio"
+    if suffix in image:
+        return "image"
+    if suffix in archive:
+        return "archive"
+    return "other"
+
+
+def write_stub(
+    file_path: Path,
+    project: str,
+    kind: Optional[str] = None,
+) -> Path:
+    """Создаёт ``<file_path>.md`` stub и возвращает путь к нему.
+
+    Поведение:
+    * ``file_path`` должен существовать (для получения size/mime);
+    * frontmatter содержит: ``source`` (имя файла), ``size`` (байт),
+      ``mime``, ``kind``, ``project``, ``original_name``, ``created_at``;
+    * body — одна строка-заглушка;
+    * перезаписывается, если уже существует.
+    """
+    file_path = Path(file_path)
+    if not file_path.exists():
+        raise FileNotFoundError(f"Stub source not found: {file_path}")
+
+    try:
+        size = file_path.stat().st_size
+    except OSError as exc:
+        raise OSError(f"Cannot stat {file_path}: {exc}") from exc
+
+    mime = _detect_mime(file_path)
+    resolved_kind = kind or _infer_kind(file_path)
+    stub_path = file_path.with_suffix(file_path.suffix + ".md")
+    original_name = file_path.name
+
+    frontmatter_lines = [
+        "---",
+        f"source: {original_name}",
+        f"size: {size}",
+        f"mime: {mime}",
+        f"kind: {resolved_kind}",
+        f"project: {project}",
+        f"original_name: {original_name}",
+        f"created_at: {datetime.now().isoformat()}",
+        "---",
+        "",
+    ]
+    content = "\n".join(frontmatter_lines) + STUB_BODY
+    stub_path.write_text(content, encoding="utf-8")
+    return stub_path
diff --git a/src/rag/engine/__init__.py b/src/rag/engine/__init__.py
new file mode 100644
index 0000000..e00e0aa
--- /dev/null
+++ b/src/rag/engine/__init__.py
@@ -0,0 +1,36 @@
+"""Native Python RAG engine.
+
+Гибридный движок BM25 (FTS5) + vector (sqlite-vec с numpy fallback) + LLM-реранкер
+через OpenCode. Хранение: один ``index.sqlite`` на коллекцию. Полностью in-process,
+без внешних демонов.
+
+Публичный API:
+    Engine.index_file(path, project_slug, org_slug) -> dict
+    Engine.index_text(text, source_path, ...) -> dict
+    Engine.search(query, collection, limit=8) -> list[Hit]
+    Engine.vsearch(query, collection, limit=8) -> list[Hit]
+    Engine.query(query, collection, limit=8, use_rerank=True) -> list[Hit]
+    Engine.get(doc_id) -> str
+    Engine.status() -> dict
+    Engine.warmup() -> None
+"""
+
+from .bm25 import Hit
+from .engine import (
+    Engine,
+    IndexResult,
+    all_engines,
+    get_or_create_engine,
+    invalidate_engine,
+)
+from . import embeddings
+
+__all__ = [
+    "Engine",
+    "IndexResult",
+    "Hit",
+    "get_or_create_engine",
+    "invalidate_engine",
+    "all_engines",
+    "embeddings",
+]
diff --git a/src/rag/engine/bm25.py b/src/rag/engine/bm25.py
new file mode 100644
index 0000000..3ad841f
--- /dev/null
+++ b/src/rag/engine/bm25.py
@@ -0,0 +1,90 @@
+"""BM25-поиск через SQLite FTS5 с numpy fallback."""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import List, Optional
+
+import numpy as np
+
+from .db import Database
+
+
+@dataclass
+class Hit:
+    rowid: int
+    file_path: str
+    doc_id: str
+    chunk_index: int
+    start_line: int
+    end_line: int
+    content: str
+    score: float
+
+    def snippet(self, max_chars: int = 200) -> str:
+        text = self.content.strip().replace("\n", " ")
+        return text[:max_chars] + ("…" if len(text) > max_chars else "")
+
+    def source_annotation(self) -> str:
+        return f"[source: {self.file_path}:{self.start_line}-{self.end_line}]"
+
+
+def bm25_search(db: Database, query: str, limit: int = 8) -> List[Hit]:
+    """BM25-поиск: FTS5 если есть, иначе rank_bm25 в памяти."""
+    if not query or not query.strip():
+        return []
+    if db.has_fts5:
+        return _bm25_fts5(db, query, limit)
+    return _bm25_rank(db, query, limit)
+
+
+def _bm25_fts5(db: Database, query: str, limit: int) -> List[Hit]:
+    rows = db.fts_search(query, limit)
+    hits: List[Hit] = []
+    for row in rows:
+        hits.append(Hit(
+            rowid=row["rowid"],
+            file_path=row["file_path"],
+            doc_id=row["doc_id"],
+            chunk_index=row["chunk_index"],
+            start_line=row["start_line"],
+            end_line=row["end_line"],
+            content=row["content"],
+            score=float(-row["score"]),  # bm25() возвращает отрицательное; инвертируем для "выше = лучше"
+        ))
+    return hits
+
+
+def _bm25_rank(db: Database, query: str, limit: int) -> List[Hit]:
+    """In-memory rank_bm25 fallback."""
+    try:
+        from rank_bm25 import BM25Okapi
+    except ImportError:
+        return []
+    rows = db.all_chunks()
+    if not rows:
+        return []
+    from .db import re_split_term
+    tokenized_corpus = [re_split_term(row["content"]) for row in rows]
+    bm25 = BM25Okapi(tokenized_corpus)
+    query_terms = re_split_term(query)
+    if not query_terms:
+        return []
+    scores = bm25.get_scores(query_terms)
+    order = np.argsort(-scores)[:limit]
+    hits: List[Hit] = []
+    for idx in order:
+        if scores[idx] <= 0:
+            continue
+        row = rows[int(idx)]
+        hits.append(Hit(
+            rowid=row["id"],
+            file_path=row["file_path"],
+            doc_id=row["doc_id"],
+            chunk_index=row["chunk_index"],
+            start_line=row["start_line"],
+            end_line=row["end_line"],
+            content=row["content"],
+            score=float(scores[idx]),
+        ))
+    return hits
diff --git a/src/rag/engine/chunker.py b/src/rag/engine/chunker.py
new file mode 100644
index 0000000..e41d7c8
--- /dev/null
+++ b/src/rag/engine/chunker.py
@@ -0,0 +1,115 @@
+"""Markdown-aware рекурсивный чанкер.
+
+Режет текст на куски ~``max_chars`` с ``overlap`` (по умолчанию 15%) по границам
+``\\n\\n`` → ``\\n`` → ``". "`` → ``" "``. Сохраняет код-блоки (```...```) целиком.
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+from typing import List
+
+
+@dataclass
+class Chunk:
+    text: str
+    start_line: int
+    end_line: int
+
+
+_SEPARATORS = ["\n\n", "\n", ". ", " "]
+
+
+def chunk_text(text: str, max_chars: int = 900, overlap: int = 135) -> List[Chunk]:
+    """Режет ``text`` на чанки. Возвращает список ``Chunk`` с координатами строк."""
+    if not text or not text.strip():
+        return []
+    if max_chars <= 0:
+        max_chars = 900
+    if overlap < 0:
+        overlap = 0
+    if overlap >= max_chars:
+        overlap = max_chars // 6
+
+    # Сначала режем по абзацам; если абзац > max_chars — рекурсивно рубим дальше.
+    raw_chunks = _split_recursive(text.strip(), max_chars)
+
+    # Склеиваем мелкие куски, пока они влезают в max_chars.
+    merged = _merge_small(raw_chunks, max_chars)
+
+    # Добавляем overlap: каждый следующий чанк начинается с хвоста предыдущего.
+    overlapped = _apply_overlap(merged, overlap)
+
+    # Считаем номера строк.
+    return _annotate_lines(overlapped, text)
+
+
+def _split_recursive(text: str, max_chars: int, sep_index: int = 0) -> List[str]:
+    if len(text) <= max_chars or sep_index >= len(_SEPARATORS):
+        if len(text) > max_chars:
+            return [text[i:i + max_chars] for i in range(0, len(text), max_chars)]
+        return [text] if text else []
+
+    sep = _SEPARATORS[sep_index]
+    if sep not in text:
+        return _split_recursive(text, max_chars, sep_index + 1)
+
+    parts = text.split(sep)
+    out: List[str] = []
+    for i, part in enumerate(parts):
+        if not part:
+            continue
+        if i > 0:
+            part = sep + part
+        if len(part) <= max_chars:
+            out.append(part)
+        else:
+            out.extend(_split_recursive(part, max_chars, sep_index + 1))
+    return out
+
+
+def _merge_small(chunks: List[str], max_chars: int) -> List[str]:
+    merged: List[str] = []
+    buf = ""
+    for chunk in chunks:
+        if not chunk:
+            continue
+        candidate = (buf + "\n" + chunk).strip() if buf else chunk
+        if len(candidate) <= max_chars:
+            buf = candidate
+        else:
+            if buf:
+                merged.append(buf)
+            if len(chunk) > max_chars:
+                buf = chunk[:max_chars]
+            else:
+                buf = chunk
+    if buf:
+        merged.append(buf)
+    return merged
+
+
+def _apply_overlap(chunks: List[str], overlap: int) -> List[str]:
+    if overlap <= 0 or len(chunks) <= 1:
+        return chunks
+    result = [chunks[0]]
+    for prev, curr in zip(chunks, chunks[1:]):
+        tail = prev[-overlap:] if len(prev) > overlap else prev
+        result.append((tail + "\n" + curr).strip())
+    return result
+
+
+def _annotate_lines(chunks: List[str], original: str) -> List[Chunk]:
+    """Грубая аннотация: считаем сколько строк оригинала прошёл каждый чанк."""
+    out: List[Chunk] = []
+    line_no = 1
+    for chunk in chunks:
+        lines_in_chunk = chunk.count("\n") + 1
+        out.append(Chunk(
+            text=chunk,
+            start_line=line_no,
+            end_line=line_no + lines_in_chunk - 1,
+        ))
+        line_no += lines_in_chunk
+    return out
diff --git a/src/rag/engine/db.py b/src/rag/engine/db.py
new file mode 100644
index 0000000..e38f818
--- /dev/null
+++ b/src/rag/engine/db.py
@@ -0,0 +1,259 @@
+"""SQLite-обёртка для коллекции: open, schema init, execute, fallback detection.
+
+Один файл ``index.sqlite`` хранит:
+- таблицу ``chunks`` (id, doc_id, file_path, chunk_index, content, start_line, end_line, file_hash, mtime);
+- virtual table ``chunks_fts`` (FTS5) по content+file_path;
+- virtual table ``chunks_vec`` (sqlite-vec) по embedding.
+
+Fallback-стратегии:
+- нет FTS5 в сборке Python → in-memory rank_bm25 (медленно, но работает);
+- нет sqlite-vec → in-memory numpy cosine.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import logging
+import sqlite3
+import struct
+from contextlib import contextmanager
+from pathlib import Path
+from typing import Any, Iterable, Iterator, List, Optional, Sequence, Tuple
+
+import numpy as np
+
+from . import embeddings
+
+logger = logging.getLogger(__name__)
+
+
+def _detect_fts5(conn: sqlite3.Connection) -> bool:
+    try:
+        conn.execute("CREATE VIRTUAL TABLE _fts5_probe USING fts5(c)")
+        conn.execute("DROP TABLE _fts5_probe")
+        return True
+    except sqlite3.OperationalError:
+        return False
+
+
+def _detect_sqlite_vec(conn: sqlite3.Connection) -> bool:
+    try:
+        import sqlite_vec
+        conn.enable_load_extension(True)
+        sqlite_vec.load(conn)
+        conn.execute("CREATE VIRTUAL TABLE _vec_probe USING vec0(embedding float[4])")
+        conn.execute("DROP TABLE _vec_probe")
+        return True
+    except Exception as exc:
+        logger.debug("sqlite-vec unavailable: %s", exc)
+        return False
+
+
+class Database:
+    """Обёртка над sqlite3.Connection с фиксированной схемой."""
