Guide

A fastrag pipeline has four stages: ingest a source into documents, index them into a vector collection, retrieve the ones relevant to a question, and query an answer out of them. Benchmark measures the whole chain against a labelled dataset, and every stage is built from typed, swappable pieces you construct and pass in — nothing is configured by string or discovered by import.

Read Concepts first regardless of what you came here to build — it names the four kinds of thing every other page assumes you already recognise. After that, the pages below follow the pipeline in order, and each stands on its own once you know the vocabulary:

• Install
• Quickstart
• Concepts
• Ingestion
• Structure-aware ingestion
• Caching
• Indexing
• Retrieval
• Querying
• Benchmarking
• Configuration
• Customizing
• Examples
• Serving

• New to the package? Install then Quickstart gets a real question answered in about 40 lines.

• Building an ingestion pipeline? Ingestion → Caching → Indexing, in that order — each stage’s cache and identity rules depend on the one before it.

• Already indexed, now querying? Retrieval → Querying.

• Comparing configurations? Benchmarking turns “I think 500 is better than 800” into a table.

• Exposing this over HTTP? Serving is a separate audience — it assumes no Python on the other end at all.

• None of the built-ins fit? Customizing is the extension story for every kind in Concepts.

See also

API reference for the generated API; Contributing if the piece you need does not exist yet and you are thinking about adding it.