PRODUCTION SCENARIO
An aerospace client searches 500,000 maintenance manuals. Technicians query exact part serial numbers such as AN960-JD10 and conceptual symptoms such as hydraulic fluid cavitation during high-altitude descent. Pure dense vector retrieval answers the symptom queries but misses the serial numbers.
Which change recovers exact-identifier matching without losing symptom search?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Add a BM25 index beside the vectors and fuse the two ranked lists
Dense embeddings generalise over meaning and lose rare alphanumeric tokens, while BM25 matches the exact string, so fusing the two ranked lists keeps each strength. A 3,072-dimension model changes the vector's width, not how little signal a one-off serial number carries. Fine-tuning covers only the identifiers present at training time and must be redone for every new part. Reranking reorders what the dense stage returned, and the identifier chunk was never among it.
Official Source: https://www.anthropic.com/engineering/contextual-retrieval ↗