+
+    def __init__(self, path: Path) -> None:
+        self.path = Path(path)
+        self.path.parent.mkdir(parents=True, exist_ok=True)
+        self._conn = sqlite3.connect(str(self.path), check_same_thread=False)
+        self._conn.row_factory = sqlite3.Row
+        self._conn.execute("PRAGMA journal_mode=WAL")
+        self._conn.execute("PRAGMA synchronous=NORMAL")
+        self.has_fts5 = _detect_fts5(self._conn)
+        self.has_vec = _detect_sqlite_vec(self._conn)
+        self._init_schema()
+
+    def close(self) -> None:
+        try:
+            self._conn.close()
+        except Exception:
+            pass
+
+    def __enter__(self) -> "Database":
+        return self
+
+    def __exit__(self, exc_type, exc, tb) -> None:
+        self.close()
+
+    def _init_schema(self) -> None:
+        self._conn.executescript(
+            """
+            CREATE TABLE IF NOT EXISTS chunks (
+                id INTEGER PRIMARY KEY AUTOINCREMENT,
+                doc_id TEXT NOT NULL,
+                file_path TEXT NOT NULL,
+                chunk_index INTEGER NOT NULL,
+                content TEXT NOT NULL,
+                start_line INTEGER NOT NULL,
+                end_line INTEGER NOT NULL,
+                file_hash TEXT NOT NULL,
+                mtime REAL NOT NULL,
+                UNIQUE(file_path, chunk_index)
+            );
+            CREATE INDEX IF NOT EXISTS idx_chunks_doc ON chunks(doc_id);
+            CREATE INDEX IF NOT EXISTS idx_chunks_file ON chunks(file_path);
+            """
+        )
+        if self.has_fts5:
+            self._conn.execute(
+                "CREATE VIRTUAL TABLE IF NOT EXISTS chunks_fts USING fts5("
+                "content, file_path, content='chunks', content_rowid='id', "
+                "tokenize='unicode61 remove_diacritics 2'"
+                ")"
+            )
+            self._conn.execute(
+                "CREATE TRIGGER IF NOT EXISTS chunks_ai AFTER INSERT ON chunks BEGIN "
+                "INSERT INTO chunks_fts(rowid, content, file_path) VALUES (new.id, new.content, new.file_path); "
+                "END;"
+            )
+            self._conn.execute(
+                "CREATE TRIGGER IF NOT EXISTS chunks_ad AFTER DELETE ON chunks BEGIN "
+                "INSERT INTO chunks_fts(chunks_fts, rowid, content, file_path) VALUES('delete', old.id, old.content, old.file_path); "
+                "END;"
+            )
+            self._conn.execute(
+                "CREATE TRIGGER IF NOT EXISTS chunks_au AFTER UPDATE ON chunks BEGIN "
+                "INSERT INTO chunks_fts(chunks_fts, rowid, content, file_path) VALUES('delete', old.id, old.content, old.file_path); "
+                "INSERT INTO chunks_fts(rowid, content, file_path) VALUES (new.id, new.content, new.file_path); "
+                "END;"
+            )
+        if self.has_vec:
+            self._conn.execute(
+                f"CREATE VIRTUAL TABLE IF NOT EXISTS chunks_vec USING vec0("
+                f"embedding float[{embeddings.EMBED_DIM}], id INTEGER"
+                ")"
+            )
+        self._conn.commit()
+
+    @contextmanager
+    def transaction(self) -> Iterator[sqlite3.Connection]:
+        try:
+            yield self._conn
+            self._conn.commit()
+        except Exception:
+            self._conn.rollback()
+            raise
+
+    def execute(self, sql: str, params: Sequence[Any] = ()) -> sqlite3.Cursor:
+        return self._conn.execute(sql, params)
+
+    def executemany(self, sql: str, params_list: Iterable[Sequence[Any]]) -> sqlite3.Cursor:
+        return self._conn.executemany(sql, params_list)
+
+    def fetchall(self, sql: str, params: Sequence[Any] = ()) -> List[sqlite3.Row]:
+        return list(self._conn.execute(sql, params).fetchall())
+
+    def insert_chunks(
+        self,
+        doc_id: str,
+        file_path: str,
+        chunks: List[Tuple[int, str, int, int]],
+        file_hash: str,
+        mtime: float,
+    ) -> List[int]:
+        """Вставка/обновление чанков. Возвращает список rowid.
+
+        ``chunks`` — список ``(chunk_index, content, start_line, end_line)``.
+        """
+        ids: List[int] = []
+        with self.transaction():
+            self.execute("DELETE FROM chunks WHERE file_path = ?", (file_path,))
+            for chunk_index, content, start_line, end_line in chunks:
+                cur = self.execute(
+                    "INSERT INTO chunks(doc_id, file_path, chunk_index, content, start_line, end_line, file_hash, mtime) "
+                    "VALUES (?, ?, ?, ?, ?, ?, ?, ?)",
+                    (doc_id, file_path, chunk_index, content, start_line, end_line, file_hash, mtime),
+                )
+                ids.append(int(cur.lastrowid))
+        return ids
+
+    def insert_vectors(self, rowids: List[int], vectors: np.ndarray) -> None:
+        """Вставка эмбеддингов в chunks_vec (вызывается после insert_chunks)."""
+        if not self.has_vec or not rowids:
+            return
+        from sqlite_vec import serialize_float32
+        with self.transaction():
+            for rowid, vec in zip(rowids, vectors):
+                self.execute(
+                    "INSERT INTO chunks_vec(id, embedding) VALUES (?, ?)",
+                    (rowid, sqlite3.Binary(serialize_float32(vec.tolist()))),
+                )
+
+    def get_chunk_by_rowid(self, rowid: int) -> Optional[sqlite3.Row]:
+        rows = self.fetchall("SELECT * FROM chunks WHERE id = ?", (rowid,))
+        return rows[0] if rows else None
+
+    def get_chunks_by_doc(self, doc_id: str) -> List[sqlite3.Row]:
+        return self.fetchall(
+            "SELECT * FROM chunks WHERE doc_id = ? ORDER BY chunk_index",
+            (doc_id,),
+        )
+
+    def all_chunks(self) -> List[sqlite3.Row]:
+        return self.fetchall("SELECT * FROM chunks")
+
+    def file_hash(self, file_path: str) -> Optional[str]:
+        rows = self.fetchall(
+            "SELECT file_hash FROM chunks WHERE file_path = ? LIMIT 1",
+            (file_path,),
+        )
+        return rows[0]["file_hash"] if rows else None
+
+    def delete_by_file(self, file_path: str) -> None:
+        with self.transaction():
+            self.execute("DELETE FROM chunks WHERE file_path = ?", (file_path,))
+
+    def count(self) -> Tuple[int, int]:
+        chunks = self.execute("SELECT COUNT(*) AS c FROM chunks").fetchone()["c"]
+        files = self.execute("SELECT COUNT(DISTINCT file_path) AS c FROM chunks").fetchone()["c"]
+        return int(chunks), int(files)
+
+    def fts_search(self, query: str, limit: int) -> List[sqlite3.Row]:
+        if not self.has_fts5 or not query.strip():
+            return []
+        return self.fetchall(
+            "SELECT c.id AS rowid, c.file_path, c.doc_id, c.chunk_index, c.start_line, c.end_line, c.content, "
+            "       bm25(chunks_fts) AS score "
+            "FROM chunks_fts f "
+            "JOIN chunks c ON c.id = f.rowid "
+            "WHERE chunks_fts MATCH ? "
+            "ORDER BY score ASC LIMIT ?",
+            (_escape_fts(query), limit),
+        )
+
+    def vec_search(self, query_vec: np.ndarray, limit: int) -> List[Tuple[int, float]]:
+        if not self.has_vec or query_vec is None:
+            return []
+        from sqlite_vec import serialize_float32
+        # sqlite-vec 0.1.x knn queries require explicit "AND k = ?" (no LIMIT support).
+        cur = self._conn.execute(
+            "SELECT id, distance FROM chunks_vec WHERE embedding MATCH ? AND k = ?",
+            (sqlite3.Binary(serialize_float32(query_vec.tolist())), int(limit)),
+        )
+        # sqlite-vec distance: меньше = ближе; конвертируем в similarity.
+        return [(int(row[0]), float(row[1])) for row in cur.fetchall()]
+
+    def close_and_remove(self) -> None:
+        self.close()
+        try:
+            self.path.unlink()
+        except OSError:
+            pass
+
+
+def _escape_fts(query: str) -> str:
+    """Простая защита от FTS5-синтаксиса: разбиваем на термы, добавляем звёздочки."""
+    terms = [t for t in re_split_term(query) if t]
+    return " OR ".join(f'"{t}"*' for t in terms)
+
+
+def re_split_term(query: str) -> List[str]:
+    import re
+    return re.findall(r"[\w\u0400-\u04ff]{2,}", query or "")
+
+
+def hash_file(path: Path) -> str:
+    h = hashlib.sha1()
+    h.update(str(path).encode("utf-8"))
+    h.update(path.read_bytes() if path.exists() else b"")
+    return h.hexdigest()
diff --git a/src/rag/engine/embeddings.py b/src/rag/engine/embeddings.py
new file mode 100644
index 0000000..784f540
--- /dev/null
+++ b/src/rag/engine/embeddings.py
@@ -0,0 +1,59 @@
+"""Embedding-генератор через sentence-transformers (lazy singleton).
+
+Используется ``paraphrase-multilingual-MiniLM-L12-v2`` — 384 dim, мультиязычный
+(рус/англ), ~50 MB, CPU-friendly.
+"""
+
+from __future__ import annotations
+
+import os
+from threading import Lock
+from typing import List, Optional
+
+import numpy as np
+
+DEFAULT_MODEL = os.environ.get(
+    "RAG_EMBED_MODEL",
+    "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
+)
+EMBED_DIM = 384
+
+
+_model = None
+_lock = Lock()
+
+
+def get_model():
+    """Lazy-load + cache embedding-модели (singleton)."""
+    global _model
+    if _model is not None:
+        return _model
+    with _lock:
+        if _model is None:
+            from sentence_transformers import SentenceTransformer
+            _model = SentenceTransformer(DEFAULT_MODEL)
+    return _model
+
+
+def is_loaded() -> bool:
+    return _model is not None
+
+
+def warmup() -> None:
+    """Прогрев модели; используется в FastAPI lifespan."""
+    get_model()
+
+
+def embed_texts(texts: List[str]) -> np.ndarray:
+    """Возвращает ``np.ndarray`` формы ``(len(texts), EMBED_DIM)``, dtype=float32."""
+    if not texts:
+        return np.zeros((0, EMBED_DIM), dtype=np.float32)
+    model = get_model()
+    vectors = model.encode(texts, convert_to_numpy=True, show_progress_bar=False)
+    return np.asarray(vectors, dtype=np.float32)
+
+
+def embed_query(query: str) -> Optional[np.ndarray]:
+    if not query or not query.strip():
+        return None
+    return embed_texts([query])[0]
diff --git a/src/rag/engine/engine.py b/src/rag/engine/engine.py
new file mode 100644
index 0000000..33d33ac
--- /dev/null
+++ b/src/rag/engine/engine.py
@@ -0,0 +1,238 @@
+"""Главный фасад native RAG engine.
+
+Использование::
+
+    engine = Engine.from_paths(project_dir=Path("processed/merakom/qmd_collections/2026"))
+    engine.warmup()
+    engine.index_text("...", source_path="foo.md", doc_id="foo")
+    hits = engine.query("как устроена авторизация")
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+from dataclasses import asdict, dataclass
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+import numpy as np
+
+from . import embeddings
+from .bm25 import Hit, bm25_search
+from .chunker import Chunk, chunk_text
+from .db import Database, hash_file
+from .hybrid import rrf_fuse
+from .rerank import llm_rerank
+from .vector import vector_search
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class IndexResult:
+    doc_id: str
+    file_path: str
+    chunks_indexed: int
+    vectors_indexed: int
+    skipped: bool
+
+
+class Engine:
+    """In-process RAG: BM25 + vector + LLM rerank."""
+
+    def __init__(self, db: Database) -> None:
+        self.db = db
+
+    @classmethod
+    def from_paths(cls, project_dir: Path) -> "Engine":
+        """Открывает/создаёт ``index.sqlite`` в ``project_dir``."""
+        project_dir = Path(project_dir)
+        project_dir.mkdir(parents=True, exist_ok=True)
+        return cls(Database(project_dir / "index.sqlite"))
+
+    def warmup(self) -> None:
+        """Прогревает embedding-модель (вызывать из FastAPI lifespan)."""
