PRODUCTION SCENARIO
A bank's RAG index holds 900 quarterly filings split at section headings into 600-token chunks. One chunk reads 'Net interest margin fell 20 basis points' with no issuer or quarter, so questions about a named bank's margin pull the wrong filing.
Which ingestion change targets the missing-context failure?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Prepend a model-written, chunk-specific context drawn from the whole filing
Contextual retrieval has a model write a 50-to-100-token chunk-specific context from the whole filing and prepends it before indexing, which is the only change here that puts the issuer and quarter into the chunk. Anthropic measured a 35 percent cut in top-20 retrieval failures from contextual embeddings alone, 49 percent once contextual BM25 is added. A keyword index, a 3,072-dimension model and a 1,200-token chunk all still read text that names neither.
Official Source: https://www.anthropic.com/engineering/contextual-retrieval ↗