Contextualized Search

Contextualized search (Anthropic) addresses loss of context during chunking: before indexing, the LLM generates a prefix summary for each chunk (“excerpt from ACME report Q2 2025, key metrics section”), the prefix is joined with the text and indexed. Simultaneously boosts both sparse retrieval (exact terms in the prefix) and dense (semantic background). Reduces failed searches by 49% over BM25 and 67% over a reranker. Important to not confuse: this happens at indexing time and adds context; context compression happens at retrieval time and removes redundancy.

Related: [Document Chunking], [Hybrid Search], [Three-Level Memory Ranking]