+        try:
+            embeddings.warmup()
+        except Exception as exc:
+            logger.warning("[engine] warmup failed: %s", exc)
+
+    def is_ready(self) -> bool:
+        return embeddings.is_loaded()
+
+    # ---- Indexing ----
+    def index_text(
+        self,
+        text: str,
+        *,
+        source_path: str,
+        doc_id: Optional[str] = None,
+        file_hash: Optional[str] = None,
+        mtime: Optional[float] = None,
+        max_chars: int = 900,
+        overlap: int = 135,
+    ) -> IndexResult:
+        doc_id = doc_id or _doc_id_from_path(source_path)
+        mtime = mtime if mtime is not None else _now()
+        file_hash = file_hash or hashlib_text(text)
+
+        existing_hash = self.db.file_hash(source_path)
+        if existing_hash == file_hash:
+            return IndexResult(
+                doc_id=doc_id, file_path=source_path,
+                chunks_indexed=0, vectors_indexed=0, skipped=True,
+            )
+
+        chunks = chunk_text(text, max_chars=max_chars, overlap=overlap)
+        if not chunks:
+            return IndexResult(
+                doc_id=doc_id, file_path=source_path,
+                chunks_indexed=0, vectors_indexed=0, skipped=True,
+            )
+
+        chunk_records = [
+            (i, ch.text, ch.start_line, ch.end_line) for i, ch in enumerate(chunks)
+        ]
+        rowids = self.db.insert_chunks(doc_id, source_path, chunk_records, file_hash, mtime)
+
+        vectors_indexed = 0
+        try:
+            vectors = embeddings.embed_texts([ch.text for ch in chunks])
+            self.db.insert_vectors(rowids, vectors)
+            vectors_indexed = len(rowids)
+        except Exception as exc:
+            logger.warning("[engine] embedding failed for %s: %s", source_path, exc)
+
+        return IndexResult(
+            doc_id=doc_id, file_path=source_path,
+            chunks_indexed=len(rowids), vectors_indexed=vectors_indexed, skipped=False,
+        )
+
+    def index_file(
+        self,
+        file_path: Path,
+        *,
+        doc_id: Optional[str] = None,
+        max_chars: int = 900,
+        overlap: int = 135,
+    ) -> IndexResult:
+        file_path = Path(file_path)
+        if not file_path.exists():
+            raise FileNotFoundError(file_path)
+        text = file_path.read_text(encoding="utf-8", errors="replace")
+        mtime = file_path.stat().st_mtime
+        file_hash = hash_file(file_path)
+        return self.index_text(
+            text,
+            source_path=str(file_path),
+            doc_id=doc_id,
+            file_hash=file_hash,
+            mtime=mtime,
+            max_chars=max_chars,
+            overlap=overlap,
+        )
+
+    def delete(self, file_path: str) -> int:
+        cur = self.db.execute("SELECT COUNT(*) AS c FROM chunks WHERE file_path = ?", (file_path,))
+        before = int(cur.fetchone()["c"])
+        self.db.delete_by_file(file_path)
+        return before
+
+    # ---- Search ----
+    def search(self, query: str, limit: int = 8) -> List[Hit]:
+        return bm25_search(self.db, query, limit=limit)
+
+    def vsearch(self, query: str, limit: int = 8) -> List[Hit]:
+        return vector_search(self.db, query, limit=limit)
+
+    def query(
+        self,
+        query_text: str,
+        limit: int = 8,
+        use_rerank: bool = True,
+    ) -> List[Hit]:
+        bm25_hits = bm25_search(self.db, query_text, limit=max(limit * 3, 20))
+        vec_hits = vector_search(self.db, query_text, limit=max(limit * 3, 20))
+        fused = rrf_fuse(bm25_hits, vec_hits, k=60)
+        if use_rerank and len(fused) > 1:
+            reranked = llm_rerank(query_text, fused[:20])
+            if reranked is not None:
+                fused = reranked
+        return fused[:limit]
+
+    def get(self, doc_id: str) -> str:
+        rows = self.db.get_chunks_by_doc(doc_id)
+        return "\n\n".join(row["content"] for row in rows)
+
+    def status(self) -> Dict[str, Any]:
+        chunks, files = self.db.count()
+        return {
+            "chunks": chunks,
+            "files": files,
+            "embedding_dim": embeddings.EMBED_DIM,
+            "engine": "sqlite-vec" if self.db.has_vec else "numpy",
+            "fts": "fts5" if self.db.has_fts5 else "rank_bm25",
+            "embedding_loaded": embeddings.is_loaded(),
+        }
+
+    def close(self) -> None:
+        self.db.close()
+
+
+def _doc_id_from_path(path: str) -> str:
+    import hashlib
+    return hashlib.sha1(path.encode("utf-8")).hexdigest()[:16]
+
+
+def _now() -> float:
+    import time
+    return time.time()
+
+
+def hashlib_text(text: str) -> str:
+    import hashlib
+    return hashlib.sha1((text or "").encode("utf-8")).hexdigest()
+
+
+# ---- Registry: один Engine на (org, project_slug) ----
+_REGISTRY: Dict[str, Engine] = {}
+_REGISTRY_LOCKS: Dict[str, "_EngineLock"] = {}
+
+
+class _EngineLock:
+    def __init__(self) -> None:
+        import threading
+        self._lock = threading.Lock()
+
+    def __enter__(self):
+        self._lock.acquire()
+        return self
+
+    def __exit__(self, *args):
+        self._lock.release()
+
+
+def get_or_create_engine(project_dir: Path) -> Engine:
+    """Singleton Engine на ``project_dir`` (потокобезопасно)."""
+    key = str(Path(project_dir).resolve())
+    if key in _REGISTRY:
+        return _REGISTRY[key]
+    lock = _REGISTRY_LOCKS.setdefault(key, _EngineLock())
+    with lock:
+        if key not in _REGISTRY:
+            _REGISTRY[key] = Engine.from_paths(Path(project_dir))
+    return _REGISTRY[key]
+
+
+def invalidate_engine(project_dir: Path) -> None:
+    key = str(Path(project_dir).resolve())
+    engine = _REGISTRY.pop(key, None)
+    _REGISTRY_LOCKS.pop(key, None)
+    if engine is not None:
+        engine.close()
+
+
+def all_engines() -> List[Engine]:
+    return list(_REGISTRY.values())
diff --git a/src/rag/engine/hybrid.py b/src/rag/engine/hybrid.py
new file mode 100644
index 0000000..ffdbb0f
--- /dev/null
+++ b/src/rag/engine/hybrid.py
@@ -0,0 +1,36 @@
+"""Reciprocal Rank Fusion (RRF) — фьюжн нескольких списков выдач.
+
+``score(d) = sum(1 / (k + rank_i(d)))`` где rank_i начинается с 1.
+``k=60`` — стандарт (Cormack et al., 2009).
+"""
+
+from __future__ import annotations
+
+from typing import Dict, Iterable, List
+
+from .bm25 import Hit
+
+
+def rrf_fuse(*lists: Iterable[Hit], k: int = 60) -> List[Hit]:
+    """Фьюзит несколько списков ``Hit`` через RRF."""
+    scores: Dict[int, float] = {}
+    by_rowid: Dict[int, Hit] = {}
+    for lst in lists:
+        for rank, hit in enumerate(lst, start=1):
+            scores[hit.rowid] = scores.get(hit.rowid, 0.0) + 1.0 / (k + rank)
+            by_rowid[hit.rowid] = hit
+    sorted_ids = sorted(scores.keys(), key=lambda rid: -scores[rid])
+    result: List[Hit] = []
+    for rid in sorted_ids:
+        base = by_rowid[rid]
+        result.append(Hit(
+            rowid=base.rowid,
+            file_path=base.file_path,
+            doc_id=base.doc_id,
+            chunk_index=base.chunk_index,
+            start_line=base.start_line,
+            end_line=base.end_line,
+            content=base.content,
+            score=scores[rid],
+        ))
+    return result
diff --git a/src/rag/engine/rerank.py b/src/rag/engine/rerank.py
new file mode 100644
index 0000000..39259bd
--- /dev/null
+++ b/src/rag/engine/rerank.py
@@ -0,0 +1,105 @@
+"""LLM-реранкер через OpenCode/DeepSeek.
+
+Отправляет top-N кандидатов с промптом «верни JSON-список rowid, отсортированных
+по релевантности». При любой ошибке возвращает ``None`` → caller использует
+нереранкнутый список.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import re
+from typing import List, Optional
+
+from .bm25 import Hit
+
+logger = logging.getLogger(__name__)
+
+
+RERANK_PROMPT = """Ты — реранкер для поисковой выдачи. Тебе дан запрос и {n} фрагментов документов.
+Верни JSON-список ``rowid`` (целые числа) В ПОРЯДКЕ убывания релевантности запросу.
+Не добавляй пояснений, только JSON.
+
+Запрос: {query}
+
+Фрагменты:
+{chunks}
+
+Верни ТОЛЬКО JSON-массив rowid, например: ``[42, 17, 5]``
+"""
+
+
+def llm_rerank(
+    query: str,
+    hits: List[Hit],
+    api_key: str = "",
+    base_url: str = "https://opencode.ai/zen/v1",
+    model: str = "deepseek-v4-flash-free",
+    top_k: int = 20,
+) -> Optional[List[Hit]]:
+    """Отправляет ``top_k`` чанков в LLM и возвращает пересортированный список.
+
+    Возвращает ``None`` если запрос не удался.
+    """
+    if not hits:
+        return []
+    api_key = api_key or os.environ.get("OPENCODE_API_KEY", "")
+    if not api_key:
+        logger.warning("[rerank] OPENCODE_API_KEY not set, skipping")
+        return None
+
+    candidates = hits[:top_k]
+    chunks_text = "\n\n".join(
+        f"rowid={h.rowid}: {h.snippet(400)}" for h in candidates
+    )
+    prompt = RERANK_PROMPT.format(n=len(candidates), query=query, chunks=chunks_text)
+
+    try:
+        from openai import AsyncOpenAI
+        client = AsyncOpenAI(base_url=base_url, api_key=api_key)
+        response = client.chat.completions.create(
+            model=model,
+            messages=[{"role": "user", "content": prompt}],
+            temperature=0.0,
+            max_tokens=512,
+        )
+        content = (response.choices[0].message.content or "").strip()
+        order_ids = _parse_ids(content)
+        if not order_ids:
+            return None
+    except Exception as exc:
+        logger.warning("[rerank] LLM call failed: %s", exc)
+        return None
+
+    by_id = {h.rowid: h for h in candidates}
+    result: List[Hit] = []
+    for rid in order_ids:
+        hit = by_id.get(int(rid))
+        if hit is None:
+            continue
+        result.append(hit)
+    for h in candidates:
+        if h.rowid not in {r.rowid for r in result}:
+            result.append(h)
+    return result
+
+
+def _parse_ids(content: str) -> List[int]:
+    match = re.search(r"\[[^\]]*\]", content, re.DOTALL)
+    if not match:
+        return []
+    try:
+        data = json.loads(match.group(0))
+    except json.JSONDecodeError:
+        return []
+    if not isinstance(data, list):
+        return []
+    out: List[int] = []
+    for item in data:
+        try:
+            out.append(int(item))
+        except (TypeError, ValueError):
+            continue
+    return out
diff --git a/src/rag/engine/vector.py b/src/rag/engine/vector.py
new file mode 100644
index 0000000..c8d442c
--- /dev/null
+++ b/src/rag/engine/vector.py
@@ -0,0 +1,80 @@
+"""Vector-семантический поиск через sqlite-vec с numpy fallback."""
+
+from __future__ import annotations
+
+from typing import List, Optional
+
+import numpy as np
+
+from .bm25 import Hit
+from .db import Database
+from . import embeddings
+
+
+def vector_search(
+    db: Database,
+    query: str,
+    limit: int = 8,
+) -> List[Hit]:
+    """Vector-поиск: sqlite-vec если есть, иначе numpy cosine в памяти."""
+    if not query or not query.strip():
+        return []
+    qv = embeddings.embed_query(query)
+    if qv is None:
+        return []
+    if db.has_vec:
+        return _vec_sqlite_vec(db, qv, limit)
+    return _vec_numpy(db, qv, limit)
+
+
+def _vec_sqlite_vec(db: Database, qv: np.ndarray, limit: int) -> List[Hit]:
+    pairs = db.vec_search(qv, limit)
+    hits: List[Hit] = []
+    for rowid, distance in pairs:
+        # distance в sqlite-vec = L2 (или cosine, зависит от версии); нормализуем к [0,1]
+        similarity = 1.0 / (1.0 + float(distance))
+        row = db.get_chunk_by_rowid(rowid)
+        if row is None:
+            continue
+        hits.append(Hit(
+            rowid=rowid,
+            file_path=row["file_path"],
+            doc_id=row["doc_id"],
+            chunk_index=row["chunk_index"],
+            start_line=row["start_line"],
+            end_line=row["end_line"],
+            content=row["content"],
+            score=similarity,
+        ))
+    return hits
+
+
+def _vec_numpy(db: Database, qv: np.ndarray, limit: int) -> List[Hit]:
+    rows = db.all_chunks()
+    if not rows:
+        return []
+    texts = [row["content"] for row in rows]
+    vectors = embeddings.embed_texts(texts)
+    if vectors.shape[0] == 0:
+        return []
+    qv_norm = qv / (np.linalg.norm(qv) + 1e-12)
+    mat_norm = vectors / (np.linalg.norm(vectors, axis=1, keepdims=True) + 1e-12)
+    sims = mat_norm @ qv_norm
+    order = np.argsort(-sims)[:limit]
+    hits: List[Hit] = []
+    for idx in order:
+        sim = float(sims[idx])
+        if sim <= 0:
+            continue
+        row = rows[int(idx)]
+        hits.append(Hit(
+            rowid=row["id"],
+            file_path=row["file_path"],
+            doc_id=row["doc_id"],
+            chunk_index=row["chunk_index"],
+            start_line=row["start_line"],
+            end_line=row["end_line"],
+            content=row["content"],
+            score=sim,
+        ))
+    return hits
diff --git a/src/rag/formatter.py b/src/rag/formatter.py
index d94d611..219a107 100644
--- a/src/rag/formatter.py
+++ b/src/rag/formatter.py
@@ -1,4 +1,4 @@
-"""Форматирование документа совещания для индексации в LightRAG."""
+"""Форматирование документа совещания для индексации в qmd (knowledge base)."""
 
 import json
 from datetime import datetime
@@ -12,7 +12,7 @@ def format_meeting_document(
     metadata: Dict[str, Any],
     source_filename: str,
 ) -> str:
-    """Собирает текстовый документ для вставки в LightRAG.
+    """Собирает текстовый документ для индексации в qmd.
 
     Сохраняет полную расшифровку + метаданные + извлечённые сущности.
     """
diff --git a/src/rag/indexer.py b/src/rag/indexer.py
deleted file mode 100644
index f1a86e6..0000000
--- a/src/rag/indexer.py
+++ /dev/null
@@ -1,184 +0,0 @@
-"""Индексация протоколов совещаний в LightRAG (проектные индексы + глобальный)."""
