179 questions
No questions match those filters.
Why is pure vector search not enough for an enterprise corpus?
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansDense retrieval encodes meaning, and it is systematically weak on exact tokens — error codes, field names, clause numbers, account identifiers. Those are often the highest-signal terms in an enterprise document, especially in regulated industries where the answer hinges on a specific clause number or field name, not on paraphrased meaning.
The fix is hybrid retrieval: run a lexical method (BM25) and a dense vector search in parallel, then fuse the two ranked lists with Reciprocal Rank Fusion. RRF combines by rank rather than by raw score, which sidesteps the problem that BM25 scores and cosine-similarity scores live on incomparable scales and can’t just be averaged.