Hybrid Search: Combining BM25 Keyword Scores With Vector Similarity
How hybrid search blends BM25 keyword ranking with dense vector similarity to catch exact terms and paraphrase alike, plus fusion techniques.
Published on • November 22, 2026
AI Assistant

Introduction
Keyword search finds ERROR_429 flawlessly and misses “rate limit” paraphrases. Vector search does the reverse. Hybrid search runs both and fuses the rankings — which is why it has become the default retrieval mode for serious search products and RAG pipelines.
Why Neither Alone Is Enough
| Query type | BM25 | Vector |
|---|---|---|
| ”how to reset my password” vs doc “credential recovery procedure” | ❌ no overlap | ✅ matches |
| ”ECONNRESET” error code | ✅ exact hit | ❌ diluted in semantic space |
| Product SKU “AX-9912-B” | ✅ exact | ❌ often missing |
| ”movies like Inception but funnier” | ❌ | ✅ captures intent |
Dense vectors generalise but blur rare tokens; BM25 is precise but brittle to paraphrase. Their failure modes are almost complementary — the definition of a good ensemble.
The Hybrid Architecture
┌─ BM25 (inverted index) ──── top-k keyword ranks ─┐
query ──────────┤ ├─ fusion → top-k
└─ embed → ANN index ──────── top-k vector ranks ──┘
- The query hits both retrieval legs in parallel.
- Each returns its own ranked top-k (often overfetch to k×5–k×10).
- A fusion method merges the two rankings into one.
Most systems (Elasticsearch, OpenSearch, Weaviate, Qdrant, Vespa, Pinecone) now ship this natively; LangChain-style RAG stacks compose it from retrievers.
Fusion Techniques
Score normalisation + weighted sum
Rescale each leg’s scores (min-max or z-score) to a common range, then score = α·bm25 + (1−α)·vector.
- ✅ Intuitive, tunable with one α
- ❌ score distributions differ per query; normalisation is fiddly; BM25 is unbounded while cosine is [-1,1]
Reciprocal rank fusion (RRF)
Ignore raw scores; use ranks only:
RRF(d) = Σ 1 / (rank_leg(d) + c) # c ≈ 60
- ✅ No score calibration needed — robust default
- ✅ Works even when the two legs’ score scales are incomparable
- (RRF gets its own dedicated treatment in this series — see topic #21)
Learn-to-rank / rerank stage
Fuse candidates, then pass the union to a cross-encoder reranker that scores query–document pairs directly. This is the highest-quality pattern: hybrid recall for candidates, reranker precision for final ordering.
Practical Tuning
- Overfetch both legs (k×10) so fusion has material to work with — fusing only top-5 too early starves the blend.
- Set α by query type: exact-ID-heavy corpora lean BM25 (α up); paraphrase-heavy lean vector. Some systems switch legs per query via a classifier.
- Watch tokenisation for rare identifiers — ensure BM25 indexes SKUs, error codes, and part numbers as tokens (or add a keyword subfield).
- Deduplicate — the same chunk often appears in both legs; fuse by document ID.
- Evaluate end-to-end: hybrid should beat both single legs on your eval set, or it’s complexity for nothing.
Hybrid in RAG
In retrieval-augmented generation, hybrid search reduces the two classic failure modes: missing context (model hallucinates because retrieval failed) and wrong context (irrelevant chunks poisoning the prompt). Combined with metadata filters and a reranker, it is the current best-practice retrieval stack.
Conclusion
Hybrid search is not “two searches, pick one” — it is a fusion problem: run BM25 for exactness, vectors for meaning, then blend ranks (RRF is the safe default) and optionally rerank. When the two legs’ weaknesses are each other’s strengths, the ensemble wins.