Hybrid search: the best of keywords and meaning
Hybrid search runs a keyword search and a semantic search side by side, then merges the two rankings. It finds both the documents that contain the exact term and the ones that say the same thing in other words.
Data as of October 22, 2026
Key figures
| Level | Advanced |
|---|
Definition
Put simply: you query two engines at once, one looking for exact words and one looking for meaning, and you combine their answers. Keyword search (usually the BM25 algorithm) shines on rare terms and identifiers. Semantic search compares embeddings, vectors that capture meaning, and catches paraphrases. Each has blind spots the other covers.
How the results are merged
The two scores are not directly comparable: a BM25 score and a cosine similarity live on different scales. Two families of fusion dominate:
- Rank-based fusion: the best known is Reciprocal Rank Fusion (RRF), described by Cormack and colleagues in 2009. For each list, a document gets 1 / (k + rank), and the contributions are summed. In the Elasticsearch documentation, k defaults to 60. The advantage: no score calibration needed.
- Normalized score fusion: each score is rescaled to a common range before weighting. Weaviate offers both approaches, with an alpha parameter to balance vectors and keywords.
A concrete example
A printer manufacturer’s support desk receives: “My printer shows E-4012 and won’t turn on.” Semantic search finds articles about “power issues” but ranks the article that mentions code E-4012 poorly. BM25 finds that article but misses a guide titled “device won’t start.” Fusion puts both at the top, so the model in a RAG pipeline gets the right context.
Why it matters
The BEIR benchmark (2021) showed that BM25 remains a robust baseline against dense retrieval models across varied domains. And in its Contextual Retrieval write-up (September 2024), Anthropic reports that combining embeddings with BM25 reduced retrieval failures further than embeddings alone on its test sets. Final quality also depends on how documents are chunked and on an optional reranking step.
Common misconceptions
- “Embeddings made keywords obsolete.” For codes, names and exact references, the opposite is often true.
- “Hybrid means twice as slow.” Both queries run in parallel; in practice the overhead is usually small compared with the model call.
- “One setting fits every corpus.” The right balance depends on your documents and your queries, and it has to be measured.
Related terms: vector database, semantic search, reranking, RAG.
Related terms
- Semantic search · Search that finds content by meaning rather than exact words, by comparing numerical representations of the texts.
- Embedding · A list of numbers (a vector) that captures the meaning of a text or image, so content can be compared by how close in meaning it is.
- RAG (retrieval-augmented generation) · A technique that first retrieves relevant passages from your documents, then hands them to the model so its answer is based on them.
- Reranking · A second search stage in which a more precise model rereads the top candidates and puts them back in the right order of relevance.
- Vector database · A database that stores embeddings and quickly finds the ones closest to a query, which means the most relevant pieces of content.
- Chunking · Splitting long documents into small, coherent passages so they can be indexed and only the useful pieces are sent to the model.
Full glossary
Frequently asked questions
When is hybrid search really worth it?
Mostly when queries mix natural language with precise identifiers: product SKUs, error codes, customer names, legal references. Meaning-based search alone tends to miss those exact matches.
What is BM25?
A classic keyword-ranking formula that scores a document by how often the query terms appear, adjusted for how rare those terms are in the collection and for document length. It remains a strong baseline.
How do I set the balance between keywords and vectors?
Some tools expose a weighting parameter, such as alpha in Weaviate, where 0 means keywords only and 1 means vectors only. The right value comes from testing on a set of real queries with known good answers.
Does hybrid search replace reranking?
No, they complement each other. Hybrid search improves the candidate list; reranking then rereads those candidates with a more precise model and puts them in the right order.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.