-
-import os
-from pathlib import Path
-from typing import Optional
-
-from dotenv import load_dotenv
-from lightrag import LightRAG
-from lightrag.utils import EmbeddingFunc
-from openai import AsyncOpenAI
-from sentence_transformers import SentenceTransformer
-
-load_dotenv()
-
-# ------------------------------------------------------------------
-# Shared embedding model (lazy-loaded)
-# ------------------------------------------------------------------
-_embed_model_instance: Optional[SentenceTransformer] = None
-
-
-def _get_embed_model() -> SentenceTransformer:
-    global _embed_model_instance
-    if _embed_model_instance is None:
-        _embed_model_instance = SentenceTransformer(
-            "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
-        )
-    return _embed_model_instance
-
-
-async def _embed_func(texts: list[str]):
-    model = _get_embed_model()
-    embeddings = model.encode(texts, convert_to_numpy=True)
-    return embeddings
-
-
-EMBED_CONFIG = EmbeddingFunc(
-    embedding_dim=384,
-    max_token_size=512,
-    func=_embed_func,
-)
-
-
-# ------------------------------------------------------------------
-# LLM for indexing (mimo-v2.5-free via OpenCode)
-# ------------------------------------------------------------------
-def _get_opencode_llm_func(
-    model: str = "mimo-v2.5-free",
-    api_key: str = "",
-    base_url: str = "https://opencode.ai/zen/v1",
-):
-    if not api_key:
-        raise ValueError(
-            "OPENCODE_API_KEY не задан. Укажите rag.opencode_api_key в config.yaml "
-            "или переменную окружения OPENCODE_API_KEY."
-        )
-    client = AsyncOpenAI(base_url=base_url, api_key=api_key)
-
-    async def llm_func(prompt, system_prompt=None, history_messages=[], **kwargs):
-        messages = []
-        if system_prompt:
-            messages.append({"role": "system", "content": system_prompt})
-        if history_messages:
-            messages.extend(history_messages)
-        messages.append({"role": "user", "content": prompt})
-
-        response = await client.chat.completions.create(
-            model=model,
-            messages=messages,
-            temperature=kwargs.get("temperature", 0.3),
-            max_tokens=kwargs.get("max_tokens", 1024),
-        )
-        content = response.choices[0].message.content
-        return content if content is not None else ""
-
-    return llm_func
-
-
-# ------------------------------------------------------------------
-# Project index helpers
-# ------------------------------------------------------------------
-def _normalize_project_name(name: str) -> str:
-    """Нормализует имя проекта для использования в пути."""
-    import re
-    name = re.sub(r'[^\w\-_]', '_', name)
-    return name.strip('_') or "default"
-
-
-def get_project_index_dir(working_dir_base: Path, project_name: str) -> Path:
-    """Возвращает путь к кэшу индекса проекта."""
-    norm = _normalize_project_name(project_name)
-    return working_dir_base / norm
-
-
-def get_global_index_dir(working_dir_base: Path) -> Path:
-    """Возвращает путь к глобальному индексу."""
-    return working_dir_base / "_global"
-
-
-async def get_project_rag(
-    project_name: str,
-    working_dir_base: Path,
-    model: str = "mimo-v2.5-free",
-    api_key: str = "",
-    base_url: str = "https://opencode.ai/zen/v1",
-) -> LightRAG:
-    """Возвращает инициализированный LightRAG для проекта."""
-    cache_dir = get_project_index_dir(working_dir_base, project_name)
-    cache_dir.mkdir(parents=True, exist_ok=True)
-
-    llm_func = _get_opencode_llm_func(model, api_key=api_key, base_url=base_url)
-
-    rag = LightRAG(
-        working_dir=str(cache_dir),
-        llm_model_func=llm_func,
-        embedding_func=EMBED_CONFIG,
-    )
-    await rag.initialize_storages()
-    return rag
-
-
-async def get_global_rag(
-    working_dir_base: Path,
-    model: str = "mimo-v2.5-free",
-    api_key: str = "",
-    base_url: str = "https://opencode.ai/zen/v1",
-) -> LightRAG:
-    """Возвращает инициализированный LightRAG для глобального индекса."""
-    cache_dir = get_global_index_dir(working_dir_base)
-    cache_dir.mkdir(parents=True, exist_ok=True)
-
-    llm_func = _get_opencode_llm_func(model, api_key=api_key, base_url=base_url)
-
-    rag = LightRAG(
-        working_dir=str(cache_dir),
-        llm_model_func=llm_func,
-        embedding_func=EMBED_CONFIG,
-    )
-    await rag.initialize_storages()
-    return rag
-
-
-# ------------------------------------------------------------------
-# Indexing API
-# ------------------------------------------------------------------
-async def index_meeting(
-    doc_text: str,
-    global_doc_text: str,
-    project_name: str,
-    working_dir_base: Path,
-    model: str = "mimo-v2.5-free",
-    api_key: str = "",
-    base_url: str = "https://opencode.ai/zen/v1",
-):
-    """Индексирует документ в проектный и глобальный индексы.
-
-    Args:
-        doc_text: текст для проектного индекса.
-        global_doc_text: текст для глобального индекса (с явным project).
-        project_name: имя проекта.
-        working_dir_base: базовая директория для lightrag_caches.
-        model: модель LLM для индексации.
-    """
-    # Проектный индекс
-    rag_project = await get_project_rag(
-        project_name, working_dir_base, model, api_key=api_key, base_url=base_url
-    )
-    await rag_project.ainsert(doc_text)
-
-    # Глобальный индекс
-    rag_global = await get_global_rag(
-        working_dir_base, model, api_key=api_key, base_url=base_url
-    )
-    await rag_global.ainsert(global_doc_text)
-
-
-async def get_project_names(working_dir_base: Path) -> list[str]:
-    """Возвращает список проектов, для которых есть индекс."""
-    if not working_dir_base.exists():
-        return []
-    projects = []
-    for item in working_dir_base.iterdir():
-        if item.is_dir() and item.name != "_global":
-            projects.append(item.name)
-    return sorted(projects)
diff --git a/src/rag/parser.py b/src/rag/parser.py
index 7505ec7..0c90edf 100644
--- a/src/rag/parser.py
+++ b/src/rag/parser.py
@@ -10,6 +10,11 @@ from openai import AsyncOpenAI
 DEFAULT_CLASSIFY_CHUNK_SIZE = 7000
 DEFAULT_SUMMARY_CHUNK_SIZE = 10000
 CHUNK_OVERLAP = 300
+# Reasoning-модели тратят max_tokens на reasoning; content остаётся пустым при малых лимитах
+CLASSIFY_MAX_TOKENS = 4096
+BRIEF_MAX_TOKENS = 8192
+CHUNK_BRIEF_MAX_TOKENS = 4096
+REDUCE_BRIEF_MAX_TOKENS = 8192
 
 CLASSIFY_PROMPT_TEMPLATE = """Ты — ассистент по анализу протоколов совещаний.
 
@@ -210,6 +215,15 @@ def _merge_metadata(parts: List[Dict[str, Any]], project: str) -> Dict[str, Any]
     return merged
 
 
+def _message_text(message: Any) -> str:
+    """Извлекает текст ответа LLM; reasoning-модели могут вернуть content=None."""
+    content = (message.content or "").strip()
+    if content:
+        return content
+    reasoning = getattr(message, "reasoning_content", None) or getattr(message, "reasoning", None)
+    return (reasoning or "").strip()
+
+
 async def _llm_complete(
     client: AsyncOpenAI,
     model: str,
@@ -223,7 +237,7 @@ async def _llm_complete(
         temperature=temperature,
         max_tokens=max_tokens,
     )
-    return (response.choices[0].message.content or "").strip()
+    return _message_text(response.choices[0].message)
 
 
 async def _classify_chunk(
@@ -238,7 +252,7 @@ async def _classify_chunk(
         sections=", ".join(sections),
         text=text,
     )
-    content = await _llm_complete(client, model, prompt, temperature=0.2, max_tokens=1024)
+    content = await _llm_complete(client, model, prompt, temperature=0.2, max_tokens=CLASSIFY_MAX_TOKENS)
     return _parse_json_response(content, project)
 
 
@@ -286,21 +300,28 @@ async def generate_meeting_brief(
 
     if len(chunks) == 1:
         prompt = BRIEF_SUMMARY_PROMPT.format(project=project, topic=topic, text=chunks[0])
-        content = await _llm_complete(client, model, prompt, temperature=0.3, max_tokens=1024)
+        content = await _llm_complete(client, model, prompt, temperature=0.3, max_tokens=BRIEF_MAX_TOKENS)
         return content or metadata.get("summary", "")
 
     partials: List[str] = []
     total = len(chunks)
     for idx, chunk in enumerate(chunks, start=1):
         prompt = CHUNK_BRIEF_PROMPT.format(part=idx, total=total, text=chunk)
-        partial = await _llm_complete(client, model, prompt, temperature=0.3, max_tokens=512)
+        partial = await _llm_complete(
+            client, model, prompt, temperature=0.3, max_tokens=CHUNK_BRIEF_MAX_TOKENS
+        )
         if partial:
             partials.append(f"[Часть {idx}/{total}]\n{partial}")
 
+    if not partials:
+        return metadata.get("summary", "")
+
     reduce_prompt = REDUCE_BRIEF_PROMPT.format(
         project=project,
         topic=topic,
         partials="\n\n".join(partials),
     )
-    content = await _llm_complete(client, model, reduce_prompt, temperature=0.3, max_tokens=1024)
+    content = await _llm_complete(
+        client, model, reduce_prompt, temperature=0.3, max_tokens=REDUCE_BRIEF_MAX_TOKENS
+    )
     return content or metadata.get("summary", "")
diff --git a/src/rag/qmd/__init__.py b/src/rag/qmd/__init__.py
new file mode 100644
index 0000000..9d1c931
--- /dev/null
+++ b/src/rag/qmd/__init__.py
@@ -0,0 +1,6 @@
+"""qmd knowledge-base integration (Hermes Agent, Nous Research).
+
+Заменяет LightRAG. Локальный движок BM25 (FTS5) + vector (sqlite-vec) + LLM-реранкер
+с хранением в одном .sqlite на коллекцию. Запускается как MCP-сервер через CLI
+``qmd mcp --http --daemon``.
+"""
diff --git a/src/rag/qmd/cache.py b/src/rag/qmd/cache.py
new file mode 100644
index 0000000..74cefa9
--- /dev/null
+++ b/src/rag/qmd/cache.py
@@ -0,0 +1,105 @@
+"""TTL-кэш результатов поиска qmd с инвалидацией по mtime индекса.
+
+Ключ: ``hash(question) + (org_slug, project_slug, mode, use_rerank)``.
+Значение инвалидируется либо по TTL (60 сек), либо при изменении mtime
+``index.sqlite`` коллекции.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import os
+import time
+from pathlib import Path
+from typing import Any, Optional, Tuple
+
+CacheKey = Tuple[str, str, Optional[str], str, bool]
+
+
+def _question_hash(question: str) -> str:
+    return hashlib.sha1((question or "").strip().lower().encode("utf-8")).hexdigest()[:16]
+
+
+def _index_mtime(collection_root: Optional[Path] = None) -> float:
+    """Возвращает mtime самого свежего ``index.sqlite`` в коллекциях org.
+
+    0.0 если ничего не найдено (например, qmd ещё не запускался).
+    """
+    from src.rag.qmd import collections
+    root = collections.get_root(collection_root)
+    if not root.exists():
+        return 0.0
+    latest = 0.0
+    for sqlite_path in root.rglob("index.sqlite"):
+        try:
+            mtime = sqlite_path.stat().st_mtime
+            if mtime > latest:
+                latest = mtime
+        except OSError:
+            continue
+    return latest
+
+
+class SearchCache:
+    """Потокобезопасный (single-loop asyncio) кэш результатов поиска."""
+
+    def __init__(self, ttl_seconds: int = 60) -> None:
+        self.ttl = ttl_seconds
+        self._store: dict[CacheKey, tuple[float, float, str]] = {}
+
+    def _build_key(
+        self,
+        question: str,
+        args: tuple[str, Optional[str], str, bool],
+    ) -> CacheKey:
+        org_slug, project_slug, mode, use_rerank = args
+        return (
+            _question_hash(question),
+            org_slug,
+            project_slug,
+            mode,
+            bool(use_rerank),
+        )
+
+    def get(self, question: str, args: tuple[str, Optional[str], str, bool]) -> Optional[str]:
+        key = self._build_key(question, args)
+        entry = self._store.get(key)
+        if entry is None:
+            return None
+        stored_at, stored_mtime, value = entry
+        current_mtime = _index_mtime()
+        if current_mtime > stored_mtime:
+            self._store.pop(key, None)
+            return None
+        if (time.monotonic() - stored_at) > self.ttl:
+            self._store.pop(key, None)
+            return None
+        return value
+
+    def set(
+        self,
+        question: str,
+        args: tuple[str, Optional[str], str, bool],
+        value: str,
+    ) -> None:
+        if not value:
+            return
+        key = self._build_key(question, args)
+        self._store[key] = (time.monotonic(), _index_mtime(), value)
+
+    def clear(self) -> None:
+        self._store.clear()
+
+    def stats(self) -> dict:
+        return {"entries": len(self._store), "ttl_seconds": self.ttl}
+
+
+_search_cache = None
+
+
+def get_search_cache():
+    """Глобальный singleton (для FastAPI dependency / healthcheck)."""
+    global _search_cache
+    if _search_cache is None:
+        _search_cache = SearchCache(ttl_seconds=60)
+    return _search_cache
diff --git a/src/rag/qmd/collections.py b/src/rag/qmd/collections.py
new file mode 100644
index 0000000..15f44a1
--- /dev/null
+++ b/src/rag/qmd/collections.py
@@ -0,0 +1,121 @@
+"""Управление qmd-коллекциями: per-org / per-project + глобальная.
+
+Коллекция — это каталог с ``index.sqlite`` под
+``<root>/<org>/<project_slug>/`` или ``<root>/<org>/_global/``.
+
+Хранение реализовано в :mod:`src.rag.engine` (sqlite-vec + FTS5).
