Someone searches for a specific error code, a part number, or an unusual surname. Dense embedding search — which matches on meaning — has no strong signal for a string it has essentially never seen, and returns documents that are topically adjacent and do not contain it. Keyword search finds it immediately, because the string is either present or absent.
That failure mode is not an edge case. It is a large share of real search traffic in any technical or commercial corpus.
What changed in 2026
- Hybrid became the default recommendation. Guidance shifted from dense-only to hybrid as the standard starting architecture rather than an optimization.
- Vector databases added keyword support natively. Running both retrieval methods in one system, rather than maintaining two, became common.
- Rank fusion consolidated. Reciprocal rank fusion emerged as the robust default for combining results, displacing score-weighted approaches.
- Learned sparse retrieval matured. Methods producing sparse representations with learned term weights offered a middle path, covered in sparse embeddings explained.
Where each method wins
| Query type |
Dense embeddings |
Keyword search |
| Conceptual question, different wording |
Strong |
Weak |
| Synonyms and paraphrase |
Strong |
Weak |
| Exact identifier or error code |
Weak |
Strong |
| Rare proper noun |
Weak |
Strong |
| Acronym or product name |
Variable |
Strong |
| Cross-lingual matching |
Strong with the right model |
Weak |
| Very short queries |
Variable |
Often strong |
| Long descriptive queries |
Strong |
Weaker |
The complementarity is nearly ideal. The queries dense retrieval handles badly are precisely the ones keyword search handles trivially, and vice versa. That is why combining them produces a larger improvement than either method's own tuning typically delivers.
How fusion works
The naive approach — normalize both scores and add them — does not work well, because the two methods produce scores on incomparable scales with different distributions. A tuned weighting on one corpus fails on another.
Reciprocal rank fusion sidesteps this by ignoring scores entirely and using only rank position. Each document gets a contribution based on where it placed in each result list, and those contributions sum. A document ranked highly by both methods rises to the top; a document ranked highly by one and absent from the other still scores reasonably.
This is robust because rank is comparable across methods in a way raw score is not, and it requires almost no tuning. It is the sensible default, and the small constant it uses to damp the contribution of low ranks is one of the few parameters worth touching.
Add reranking on top when quality matters more than latency. Retrieve a broader candidate set with hybrid search, then rerank the top results with a cross-encoder that scores query and document together — the approach in cross-encoder reranking.
Common mistakes
- Dense-only retrieval. Fails on exact matches in a way users notice immediately.
- Score normalization for fusion. Fragile across corpora; use rank fusion.
- Retrieving the same count from both methods and merging naively. Fusion handles the combination; retrieve generously from each.
- Not measuring by query type. Aggregate recall hides that identifier queries are failing.
- Keyword-only in a conversational product. Users paraphrase, and lexical matching misses it.
FAQ
Do I need two separate systems?
Increasingly not. Most vector databases and search engines now support both retrieval modes in one system with fusion built in.
How much does hybrid improve recall?
It varies by corpus and query mix, and it is typically one of the larger single improvements available in a retrieval pipeline, particularly where identifiers and proper nouns are common.
Is BM25 specifically required?
It is the well-tested default for lexical scoring. Alternatives exist, and the important part is having a lexical channel at all.
Does hybrid slow retrieval down?
Two searches instead of one, usually run in parallel. The added latency is small relative to generation.
Where to go next
For the middle ground, read sparse embeddings explained. For improving the final ordering, cross-encoder reranking, and for measuring the gain, RAG evaluation metrics.