Installation¶

AutoRAG is distributed from GitHub, not PyPI. The base install carries only typer, pydantic, langchain-core, and langchain-ollama — anything heavier (Whisper, pyannote, Chroma, UMAP, yt-dlp, FastAPI) is gated behind an install extra so the import from autorag import AutoRAG stays fast.

Choosing your extras¶

Pick the smallest set that unlocks the methods you intend to call:

Extra	Adds	Use when you want…
`audio`	whisperx, torch, imageio-ffmpeg	…to call `AutoRAG.transcribe` or `build_agent`.
`diarize`	pyannote.audio, huggingface-hub	…speaker labels on every word (combine with `audio`).
`youtube`	yt-dlp	…to pass a YouTube URL as `file=`.
`rag`	chromadb, umap-learn, scikit-learn, pydantic_sqlite, numpy	…`persist_transcription`, `persist_topics`, document ingest/query, or the `/viz` page.
`server`	fastapi, uvicorn[standard]	…`autorag serve` or the HTTP API.
`broker`	pika	…the async, RabbitMQ-driven pipeline (`autorag.services`) — the `/jobs/*` endpoints and the GPU/IO workers. The synchronous endpoints and CLI never need it.
`all`	everything above	…the full local-dev stack.

[diarize] rides on top of [audio] — pyannote needs the same torch + ffmpeg stack. Install them together.

Installing from a tagged release¶

# Audio → topics agent only
pip install "autorag[audio,diarize] @ git+https://github.com/AutoLogger/AutoRAG@v0.7.0"

# Audio + YouTube URL support
pip install "autorag[audio,diarize,youtube] @ git+https://github.com/AutoLogger/AutoRAG@v0.7.0"

# Full stack (audio, diarize, rag, server, youtube)
pip install "autorag[all] @ git+https://github.com/AutoLogger/AutoRAG@v0.7.0"

Calling a method whose extra is missing raises MissingExtraError with a hint naming the pip install command that fixes it.

Local development¶

Inside a checkout of the repository, AutoRAG uses uv (not pip):

uv sync --all-extras       # install everything
uv sync --group docs       # add the docs build deps
uv run pytest              # run the test suite
uv run autorag --help      # invoke the CLI

See Packaging and distribution for the release flow.

Required external services¶

AutoRAG calls Ollama for LLM chat and embeddings; you need a local (or remote) Ollama running before invoking generate_topics, query, or the /viz page. Diarization needs an HF token; the optional async pipeline needs a broker:

Ollama — AUTORAG_OLLAMA_BASE_URL (default http://localhost:11434) and AUTORAG_EMBED_MODEL (default nomic-embed-text). LLM chat uses gemma4:latest by default (a thinking-capable model; the agent disables thinking by default — see the audio-pipeline-design internals page).
Hugging Face — HF_TOKEN is required for the gated pyannote/speaker-diarization-3.1 model. Without it, every word is labelled "0".
RabbitMQ — only for the async [broker] path; AUTORAG_BROKER_URL (default amqp://localhost:5672). The synchronous SDK / CLI / API never need it. The whole async stack (RabbitMQ + Ollama + workers + docker-socket-proxy) is brought up on the host with one command, ./scripts/stack.sh up (.env is optional — only HF_TOKEN for diarization), or point AUTORAG_BROKER_URL at an external broker. See Running the HTTP server for the full deployment.