Why it matters

Contextual retrieval is proven improvement. Understanding shapes modern RAG.

Advertisement

The architecture

LLM: summarize doc context for each chunk.

Prepend to chunk.

Embed enriched chunk.

Contextual retrievalChunkraw textLLM adds contextdoc summaryEnriched chunkfor embeddingOne-time cost per chunk; big retrieval quality gain
Contextual.
Advertisement

How it works end to end

Anthropic paper 2024.

50-100 tokens of context per chunk.

One-time preprocessing.

Combines well with hybrid + rerank.