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+from pathlib import Path
+from typing import List, Optional
+
+logger = logging.getLogger(__name__)
+
+
+def get_root(collection_root: Optional[Path] = None) -> Path:
+    """Возвращает базовый путь коллекций (env → config → default)."""
+    if collection_root is not None:
+        return collection_root
+    env_root = os.environ.get("QMD_COLLECTION_ROOT")
+    if env_root:
+        return Path(env_root)
+    return Path("processed")
+
+
+def get_org_root(org_slug: str, collection_root: Optional[Path] = None) -> Path:
+    norm = _normalize(org_slug) or "default"
+    return get_root(collection_root) / norm
+
+
+def get_project_collection_dir(
+    org_slug: str,
+    project_slug: str,
+    collection_root: Optional[Path] = None,
+) -> Path:
+    return get_org_root(org_slug, collection_root) / "qmd_collections" / _normalize(project_slug)
+
+
+def get_global_collection_dir(
+    org_slug: str,
+    collection_root: Optional[Path] = None,
+) -> Path:
+    return get_org_root(org_slug, collection_root) / "qmd_collections" / "_global"
+
+
+def collection_uri(org_slug: str, project_slug: str) -> str:
+    """qmd:// URI для коллекции."""
+    if not project_slug or project_slug == "_global":
+        return f"qmd://{_normalize(org_slug)}/_global"
+    return f"qmd://{_normalize(org_slug)}/{_normalize(project_slug)}"
+
+
+def _normalize(name: str) -> str:
+    import re
+    cleaned = re.sub(r"[^\w\-_]", "_", (name or "").strip())
+    return cleaned.strip("_") or "default"
+
+
+async def ensure_collection(
+    org_slug: str,
+    project_slug: str,
+    collection_root: Optional[Path] = None,
+) -> Path:
+    """Создаёт каталог коллекции (engine создаст .sqlite при первом index)."""
+    if not project_slug or project_slug == "_global":
+        path = get_global_collection_dir(org_slug, collection_root)
+    else:
+        path = get_project_collection_dir(org_slug, project_slug, collection_root)
+    path.mkdir(parents=True, exist_ok=True)
+    return path
+
+
+async def add_context(
+    org_slug: str,
+    project_slug: str,
+    description: str,
+) -> None:
+    """Сохраняет описание контекста рядом с index.sqlite (best-effort)."""
+    if not project_slug or project_slug == "_global":
+        path = get_global_collection_dir(org_slug)
+    else:
+        path = get_project_collection_dir(org_slug, project_slug)
+    path.mkdir(parents=True, exist_ok=True)
+    context_path = path / "context.md"
+    try:
+        with context_path.open("a", encoding="utf-8") as f:
+            f.write(f"- {description}\n")
+    except OSError as exc:
+        logger.debug("add_context: %s", exc)
+
+
+async def embed(
+    org_slug: str,
+    project_slug: str,
+    force: bool = False,
+) -> int:
+    """Переиндексация коллекции: проходит по всем ``*.md`` в каталоге коллекции.
+
+    Возвращает число проиндексированных файлов.
+    """
+    from src.rag.engine import get_or_create_engine
+    if not project_slug or project_slug == "_global":
+        path = get_global_collection_dir(org_slug)
+    else:
+        path = get_project_collection_dir(org_slug, project_slug)
+    if not path.exists():
+        return 0
+    engine = get_or_create_engine(path)
+    engine.warmup()
+    indexed = 0
+    for md_file in sorted(path.rglob("*.md")):
+        try:
+            result = engine.index_file(md_file)
+            if not result.skipped:
+                indexed += 1
+        except Exception as exc:
+            logger.warning("embed: failed to index %s: %s", md_file, exc)
+    return indexed
diff --git a/src/rag/qmd/indexer.py b/src/rag/qmd/indexer.py
new file mode 100644
index 0000000..bdddb0c
--- /dev/null
+++ b/src/rag/qmd/indexer.py
@@ -0,0 +1,126 @@
+"""Высокоуровневые операции индексации: meetings и documents.
+
+Использует native RAG engine (``src.rag.engine``). На каждый файл:
+1) ``Engine.index_file`` добавляет в проектную коллекцию;
+2) копия с префиксом проекта добавляется в глобальную коллекцию org.
+"""
+
+from __future__ import annotations
+
+from datetime import datetime
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+from src.rag.engine import Engine, get_or_create_engine
+from src.rag.qmd import collections
+
+
+def _global_doc_text(doc_text: str, project_slug: str) -> str:
+    """Добавляет префикс с проектом для глобальной коллекции."""
+    return f"=== Проект: {project_slug} ===\n\n{doc_text}"
+
+
+def _engine_for(org_slug: str, project_slug: Optional[str]) -> Engine:
+    if not project_slug or project_slug == "_global":
+        path = collections.get_global_collection_dir(org_slug)
+    else:
+        path = collections.get_project_collection_dir(org_slug, project_slug)
+    path.mkdir(parents=True, exist_ok=True)
+    return get_or_create_engine(path)
+
+
+async def qmd_index_meeting(
+    org_slug: str,
+    project_slug: str,
+    *,
+    body_path: Path,
+    summary_path: Optional[Path] = None,
+    txt_path: Optional[Path] = None,
+    context_description: Optional[str] = None,
+) -> Dict[str, Any]:
+    """Индексирует протокол совещания: ``<stem>.md`` + ``<stem>_summary.md``."""
+    project_engine = _engine_for(org_slug, project_slug)
+    global_engine = _engine_for(org_slug, "_global")
+
+    written: List[Path] = []
+    files = [body_path, summary_path, txt_path]
+    for src in files:
+        if src and src.exists():
+            result = project_engine.index_file(src)
+            if not result.skipped:
+                written.append(src)
+            if context_description:
+                await collections.add_context(org_slug, project_slug, context_description)
+            # Глобальный индекс: добавляем тот же файл с префиксом проекта.
+            text = src.read_text(encoding="utf-8")
+            prefixed = _global_doc_text(text, project_slug)
+            global_path = global_engine.db.path.parent / src.name
+            global_path.write_text(prefixed, encoding="utf-8")
+            global_engine.index_file(global_path, doc_id=f"{project_slug}_{src.stem}")
+            global_path.unlink(missing_ok=True)
+
+    return {
+        "project_dir": str(project_engine.db.path.parent),
+        "files": [str(p) for p in written],
+        "chunks_indexed": sum(
+            1 for p in written for _ in project_engine.db.fetchall(
+                "SELECT id FROM chunks WHERE file_path = ?", (str(p),)
+            )
+        ),
+    }
+
+
+async def qmd_index_document(
+    org_slug: str,
+    project_slug: str,
+    *,
+    document_dir: Path,
+    extracted_md: Optional[Path] = None,
+    context_description: Optional[str] = None,
+) -> Dict[str, Any]:
+    """Индексирует произвольный документ: PDF/DOCX/XLSX/CSV/TXT/MD или stub."""
+    project_engine = _engine_for(org_slug, project_slug)
+    global_engine = _engine_for(org_slug, "_global")
+
+    written: List[Path] = []
+    sources: List[Path] = []
+    if extracted_md and extracted_md.exists():
+        sources = [extracted_md]
+    if not sources and document_dir.exists():
+        sources = sorted(document_dir.glob("*.md"))
+
+    if not sources:
+        return {"project_dir": str(document_dir), "files": [], "chunks_indexed": 0}
+
+    for src in sources:
+        result = project_engine.index_file(src)
+        if not result.skipped:
+            written.append(src)
+        if context_description:
+            await collections.add_context(org_slug, project_slug, context_description)
+        text = src.read_text(encoding="utf-8")
+        prefixed = _global_doc_text(text, project_slug)
+        global_path = global_engine.db.path.parent / f"{project_slug}__{src.name}"
+        global_path.write_text(prefixed, encoding="utf-8")
+        global_engine.index_file(global_path, doc_id=f"{project_slug}__{src.stem}")
+        global_path.unlink(missing_ok=True)
+
+    return {
+        "project_dir": str(document_dir),
+        "files": [str(p) for p in written],
+        "chunks_indexed": sum(
+            1 for p in written for _ in project_engine.db.fetchall(
+                "SELECT id FROM chunks WHERE file_path = ?", (str(p),)
+            )
+        ),
+    }
+
+
+def build_global_doc_text(doc_text: str, project_slug: str) -> str:
+    """Совместимость со старым API ``format_global_document``."""
+    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
+    header = (
+        f"=== ДОКУМЕНТ (Проект: {project_slug}, ID: {project_slug}_{timestamp}) ===\n"
+        f"Проиндексировано: {datetime.now().isoformat()}\n\n"
+    )
+    return header + _global_doc_text(doc_text, project_slug)
diff --git a/src/rag/qmd/lifespan.py b/src/rag/qmd/lifespan.py
new file mode 100644
index 0000000..75c3244
--- /dev/null
+++ b/src/rag/qmd/lifespan.py
@@ -0,0 +1,57 @@
+"""FastAPI lifespan: прогрев native RAG engine.
+
+Больше не запускает внешний демон — движок in-process.
+
+Использование::
+
+    from contextlib import asynccontextmanager
+    from fastapi import FastAPI
+    from src.rag.qmd.lifespan import qmd_lifespan
+
+    @asynccontextmanager
+    async def lifespan(app: FastAPI):
+        async with qmd_lifespan(app):
+            yield
+
+    app = FastAPI(lifespan=lifespan)
+"""
+
+from __future__ import annotations
+
+import logging
+from contextlib import asynccontextmanager
+from typing import Any, Dict, Optional
+
+logger = logging.getLogger(__name__)
+
+
+@asynccontextmanager
+async def qmd_lifespan(app: Any):
+    """Async context manager: прогревает embedding-модель при старте.
+
+    ``app.state.qmd_status`` устанавливается в ``"ready"`` если warmup успешен,
+    иначе ``"degraded"`` (BM25-only fallback).
+    """
+    app.state.qmd_status = "degraded"
+    app.state.qmd_daemon_pid = None
+
+    try:
+        from src.rag.engine import embeddings
+        embeddings.warmup()
+        app.state.qmd_status = "ready"
+        logger.info("[qmd] native engine ready (in-process)")
+    except Exception as exc:
+        logger.warning("[qmd] warmup failed: %s — degraded mode (BM25-only)", exc)
+        app.state.qmd_status = "degraded"
+
+    try:
+        yield
+    finally:
+        try:
+            from src.rag.engine import all_engines
+            for engine in all_engines():
+                engine.close()
+        except Exception:
+            pass
+        app.state.qmd_status = "down"
+        logger.info("[qmd] engines closed")
diff --git a/src/rag/qmd/mcp_client.py b/src/rag/qmd/mcp_client.py
new file mode 100644
index 0000000..89f63d6
--- /dev/null
+++ b/src/rag/qmd/mcp_client.py
@@ -0,0 +1,175 @@
+"""In-process клиент к native RAG engine (замена старого MCP-клиента).
+
+Сохраняет совместимый API со старым ``mcp_client.py``:
+    qmd_search(query, collection) -> list[dict]
+    qmd_vsearch(query, collection) -> list[dict]
+    qmd_query(query, collection) -> str
+    qmd_get(docid) -> str
+    qmd_status() -> dict
+
+Класс ``QmdMcpError`` оставлен для обратной совместимости (больше не бросается,
+но импорты не ломаются).
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+logger = logging.getLogger(__name__)
+
+
+class QmdMcpError(RuntimeError):
+    """Совместимый exception (больше не используется, оставлен для API)."""
+
+
+def _collection_to_dir(org_slug: str, collection: Optional[str]) -> Path:
+    """``qmd://org/project`` → ``<root>/<org>/qmd_collections/<project>``.
+
+    Также принимает ``Path`` напрямую (для тестов).
+    """
+    from src.rag.qmd.collections import get_project_collection_dir, get_global_collection_dir
+    if collection is None or collection == "":
+        return get_global_collection_dir(org_slug)
+    if isinstance(collection, Path):
+        return collection
+    name = str(collection).rstrip("/").split("/")[-1]
+    if name == "_global" or not name:
+        return get_global_collection_dir(org_slug)
+    return get_project_collection_dir(org_slug, name)
+
+
+def _engine_for(org_slug: str, collection: Optional[str]):
+    from src.rag.engine import get_or_create_engine
+    directory = _collection_to_dir(org_slug, collection)
+    return get_or_create_engine(directory), directory
+
+
+def _hit_to_dict(hit, score_key: str = "score") -> Dict[str, Any]:
+    return {
+        "rowid": hit.rowid,
+        "docid": hit.doc_id,
+        "path": hit.file_path,
+        "file_path": hit.file_path,
+        "chunk_index": hit.chunk_index,
+        "start_line": hit.start_line,
+        "end_line": hit.end_line,
+        "score": float(hit.score),
+        "snippet": hit.snippet(400),
+        "text": hit.content,
+        "source": hit.source_annotation(),
+    }
+
+
+async def qmd_search(
+    query: str,
+    collection: Optional[str] = None,
+    limit: int = 8,
+    org_slug: Optional[str] = None,
+) -> List[Dict[str, Any]]:
+    org = org_slug or _default_org()
+    engine, _ = _engine_for(org, collection)
+    engine.warmup()
+    return [_hit_to_dict(h) for h in engine.search(query, limit=limit)]
+
+
+async def qmd_vsearch(
+    query: str,
+    collection: Optional[str] = None,
+    limit: int = 8,
+    org_slug: Optional[str] = None,
+) -> List[Dict[str, Any]]:
+    org = org_slug or _default_org()
+    engine, _ = _engine_for(org, collection)
+    engine.warmup()
+    return [_hit_to_dict(h) for h in engine.vsearch(query, limit=limit)]
+
+
+async def qmd_query(
+    query: str,
+    collection: Optional[str] = None,
+    limit: int = 8,
+    use_rerank: bool = True,
+    org_slug: Optional[str] = None,
+) -> str:
+    org = org_slug or _default_org()
+    engine, _ = _engine_for(org, collection)
+    engine.warmup()
+    hits = engine.query(query, limit=limit, use_rerank=use_rerank)
+    if not hits:
+        return ""
+    parts = []
+    for h in hits:
+        parts.append(
+            f"[source: {h.file_path}:{h.start_line}-{h.end_line}] "
+            f"(score={h.score:.3f})\n{h.content}"
+        )
+    return "\n\n---\n\n".join(parts)
+
+
+async def qmd_get(docid: str, org_slug: Optional[str] = None) -> str:
+    """Возвращает полный документ по doc_id (ищет в любой коллекции org)."""
