Amey-Thakur Hybrid search
Keyword search nails exact terms and fails on paraphrase; vector search does the opposite and confidently misses a product code. Hybrid keeps both and spends its complexity on combining two rankings that are not directly comparable.
Method
- Confirm you need both. If queries are mostly exact terms, keyword alone is simpler and cheaper. Hybrid is justified by real paraphrase traffic, not by novelty.
- Retrieve independently, then fuse. Run both retrievers and combine the ranked lists rather than trying to blend raw scores, which are on incompatible scales.
- Prefer rank-based fusion as the default. Reciprocal rank fusion needs no calibration and is hard to beat without tuning, which makes it the right starting point.
- Keep exact matching authoritative. Identifiers, codes, and quoted phrases should not be outranked by something semantically nearby (see typo-tolerance).
- Choose and freeze an embedding model per index. Mixing embeddings from different models within one index makes distances meaningless, and changing models means reindexing (see embeddings-selection).
- Rerank the fused head when quality justifies the cost. A cross encoder over the top results improves ordering considerably at real latency cost (see retrieval-reranking).
- Evaluate against the same judged set as keyword. Hybrid must be shown better on your queries, not assumed better (see relevance-tuning).
Boundaries
- Hybrid adds an embedding pipeline, a vector store, and reindexing obligations, which is real operational weight.
- Vector recall degrades on very short queries where there is little semantic signal to work with.
- Semantic similarity is not relevance; two documents can be close in meaning and useless for the query.