Introducing Contextual Retrieval

A passage can become difficult to find after it is separated from its document. A sentence about revenue growth, for example, may omit the company and reporting period. Anthropic's Contextual Retrieval method generates a short explanation of each passage using the whole document, then attaches that explanation before indexing it for both meaning-based and keyword searches.

In the company's reported evaluation, combining both forms of contextual indexing reduced the failure rate for retrieval among the top 20 results from 5.7% to 2.9%. That measures whether supporting material was found. It does not establish an equivalent improvement in the accuracy of answers generated from that material.

“This method can reduce the number of failed retrievals by 49% and, when combined with reranking, by 67%.”

“Contextual Retrieval solves this problem by prepending chunk-specific explanatory context to each chunk before embedding (“Contextual Embeddings”) and creating the BM25 index (“Contextual BM25”).”

“Contextual Embeddings reduced the top-20-chunk retrieval failure rate by 35% (5.7% → 3.7%).”

“the one-time cost to generate contextualized chunks is $1.02 per million document tokens”