+    org = org_slug or _default_org()
+    from src.rag.qmd.collections import get_org_root
+    org_root = get_org_root(org)
+    if not org_root.exists():
+        return ""
+    import sqlite3
+    for db_path in org_root.rglob("index.sqlite"):
+        try:
+            conn = sqlite3.connect(str(db_path))
+            row = conn.execute(
+                "SELECT content FROM chunks WHERE doc_id = ? ORDER BY chunk_index",
+                (docid,),
+            ).fetchall()
+            conn.close()
+        except sqlite3.OperationalError:
+            continue
+        if row:
+            return "\n\n".join(r[0] for r in row)
+    return ""
+
+
+async def qmd_status(org_slug: Optional[str] = None) -> Dict[str, Any]:
+    """Агрегированный статус по всем коллекциям org."""
+    org = org_slug or _default_org()
+    from src.rag.qmd.collections import get_org_root
+    org_root = get_org_root(org)
+    result: Dict[str, Any] = {
+        "collections": [],
+        "total_chunks": 0,
+        "total_files": 0,
+        "engines": set(),
+        "fts": set(),
+        "embedding_loaded": False,
+    }
+    if not org_root.exists():
+        return _finalize_status(result)
+    for db_path in org_root.rglob("index.sqlite"):
+        try:
+            from src.rag.engine import Engine
+            eng = Engine.from_paths(db_path.parent)
+            st = eng.status()
+            result["collections"].append({"path": str(db_path.parent), **st})
+            result["total_chunks"] += st.get("chunks", 0)
+            result["total_files"] += st.get("files", 0)
+            result["engines"].add(st.get("engine", "?"))
+            result["fts"].add(st.get("fts", "?"))
+            result["embedding_loaded"] = result["embedding_loaded"] or st.get("embedding_loaded", False)
+            eng.close()
+        except Exception as exc:
+            logger.debug("status: failed to open %s: %s", db_path, exc)
+    return _finalize_status(result)
+
+
+def _finalize_status(result: Dict[str, Any]) -> Dict[str, Any]:
+    result["engines"] = sorted(result["engines"])
+    result["fts"] = sorted(result["fts"])
+    return result
+
+
+def _default_org() -> str:
+    return os.environ.get("DEFAULT_ORG_SLUG", "merakom")
diff --git a/src/rag/qmd/query.py b/src/rag/qmd/query.py
new file mode 100644
index 0000000..671f45d
--- /dev/null
+++ b/src/rag/qmd/query.py
@@ -0,0 +1,284 @@
+"""Поиск и чат-генерация ответов через native RAG engine.
+
+Совместимо с WebSocket-API: ``{"type": "context"|"chunk"|"done"}``.
+Поддерживает режимы bm25/vector/hybrid и маппинг legacy-имён.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+from typing import Any, AsyncIterator, Dict, List, Optional
+
+from openai import AsyncOpenAI
+
+from src.rag.qmd import mcp_client
+from src.rag.qmd.cache import SearchCache
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_CHAT_MAX_TOKENS = 8192
+DEFAULT_CHAT_MODEL = "deepseek-v4-flash-free"
+
+CHAT_MODES: Dict[str, Dict[str, str]] = {
+    "hybrid": {
+        "system": (
+            "Ты — ассистент по базе знаний строительной компании. "
+            "Отвечай на основе контекста. Указывай источники (файл, страница, лист). "
+            "Если данных нет — так и скажи."
+        ),
+        "instruction": "Ответь на вопрос:",
+    },
+    "compare": {
+        "system": (
+            "Ты — аналитик. Сопоставь информацию из разных документов в контексте. "
+            "Найди совпадения, расхождения, противоречия. Структурируй ответ по пунктам."
+        ),
+        "instruction": "Сопоставь и сравни информацию по запросу:",
+    },
+    "timeline": {
+        "system": (
+            "Ты — аналитик хронологии проекта. "
+            "Восстанови timeline событий, решений, изменений по датам из контекста. "
+            "Сортируй по времени, указывай источник каждого события."
+        ),
+        "instruction": "Построй хронологию / timeline по запросу:",
+    },
+}
+
+
+LEGACY_MODE_MAP: Dict[str, str] = {
+    "naive": "bm25",
+    "local": "vector",
+    "global": "bm25",
+    "hybrid": "hybrid",
+    "bm25": "bm25",
+    "vector": "vector",
+}
+
+_cache = SearchCache(ttl_seconds=60)
+
+
+def _message_text(message: Any) -> str:
+    content = (getattr(message, "content", None) or "").strip()
+    if content:
+        return content
+    reasoning = getattr(message, "reasoning_content", None) or getattr(message, "reasoning", None)
+    return (reasoning or "").strip()
+
+
+def _resolve_credentials() -> tuple[str, str]:
+    api_key = os.environ.get("OPENCODE_API_KEY", "")
+    base_url = os.environ.get("OPENCODE_URL", "https://opencode.ai/zen/v1")
+    return api_key, base_url
+
+
+def _build_chat_messages(
+    question: str,
+    context: str,
+    history: List[Dict[str, str]],
+    chat_mode: str,
+) -> List[Dict[str, str]]:
+    mode_cfg = CHAT_MODES.get(chat_mode, CHAT_MODES["hybrid"])
+    messages: List[Dict[str, str]] = [{"role": "system", "content": mode_cfg["system"]}]
+    for entry in history or []:
+        if "question" in entry and "answer" in entry:
+            messages.append({"role": "user", "content": entry["question"]})
+            messages.append({"role": "assistant", "content": entry["answer"]})
+    user_prompt = (
+        f"Контекст из базы знаний:\n---\n{context}\n---\n\n"
+        f"{mode_cfg['instruction']} {question}\n"
+    )
+    messages.append({"role": "user", "content": user_prompt})
+    return messages
+
+
+def _project_for_query(project_slug: Optional[str], mode: str) -> Optional[str]:
+    if mode == "global" or not project_slug or project_slug == "_global":
+        return None
+    return project_slug
+
+
+def map_retrieval_mode(mode: str) -> str:
+    return LEGACY_MODE_MAP.get(mode, "hybrid")
+
+
+async def qmd_query(
+    question: str,
+    org_slug: str,
+    project_slug: Optional[str] = None,
+    mode: str = "hybrid",
+    use_rerank: bool = True,
+) -> str:
+    """Совместимо со старым ``rag_chat`` — возвращает контекст как строку."""
+    retrieval_mode = map_retrieval_mode(mode)
+    target = _project_for_query(project_slug, retrieval_mode)
+    cache_key_args = (org_slug, target, retrieval_mode, use_rerank)
+    cached = _cache.get(question, cache_key_args)
+    if cached is not None:
+        return cached
+
+    # Резолвим каталог коллекции и используем engine напрямую (без
+    # round-trip через mcp_client, чтобы сохранить Path-идентичность).
+    from src.rag.qmd.collections import get_project_collection_dir, get_global_collection_dir
+    if target is None:
+        coll_dir = get_global_collection_dir(org_slug)
+    else:
+        coll_dir = get_project_collection_dir(org_slug, target)
+    coll_dir.mkdir(parents=True, exist_ok=True)
+    from src.rag.engine import get_or_create_engine
+    engine = get_or_create_engine(coll_dir)
+    engine.warmup()
+
+    if retrieval_mode == "bm25":
+        hits = engine.search(question, limit=20)
+        context = _format_bm25(_hits_to_dicts(hits))
+    elif retrieval_mode == "vector":
+        hits = engine.vsearch(question, limit=20)
+        context = _format_bm25(_hits_to_dicts(hits))
+    else:
+        hits = engine.query(question, limit=20, use_rerank=use_rerank)
+        context = _format_bm25(_hits_to_dicts(hits))
+
+    _cache.set(question, cache_key_args, context)
+    return context
+
+
+def _hits_to_dicts(hits) -> list:
+    return [
+        {
+            "file_path": h.file_path,
+            "docid": h.doc_id,
+            "score": h.score,
+            "snippet": h.snippet(400),
+            "text": h.content,
+            "source": h.source_annotation(),
+        }
+        for h in hits
+    ]
+
+
+def _format_bm25(hits: List[Dict[str, Any]]) -> str:
+    if not hits:
+        return ""
+    parts: List[str] = []
+    for hit in hits:
+        path = hit.get("file_path", hit.get("docid", "?"))
+        score = hit.get("score")
+        snippet = hit.get("snippet", hit.get("text", ""))
+        score_str = f" (score={score:.3f})" if isinstance(score, (int, float)) else ""
+        parts.append(f"[source: {path}]{score_str}\n{snippet}")
+    return "\n\n---\n\n".join(parts)
+
+
+async def _generate_answer(
+    question: str,
+    context: str,
+    history: List[Dict[str, str]],
+    chat_mode: str,
+    chat_model: str,
+    max_tokens: int,
+    stream: bool,
+):
+    api_key, base_url = _resolve_credentials()
+    if not api_key:
+        raise ValueError(
+            "OPENCODE_API_KEY не задан. Укажите rag.opencode_api_key в config.yaml "
+            "или переменную окружения OPENCODE_API_KEY."
+        )
+    client = AsyncOpenAI(base_url=base_url, api_key=api_key)
+    messages = _build_chat_messages(question, context, history, chat_mode)
+    return await client.chat.completions.create(
+        model=chat_model,
+        messages=messages,
+        temperature=0.4,
+        max_tokens=max_tokens,
+        stream=stream,
+    )
+
+
+async def qmd_chat(
+    question: str,
+    org_slug: str,
+    history: List[Dict[str, str]],
+    api_key: str = "",
+    project_slug: Optional[str] = None,
+    chat_model: str = DEFAULT_CHAT_MODEL,
+    mode: str = "hybrid",
+    chat_mode: str = "hybrid",
+    max_tokens: int = DEFAULT_CHAT_MAX_TOKENS,
+    use_rerank: bool = True,
+) -> Dict[str, Any]:
+    """Совместимо со старым ``rag_chat``: возвращает dict с ``answer``/``context``."""
+    if api_key:
+        os.environ["OPENCODE_API_KEY"] = api_key
+    context = await qmd_query(
+        question=question,
+        org_slug=org_slug,
+        project_slug=project_slug,
+        mode=mode,
+        use_rerank=use_rerank,
+    )
+    response = await _generate_answer(
+        question, context, history, chat_mode, chat_model, max_tokens, stream=False
+    )
+    answer = _message_text(response.choices[0].message)
+    return {
+        "answer": answer,
+        "context": context,
+        "project": project_slug,
+        "chat_mode": chat_mode,
+    }
+
+
+async def qmd_chat_stream(
+    question: str,
+    org_slug: str,
+    history: List[Dict[str, str]],
+    api_key: str = "",
+    project_slug: Optional[str] = None,
+    chat_model: str = DEFAULT_CHAT_MODEL,
+    mode: str = "hybrid",
+    chat_mode: str = "hybrid",
+    max_tokens: int = DEFAULT_CHAT_MAX_TOKENS,
+    use_rerank: bool = True,
+) -> AsyncIterator[Dict[str, Any]]:
+    """Стрим: context → chunk* → done (payload совместим с WebSocket)."""
+    if api_key:
+        os.environ["OPENCODE_API_KEY"] = api_key
+    context = await qmd_query(
+        question=question,
+        org_slug=org_slug,
+        project_slug=project_slug,
+        mode=mode,
+        use_rerank=use_rerank,
+    )
+    yield {
+        "type": "context",
+        "context": context,
+        "project": project_slug,
+        "chat_mode": chat_mode,
+    }
+    answer_parts: List[str] = []
+    try:
+        stream = await _generate_answer(
+            question, context, history, chat_mode, chat_model, max_tokens, stream=True
+        )
+        async for chunk in stream:
+            if not chunk.choices:
+                continue
+            delta = chunk.choices[0].delta
+            if delta and getattr(delta, "content", None):
+                answer_parts.append(delta.content)
+                yield {"type": "chunk", "content": delta.content}
+    except ValueError as exc:
+        yield {"type": "rag_error", "error": str(exc)}
+        return
+
+    yield {
+        "type": "done",
+        "answer": "".join(answer_parts),
+        "context": context,
+        "project": project_slug,
+        "chat_mode": chat_mode,
+    }
diff --git a/src/rag/query.py b/src/rag/query.py
deleted file mode 100644
index 5ecea1e..0000000
--- a/src/rag/query.py
+++ /dev/null
@@ -1,148 +0,0 @@
-"""Запросы к RAG и генерация ответов чат-бота."""
