Contextual retrieval

processopen source

Prepend chunk-specific context before embedding and BM25 indexing. Anthropic measured 49% fewer failed retrievals, and 67% fewer with reranking.

Each chunk gets a short, model-written explanation of where it sits in the source document before it's embedded (Contextual Embeddings) and indexed for BM25 (Contextual BM25). Combine with a reranker for the largest gain.

Signals

No signals collected yet.

How it stacks up

Contextual retrieval next to similar tools, using only numbers both have. Bigger isn't always better: read each number's meaning first.

No fair match yet: no similar tool shares at least two of Contextual retrieval's numbers. We only compare like-for-like numbers.

Relationship map