Extras model

AutoRAG’s install extras gate which methods you can call. The base install only carries typer + pydantic + langchain-{core,ollama}.

Methods → extras

Method	Extras needed	Purpose
AutoRAG.transcribe	[audio,diarize] (+ [youtube] for URLs)	Whisper + diarization → WordSpan list. file= is a local path or a YouTube URL.
AutoRAG.generate_topics	[audio,diarize]	LLM topic extraction on a pre-computed transcript → TopicTree.
AutoRAG.build_agent	[audio,diarize]	The combined Whisper + topics Runnable[Path | str, TranscriptionResult].
AutoRAG.transcribe_blocks	[rag] on cache hit; [audio,diarize] (+ [youtube]) on miss	N-second time-bucketed transcript view. Cache reads need only [rag].
AutoRAG.persist_transcription	[rag]	Write clip row + word spans to SQLite.
AutoRAG.persist_topics	[rag]	Persist topic tree + embed topic titles into Chroma.
AutoRAG.ingest	base	Document RAG: load → chunk → embed → store.
AutoRAG.query	base	Retrieve + generate over the ingested corpus.

Extras → modules

Extra	Modules that import it	Adds
audio	autorag.whisper_runner, autorag.agent (whisper)	whisperx, torch, imageio-ffmpeg
diarize	autorag.diarize	pyannote.audio, huggingface-hub
youtube	autorag.audio_source (lazy in _download_youtube_audio)	yt-dlp
rag	autorag.chroma_store, autorag.db, autorag.viz, autorag.topic_cluster	chromadb, umap-learn, scikit-learn, numpy, pydantic_sqlite
server	autorag.api	fastapi, uvicorn[standard]
broker	autorag.services.broker (lazy pika)	pika
all	—	union of the above

The contract

A new method on AutoRAG must decide up-front which extra(s) it needs and follow the existing pattern: do the heavy import inside the method body, catch ModuleNotFoundError, and re-raise via autorag.errors._missing_extra() with the extras string. Heavy imports at module top in any of core.py, embed.py, __init__.py, store.py, or audio_source.py will fail the CI test-base job.

MissingExtraError is a subclass of ImportError, so callers that want a single except for “AutoRAG isn’t fully installed” can catch ImportError.