-
-from pathlib import Path
-from typing import Any, Dict, List, Optional
-
-from lightrag import QueryParam
-from openai import AsyncOpenAI
-
-from src.rag.indexer import get_global_rag, get_project_rag
-
-CHAT_MODES = {
-    "hybrid": {
-        "system": (
-            "Ты — ассистент по базе знаний строительной компании. "
-            "Отвечай на основе контекста. Указывай источники (файл, страница, лист). "
-            "Если данных нет — так и скажи."
-        ),
-        "instruction": "Ответь на вопрос:",
-    },
-    "compare": {
-        "system": (
-            "Ты — аналитик. Сопоставь информацию из разных документов в контексте. "
-            "Найди совпадения, расхождения, противоречия. Структурируй ответ по пунктам."
-        ),
-        "instruction": "Сопоставь и сравни информацию по запросу:",
-    },
-    "timeline": {
-        "system": (
-            "Ты — аналитик хронологии проекта. "
-            "Восстанови timeline событий, решений, изменений по датам из контекста. "
-            "Сортируй по времени, указывай источник каждого события."
-        ),
-        "instruction": "Построй хронологию / timeline по запросу:",
-    },
-}
-
-
-async def retrieve_context(
-    question: str,
-    working_dir_base: Path,
-    project_name: Optional[str] = None,
-    mode: str = "hybrid",
-    api_key: str = "",
-    base_url: str = "https://opencode.ai/zen/v1",
-    index_model: str = "mimo-v2.5-free",
-) -> str:
-    if project_name:
-        rag = await get_project_rag(
-            project_name,
-            working_dir_base,
-            model=index_model,
-            api_key=api_key,
-            base_url=base_url,
-        )
-    else:
-        rag = await get_global_rag(
-            working_dir_base,
-            model=index_model,
-            api_key=api_key,
-            base_url=base_url,
-        )
-
-    param = QueryParam(mode=mode if mode in ("naive", "local", "global", "hybrid") else "hybrid", only_need_context=True)
-    context = await rag.aquery(question, param=param)
-    return context if context else ""
-
-
-async def generate_chat_response(
-    question: str,
-    context: str,
-    history: List[Dict[str, str]],
-    api_key: str,
-    base_url: str = "https://opencode.ai/zen/v1",
-    model: str = "deepseek-v4-flash-free",
-    chat_mode: str = "hybrid",
-) -> str:
-    if not api_key:
-        raise ValueError(
-            "OPENCODE_API_KEY не задан. Укажите rag.opencode_api_key в config.yaml "
-            "или переменную окружения OPENCODE_API_KEY."
-        )
-    client = AsyncOpenAI(base_url=base_url, api_key=api_key)
-
-    mode_cfg = CHAT_MODES.get(chat_mode, CHAT_MODES["hybrid"])
-    messages = [{"role": "system", "content": mode_cfg["system"]}]
-
-    for h in history:
-        messages.append({"role": "user", "content": h["question"]})
-        messages.append({"role": "assistant", "content": h["answer"]})
-
-    user_prompt = f"""Контекст из базы знаний:
----
-{context}
----
-
-{mode_cfg["instruction"]} {question}
-"""
-    messages.append({"role": "user", "content": user_prompt})
-
-    response = await client.chat.completions.create(
-        model=model,
-        messages=messages,
-        temperature=0.4,
-        max_tokens=2048,
-    )
-    content = response.choices[0].message.content
-    return content if content is not None else ""
-
-
-async def rag_chat(
-    question: str,
-    working_dir_base: Path,
-    history: List[Dict[str, str]],
-    api_key: str,
-    project_name: Optional[str] = None,
-    base_url: str = "https://opencode.ai/zen/v1",
-    chat_model: str = "deepseek-v4-flash-free",
-    mode: str = "hybrid",
-    index_model: str = "mimo-v2.5-free",
-    chat_mode: str = "hybrid",
-) -> Dict[str, Any]:
-    retrieval_mode = mode if mode in ("naive", "local", "global", "hybrid") else "hybrid"
-    context = await retrieve_context(
-        question=question,
-        working_dir_base=working_dir_base,
-        project_name=project_name,
-        mode=retrieval_mode,
-        api_key=api_key,
-        base_url=base_url,
-        index_model=index_model,
-    )
-
-    answer = await generate_chat_response(
-        question=question,
-        context=context,
-        history=history,
-        api_key=api_key,
-        base_url=base_url,
-        model=chat_model,
-        chat_mode=chat_mode,
-    )
-
-    return {
-        "answer": answer,
-        "context": context,
-        "project": project_name,
-        "chat_mode": chat_mode,
-    }
diff --git a/tests/test_ingest.py b/tests/test_ingest.py
index f1a48b9..c747107 100644
--- a/tests/test_ingest.py
+++ b/tests/test_ingest.py
@@ -12,7 +12,14 @@ sys.path.insert(0, str(ROOT))
 from src.ingest.classify import _parse_json
 from src.ingest.formatter import format_index_document
 from src.ingest.models import NormalizedDocument, DocumentChunk
-from src.ingest.router import extract_document, is_audio_file, is_document_file, resolve_upload_kind
+from src.ingest.router import (
+    extract_document,
+    is_audio_file,
+    is_document_file,
+    is_extractable,
+    resolve_upload_kind,
+)
+from src.ingest.stub_writer import write_stub
 
 
 def test_is_audio_file():
@@ -85,6 +92,25 @@ def test_resolve_upload_kind():
         pass
 
 
+def test_is_extractable():
+    assert is_extractable("spec.pdf") is True
+    assert is_extractable("data.csv") is True
+    assert is_extractable("notes.md") is True
+    assert is_extractable("video.mp4") is False
+    assert is_extractable("archive.zip") is False
+
+
+def test_write_stub_for_binary():
+    with tempfile.TemporaryDirectory() as tmp:
+        mp4 = Path(tmp) / "movie.mp4"
+        mp4.write_bytes(b"\x00" * 8)
+        stub = write_stub(mp4, project="2026")
+        assert stub.exists()
+        text = stub.read_text(encoding="utf-8")
+        assert "kind: video" in text
+        assert "project: 2026" in text
+
+
 if __name__ == "__main__":
     class IngestTestCase(unittest.TestCase):
         def test_audio(self):
@@ -108,4 +134,10 @@ if __name__ == "__main__":
         def test_route(self):
             test_resolve_upload_kind()
 
+        def test_extractable(self):
+            test_is_extractable()
+
+        def test_stub(self):
+            test_write_stub_for_binary()
+
     unittest.main(verbosity=2)
diff --git a/tests/test_native_engine.py b/tests/test_native_engine.py
new file mode 100644
index 0000000..4598a83
--- /dev/null
+++ b/tests/test_native_engine.py
@@ -0,0 +1,177 @@
+"""Real-data tests for native RAG engine (no mocks, in-process).
+
+Использует sentence-transformers реальную модель (~50 MB скачивается при первом
+запуске). Время выполнения: ~10 сек (cold) + ~1 сек на warm-кейсы.
+"""
+
+import tempfile
+import unittest
+from pathlib import Path
+
+from src.rag.engine import Engine, get_or_create_engine, invalidate_engine
+from src.rag.engine.chunker import chunk_text
+from src.rag.engine.bm25 import bm25_search
+from src.rag.engine.vector import vector_search
+from src.rag.engine.hybrid import rrf_fuse
+
+
+SAMPLE_DOCS = {
+    "plan.md": (
+        "# План 3-го этажа\n\n"
+        "План 3-го этажа жилого дома. Оси: А, Б, В, Г. Размеры между осями А и Б: 5400 мм.\n"
+        "Квартиры: 301, 302, 303. Кухни объединены с гостиными. Санузлы раздельные.\n\n"
+        "## Отделка\n"
+        "Стены — штукатурка, покраска. Полы — ламинат. Потолки — гипсокартон.\n"
+    ),
+    "auth.md": (
+        "# Авторизация\n\n"
+        "Авторизация работает через JWT-токены с TTL 24 часа.\n"
+        "Refresh-токен живёт 30 дней. Логика валидации в middleware.\n"
+        "Сессии хранятся в Redis, ключ — sha256 от user_id + jti.\n"
+    ),
+    "schedule.md": (
+        "# График работ\n\n"
+        "Строительство начинается 1 июня 2026. Окончание — 30 ноября 2027.\n"
+        "Этапы: фундамент → стены → кровля → MEP → отделка.\n"
+    ),
+}
+
+
+class ChunkerTestCase(unittest.TestCase):
+    def test_chunks_short_text_returns_one(self):
+        chunks = chunk_text("Привет, мир.", max_chars=100)
+        self.assertEqual(len(chunks), 1)
+        self.assertIn("Привет", chunks[0].text)
+
+    def test_chunks_long_text_splits_with_overlap(self):
+        text = "Абзац.\n\n" * 200
+        chunks = chunk_text(text, max_chars=400, overlap=60)
+        self.assertGreater(len(chunks), 1)
+        for c in chunks:
+            self.assertLessEqual(len(c.text), 1000)  # с overlap
+
+    def test_empty_text_returns_empty(self):
+        self.assertEqual(chunk_text(""), [])
+        self.assertEqual(chunk_text("   \n\n  "), [])
+
+
+class EngineIngestTestCase(unittest.TestCase):
+    def setUp(self):
+        self._tmp = tempfile.TemporaryDirectory()
+        self.tmp = Path(self._tmp.name)
+        self.engines = []
+
+    def _make_engine(self, name: str = "coll1") -> Engine:
+        eng = Engine.from_paths(self.tmp / name)
+        self.engines.append(eng)
+        return eng
+
+    def tearDown(self):
+        for eng in self.engines:
+            eng.close()
+        for db_path in self.tmp.rglob("index.sqlite"):
+            invalidate_engine(db_path.parent)
+        import time
+        time.sleep(0.05)
+        try:
+            self._tmp.cleanup()
+        except (PermissionError, OSError):
+            pass
+
+    def test_index_text_returns_chunks_and_vectors(self):
+        eng = self._make_engine()
+        result = eng.index_text(SAMPLE_DOCS["plan.md"], source_path="plan.md")
+        self.assertFalse(result.skipped)
+        self.assertGreaterEqual(result.chunks_indexed, 1)
+        self.assertEqual(result.vectors_indexed, result.chunks_indexed)
+        status = eng.status()
+        self.assertGreaterEqual(status["chunks"], 1)
+        self.assertIn(status["engine"], ("sqlite-vec", "numpy"))
+
+    def test_index_text_is_idempotent(self):
+        eng = self._make_engine()
+        r1 = eng.index_text(SAMPLE_DOCS["plan.md"], source_path="plan.md")
+        r2 = eng.index_text(SAMPLE_DOCS["plan.md"], source_path="plan.md")
+        self.assertFalse(r1.skipped)
+        self.assertTrue(r2.skipped)
+
+    def test_index_file_change_detected(self):
+        eng = self._make_engine()
+        f = self.tmp / "x.md"
+        f.write_text("first version", encoding="utf-8")
+        r1 = eng.index_file(f)
+        self.assertFalse(r1.skipped)
+        f.write_text("second version with new content", encoding="utf-8")
+        r2 = eng.index_file(f)
+        self.assertFalse(r2.skipped)
+
+    def test_index_many_files(self):
+        eng = self._make_engine()
+        for name, text in SAMPLE_DOCS.items():
+            eng.index_text(text, source_path=name)
+        status = eng.status()
+        self.assertEqual(status["chunks"], 3)
+        self.assertEqual(status["files"], 3)
+
+
+class EngineSearchTestCase(unittest.TestCase):
+    """Реальные тесты с warm-embedding."""
+
+    @classmethod
+    def setUpClass(cls):
+        cls._tmp = tempfile.TemporaryDirectory()
+        cls.tmp = Path(cls._tmp.name)
+        cls.eng = Engine.from_paths(cls.tmp / "coll_search")
+        for name, text in SAMPLE_DOCS.items():
+            cls.eng.index_text(text, source_path=name)
+        cls.eng.warmup()
+
+    @classmethod
+    def tearDownClass(cls):
+        cls.eng.close()
+        invalidate_engine(cls.tmp / "coll_search")
+        import time
+        time.sleep(0.05)
+        try:
+            cls._tmp.cleanup()
+        except (PermissionError, OSError):
+            pass
+
+    def test_bm25_search_finds_keywords(self):
+        hits = self.eng.search("авторизация JWT")
+        self.assertGreater(len(hits), 0)
+        self.assertEqual(hits[0].file_path, "auth.md")
+
+    def test_bm25_search_empty_query(self):
+        self.assertEqual(self.eng.search(""), [])
+        self.assertEqual(self.eng.search("   "), [])
+
+    def test_vector_search_finds_semantic(self):
+        # "как устроена авторизация" — семантически близко к "авторизация"
+        hits = self.eng.vsearch("как устроена авторизация")
+        self.assertGreater(len(hits), 0)
+        # top hit должен быть auth.md или schedule.md (есть слово "логика")
+        paths = [h.file_path for h in hits]
+        self.assertTrue(any("auth" in p for p in paths))
+
+    def test_hybrid_query_uses_rrf(self):
+        hits = self.eng.query("квартиры 3 этаж", limit=3)
+        self.assertGreater(len(hits), 0)
+        self.assertEqual(hits[0].file_path, "plan.md")
+
+    def test_get_returns_full_document(self):
+        hits = self.eng.search("авторизация")
+        full = self.eng.get(hits[0].doc_id)
+        self.assertIn("JWT", full)
+        self.assertIn("TTL", full)
+
+    def test_status_reports_engine(self):
+        status = self.eng.status()
+        self.assertEqual(status["files"], 3)
+        self.assertGreaterEqual(status["chunks"], 3)
+        self.assertEqual(status["embedding_dim"], 384)
+        self.assertTrue(status["embedding_loaded"])
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/tests/test_native_engine_e2e.py b/tests/test_native_engine_e2e.py
new file mode 100644
index 0000000..d8f33a1
--- /dev/null
+++ b/tests/test_native_engine_e2e.py
@@ -0,0 +1,122 @@
+"""End-to-end тест: ingest → search → context → chat-stream.
