AutoRAG documentation

AutoRAG is an audio-to-topics agent and document retrieval toolkit. It ships as:

• an importable Python SDK (from autorag import AutoRAG),

• a autorag command-line tool,

• and a FastAPI server (autorag serve) with an interactive /viz WebGL page for inspecting topic embeddings.

It is built around a single autorag.core.AutoRAG facade with flat methods. Heavy dependencies (Whisper, pyannote, Chroma, UMAP, yt-dlp) are gated behind install extras so the base import remains lightweight.

Getting started

• Installation
  • Choosing your extras
  • Installing from a tagged release
  • Local development
  • Required external services
• Quickstart
  • From the CLI
  • From the Python SDK
  • From the HTTP API
  • Where to go next

User guide

• Transcription and topic extraction
  • Transcribe a local file
  • Extract topics
  • Persist
  • Cached, dependency-free reads
• YouTube URL inputs
  • How it works
  • What gets propagated
  • Canonical URL form
• Document RAG
  • Configuration
  • Status
• Visualization (/viz)
  • Running it
  • The pipeline behind the page
  • Endpoints
• Running the HTTP server
  • Endpoints
  • Calling the API
  • Async job pipeline
  • Mounting the app yourself
• CLI reference
  • autorag transcribe
  • autorag generate-topics
  • autorag blocks
  • autorag ingest
  • autorag query
  • autorag serve
  • autorag jobs

Internals

• Architecture
  • Why this matters
• Extras model
  • Methods → extras
  • Extras → modules
  • The contract
• Audio pipeline design
  • Stages
  • Whisper backend
  • Diarization
  • LLM model residency
  • Why split boundaries from summaries
• Async pipeline and deployment
  • Topology
  • Workers
  • Job state and the DB-path contract
  • Deployment
• Ollama tuning
  • Tuning contract (single source of truth)
  • OLLAMA_NUM_PARALLEL
  • OLLAMA_FLASH_ATTENTION and OLLAMA_MULTIUSER_CACHE
  • Per-slot KV-cache sizing
  • Model residency during a run
  • Resolving the Ollama URL
• Frontend (/viz)
  • Layout
  • Build flow
  • FastAPI wiring
  • Built bundle is committed
• Packaging and distribution
  • Releasing a new version
  • CI
  • Docs build
  • Third-party stubs
  • Pylance / Pyright
• Stress testing the async pipeline
  • Scenarios
  • Prerequisites
  • Running
  • Fixture staging
  • Artifact layout
  • A/B procedure (manual)
• Observability (autorag.otel)
  • Quick start
  • Span catalog
  • Metric catalog
  • Reading the per-LLM-call span (autorag.llm.call)
  • Service-name initialisation
  • Settings

API reference

• API reference
  • SDK facade (autorag.core)
  • CLI (autorag.cli)
  • HTTP server (autorag.api)
  • Audio→topics agent (autorag.agent)
  • Audio pipeline
  • RAG pipeline
  • Persistence
  • Visualization
  • Types, schemas, config, formatting

Project

• Changelog

Indices and tables

• Index

• Module Index

• Search Page