Hybrid search

What is hybrid search?

Hybrid search is a search approach that combines two different signals: keyword matching and semantic similarity. The keyword side finds the exact words, codes, names, and phrases in a document. The semantic side uses embeddings to find text with a similar meaning, even when the wording is different.

Each method misses something on its own. Keyword search is excellent for product codes, invoice numbers, error messages, and names. Vector search is better for intent, paraphrases, and messy human questions. Hybrid search runs both and merges the results into one ranked list.

This is why it became common in RAG systems. A user can ask a natural-language question, but the right answer may depend on an exact policy number, SKU, acronym, or error code.

Keyword search versus vector search

Keyword search
Classic search engines usually use ranking algorithms such as BM25. They score documents based on the words in the query, how often those words appear, and how rare they are across the corpus. This works well when the query and the document use the same terms.

Vector search
Vector search turns the query and documents into embeddings. The search engine then looks for vectors that sit close together. This works well when the user says delivery costs and the document says shipping fees.

The blind spot
Semantic search can blur exact strings. It may treat XR-4400 and XR-4500 as similar-looking tokens when the business cares about the exact one. Keyword search has the opposite problem: it may miss a useful document because the wording is different.

Hybrid search is the practical compromise: exact matches where exactness matters, semantic matches where meaning matters.

How the results are merged

The two searches return different scores, and those scores are not naturally comparable. BM25 scores are not on the same scale as vector similarity scores. Simply adding them together can produce strange rankings.

A common solution is reciprocal rank fusion, usually shortened to RRF. Instead of combining raw scores, RRF looks at where a document ranked in each result list. A document that appears high in both lists rises. A document that appears high in only one list still has a chance, but it is less dominant.

Another approach is a weighted blend. Some systems expose a parameter that shifts the balance between keyword and vector search. That gives more control, but it also gives you more tuning work.

The right blend depends on the data. Technical documentation full of error codes may need more keyword weight. Policy documents and support articles may benefit from more semantic weight.

Reranking after hybrid search

Many modern search stacks add a reranking step after hybrid retrieval. The first search pass retrieves a shortlist, often the top 20, 50, or 100 candidates. A more expensive model then reads those candidates and reorders them for the specific query.

Reranking is useful because the first pass is optimised for speed and recall. It tries to avoid missing the right document. The reranker is optimised for precision. It tries to put the best document at the top.

In a RAG application, that ordering matters. The language model can only use the snippets it receives. If the best evidence is pushed below the context limit, the final answer may be wrong even though the search system technically found the document.

Where hybrid search is useful

• Internal knowledge search. Employees search manuals, policies, tickets, contracts, and project notes with a mix of exact terms and vague questions.

• Technical support. Error codes, product names, and version numbers need exact matching, while problem descriptions need semantic matching.

• RAG over company documents. Retrieval quality controls answer quality. Hybrid search helps when documents contain acronyms, names, dates, and domain language.

• Product catalogues. Customers search with model numbers, brand names, and plain-language descriptions.

• Legal and compliance search. Exact clause numbers matter, but so do concept-level questions that use different wording from the source text.

What to watch out for

Bad keyword fields stay bad
Hybrid search does not fix poor text extraction, missing titles, broken OCR, or weak metadata. Both sides need clean input.

Embeddings set the semantic ceiling
If the embedding model handles your language or domain poorly, the vector side will miss important matches.

Access control must apply to both sides
Filter by permissions before results reach the user or the language model. It is not enough to secure only the final answer.

Measure with real queries
Do not tune the keyword/vector balance on made-up examples. Use search logs, support tickets, and real user questions.

Reranking costs latency
A reranker can improve quality, but it adds time and sometimes model cost. Use it where the extra precision is worth it.