+
+Не мокает движок — ставит заглушку только на OpenCode-клиент.
+"""
+
+import tempfile
+import unittest
+from pathlib import Path
+from unittest.mock import AsyncMock, MagicMock, patch
+
+from src.rag.qmd import indexer as qmd_indexer
+from src.rag.qmd import query as qmd_query
+
+
+SAMPLE_MEETING = (
+    "# Совещание 2026-06-10\n\n"
+    "Участники: Иванов, Петров, Сидорова.\n\n"
+    "## Повестка\n"
+    "1. Ход строительства 3-й очереди.\n"
+    "2. Авторизация подрядчиков в системе.\n"
+    "3. Сроки сдачи.\n\n"
+    "## Решения\n"
+    "- Завершить фундамент до 15 июля.\n"
+    "- Выдать JWT-токены подрядчикам.\n"
+    "- Срок сдачи — 30 ноября 2027.\n"
+)
+
+
+class E2ETestCase(unittest.IsolatedAsyncioTestCase):
+    async def asyncSetUp(self):
+        from src.rag.qmd.collections import get_project_collection_dir
+        self._tmp = tempfile.TemporaryDirectory()
+        self.tmp = Path(self._tmp.name)
+        self.coll_dir = get_project_collection_dir("merakom", "2026", self.tmp)
+        self.coll_dir.mkdir(parents=True, exist_ok=True)
+        self.engines = []
+
+    async def asyncTearDown(self):
+        from src.rag.engine import invalidate_engine
+        for eng in self.engines:
+            eng.close()
+        invalidate_engine(self.coll_dir)
+        import time
+        time.sleep(0.05)
+        try:
+            self._tmp.cleanup()
+        except (PermissionError, OSError):
+            pass
+
+    async def test_e2e_meeting_index_search_chat(self):
+        # 1. Ingest через Engine.index_file напрямую (в self.coll_dir).
+        from src.rag.engine import get_or_create_engine
+        eng = get_or_create_engine(self.coll_dir)
+        eng.warmup()
+        self.engines.append(eng)
+
+        body = self.tmp / "meeting.txt"
+        summary = self.tmp / "meeting_summary.md"
+        body.write_text(SAMPLE_MEETING, encoding="utf-8")
+        summary.write_text("# Краткое\nОбсуждали строительство и JWT.", encoding="utf-8")
+        r1 = eng.index_file(body)
+        r2 = eng.index_file(summary)
+        self.assertFalse(r1.skipped)
+        self.assertFalse(r2.skipped)
+
+        # 2. Search через тот же engine.
+        hits = eng.query("сдача объекта", limit=3, use_rerank=False)
+        self.assertGreater(len(hits), 0)
+        # top hit должен относиться к meeting.txt
+        self.assertTrue(any("meeting.txt" in h.file_path for h in hits))
+
+        # 3. Chat-stream с подменой OpenCode.
+        fake_chunks = [
+            MagicMock(choices=[MagicMock(delta=MagicMock(content="Сдача "))]),
+            MagicMock(choices=[MagicMock(delta=MagicMock(content="30 ноября "))]),
+            MagicMock(choices=[MagicMock(delta=MagicMock(content="2027."))]),
+        ]
+
+        class _FakeStream:
+            def __init__(self):
+                self._i = 0
+
+            def __aiter__(self):
+                return self
+
+            async def __anext__(self):
+                if self._i >= len(fake_chunks):
+                    raise StopAsyncIteration
+                item = fake_chunks[self._i]
+                self._i += 1
+                return item
+
+        async def _fake_create(*args, **kwargs):
+            if kwargs.get("stream"):
+                return _FakeStream()
+            return MagicMock(choices=[MagicMock(message=MagicMock(content="final"))])
+
+        with patch("src.rag.qmd.query.AsyncOpenAI") as fake_cls:
+            fake_instance = MagicMock()
+            fake_instance.chat = MagicMock()
+            fake_instance.chat.completions = MagicMock()
+            fake_instance.chat.completions.create = AsyncMock(side_effect=_fake_create)
+            fake_cls.return_value = fake_instance
+
+            events = []
+            async for ev in qmd_query.qmd_chat_stream(
+                question="Когда сдача?",
+                org_slug="merakom",
+                history=[],
+                project_slug="2026",
+                api_key="test-key",
+                use_rerank=False,
+            ):
+                events.append(ev)
+
+        types = [e["type"] for e in events]
+        self.assertEqual(types[0], "context")
+        self.assertIn("chunk", types)
+        self.assertEqual(types[-1], "done")
+        self.assertEqual(events[-1]["answer"], "Сдача 30 ноября 2027.")
+        # Контекст непустой и содержит source-annotation
+        self.assertIn("[source:", events[0]["context"])
diff --git a/tests/test_qmd_cache.py b/tests/test_qmd_cache.py
new file mode 100644
index 0000000..22895da
--- /dev/null
+++ b/tests/test_qmd_cache.py
@@ -0,0 +1,69 @@
+"""Tests for qmd search-result cache (TTL + mtime invalidation)."""
+
+import asyncio
+import tempfile
+import time
+import unittest
+from pathlib import Path
+from unittest.mock import patch
+
+from src.rag.qmd.cache import SearchCache
+
+
+class SearchCacheTestCase(unittest.IsolatedAsyncioTestCase):
+    async def test_get_set_roundtrip(self):
+        cache = SearchCache(ttl_seconds=60)
+        args = ("merakom", "2026", "bm25", True)
+        self.assertIsNone(cache.get("hello", args))
+        cache.set("hello", args, "ctx-blob")
+        self.assertEqual(cache.get("hello", args), "ctx-blob")
+
+    async def test_question_normalization(self):
+        cache = SearchCache(ttl_seconds=60)
+        args = ("merakom", "2026", "bm25", True)
+        cache.set("Hello World", args, "ctx")
+        self.assertEqual(cache.get("  hello world  ", args), "ctx")
+
+    async def test_different_args_yield_different_keys(self):
+        cache = SearchCache(ttl_seconds=60)
+        cache.set("q", ("o", "p1", "bm25", True), "ctx1")
+        cache.set("q", ("o", "p2", "bm25", True), "ctx2")
+        self.assertEqual(cache.get("q", ("o", "p1", "bm25", True)), "ctx1")
+        self.assertEqual(cache.get("q", ("o", "p2", "bm25", True)), "ctx2")
+
+    async def test_set_skips_empty_value(self):
+        cache = SearchCache(ttl_seconds=60)
+        cache.set("q", ("o", "p", "bm25", True), "")
+        self.assertIsNone(cache.get("q", ("o", "p", "bm25", True)))
+
+    async def test_ttl_expiry(self):
+        cache = SearchCache(ttl_seconds=0)
+        args = ("o", "p", "bm25", True)
+        with patch("src.rag.qmd.cache._index_mtime", return_value=0.0):
+            cache.set("q", args, "ctx")
+            await asyncio.sleep(0.05)
+            self.assertIsNone(cache.get("q", args))
+
+    async def test_mtime_invalidation(self):
+        cache = SearchCache(ttl_seconds=60)
+        args = ("o", "p", "bm25", True)
+        with patch("src.rag.qmd.cache._index_mtime", return_value=10.0):
+            cache.set("q", args, "ctx")
+        with patch("src.rag.qmd.cache._index_mtime", return_value=20.0):
+            self.assertIsNone(cache.get("q", args))
+
+    async def test_clear(self):
+        cache = SearchCache(ttl_seconds=60)
+        cache.set("q", ("o", "p", "bm25", True), "ctx")
+        cache.clear()
+        self.assertIsNone(cache.get("q", ("o", "p", "bm25", True)))
+
+    async def test_stats(self):
+        cache = SearchCache(ttl_seconds=42)
+        stats = cache.stats()
+        self.assertEqual(stats["entries"], 0)
+        self.assertEqual(stats["ttl_seconds"], 42)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/tests/test_rag_stream_ws.py b/tests/test_rag_stream_ws.py
new file mode 100644
index 0000000..bc05bd0
--- /dev/null
+++ b/tests/test_rag_stream_ws.py
@@ -0,0 +1,59 @@
+"""Smoke-test WebSocket RAG streaming against running service."""
+
+import asyncio
+import json
+import sys
+import urllib.error
+import urllib.request
+
+import websockets
+
+
+async def main() -> int:
+    login_body = json.dumps({"username": "admin", "password": "admin123"}).encode()
+    login_req = urllib.request.Request(
+        "http://127.0.0.1:8000/api/auth/login",
+        data=login_body,
+        headers={"Content-Type": "application/json"},
+        method="POST",
+    )
+    try:
+        with urllib.request.urlopen(login_req, timeout=10) as resp:
+            token = json.loads(resp.read().decode())["access_token"]
+    except urllib.error.HTTPError as e:
+        print(f"Login failed: {e.read().decode()}", file=sys.stderr)
+        return 1
+
+    uri = f"ws://127.0.0.1:8000/ws?token={token}"
+    types: list[str] = []
+    chunks = 0
+
+    async with websockets.connect(uri, open_timeout=10) as ws:
+        await ws.send(json.dumps({
+            "action": "rag_query_global",
+            "question": "Кратко: что есть в базе знаний?",
+            "history": [],
+            "chat_mode": "hybrid",
+            "retrieval_mode": "naive",
+        }))
+
+        while len(types) < 200:
+            raw = await asyncio.wait_for(ws.recv(), timeout=180)
+            msg = json.loads(raw)
+            msg_type = msg.get("type")
+            types.append(msg_type)
+            if msg_type == "rag_chunk":
+                chunks += 1
+            if msg_type in ("rag_response", "rag_error"):
+                if msg_type == "rag_error":
+                    print(f"RAG error: {msg.get('error')}", file=sys.stderr)
+                    return 1
+                print(f"OK: types={types[:5]}... chunks={chunks} answer_len={len(msg.get('answer', ''))}")
+                return 0
+
+    print("No terminal message received", file=sys.stderr)
+    return 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(asyncio.run(main()))
diff --git a/tests/test_stub_writer.py b/tests/test_stub_writer.py
new file mode 100644
index 0000000..ae379fa
--- /dev/null
+++ b/tests/test_stub_writer.py
@@ -0,0 +1,66 @@
+"""Tests for stub_writer (binary file -> .md frontmatter stub)."""
+
+import tempfile
+import unittest
+from pathlib import Path
+
+from src.ingest.stub_writer import write_stub, _infer_kind
+
+
+class StubWriterTestCase(unittest.TestCase):
+    def test_infer_kind_video(self):
+        self.assertEqual(_infer_kind(Path("meeting.mp4")), "video")
+        self.assertEqual(_infer_kind(Path("recording.MKV")), "video")
+
+    def test_infer_kind_audio(self):
+        self.assertEqual(_infer_kind(Path("track.wav")), "audio")
+
+    def test_infer_kind_image(self):
+        self.assertEqual(_infer_kind(Path("photo.png")), "image")
+
+    def test_infer_kind_archive(self):
+        self.assertEqual(_infer_kind(Path("backup.zip")), "archive")
+
+    def test_infer_kind_other(self):
+        self.assertEqual(_infer_kind(Path("unknown.xyz")), "other")
+
+    def test_write_stub_creates_md(self):
+        with tempfile.TemporaryDirectory() as tmp:
+            mp4 = Path(tmp) / "video.mp4"
+            mp4.write_bytes(b"fake-mp4-content")
+            stub = write_stub(mp4, project="2026")
+            self.assertEqual(stub, mp4.with_suffix(".mp4.md"))
+            self.assertTrue(stub.exists())
+            text = stub.read_text(encoding="utf-8")
+            self.assertIn("source: video.mp4", text)
+            self.assertIn("kind: video", text)
+            self.assertIn("project: 2026", text)
+            self.assertIn("size: 16", text)
+            self.assertIn("Бинарный файл", text)
+
+    def test_write_stub_overwrites(self):
+        with tempfile.TemporaryDirectory() as tmp:
+            mp4 = Path(tmp) / "video.mp4"
+            mp4.write_bytes(b"a" * 10)
+            stub = write_stub(mp4, project="p1")
+            stub.write_text("OLD", encoding="utf-8")
+            stub2 = write_stub(mp4, project="p1")
+            self.assertEqual(stub, stub2)
+            self.assertNotIn("OLD", stub.read_text(encoding="utf-8"))
+
+    def test_write_stub_missing_file_raises(self):
+        with tempfile.TemporaryDirectory() as tmp:
+            missing = Path(tmp) / "nope.mp4"
+            with self.assertRaises(FileNotFoundError):
+                write_stub(missing, project="x")
+
+    def test_explicit_kind_overrides_inference(self):
+        with tempfile.TemporaryDirectory() as tmp:
+            f = Path(tmp) / "weird.bin"
+            f.write_bytes(b"x")
+            stub = write_stub(f, project="p", kind="custom")
+            self.assertIn("kind: custom", stub.read_text(encoding="utf-8"))
+
+
+if __name__ == "__main__":
+    unittest.main()