Brave Search API

More Search API pages

API glossary

Hybrid search

A B C D E F G H I J K L M N O P Q R S T U V W X Y Z

Hybrid search is a retrieval approach that combines keyword (lexical) search and semantic (vector) search as two complementary methods, with results then merged into a single ranked list. By running both at once, hybrid search combines exact-term matches (which keyword search is good at) with meaning-based matches (which semantic search is good at). In mixed-query workloads, this often improves retrieval robustness compared with using either method alone.

In short: hybrid search runs keyword and semantic search together and fuses the results, giving you exact-match precision and semantic understanding in one ranking.

How hybrid search works

To help explain hybrid search, consider the real-world example of searching online for a new guitar. Lexical search can do well with exact keywords, for example if you already know you want a “J45 acoustic guitar.” By contrast, semantic search is more helpful if you have a sense of what you want the guitar to do or sound like (e.g., with a query like “folk guitar with warm tone”). In hybrid search, these two retrieval methods run side by side and are then combined:

  1. Keyword (sparse) retrieval: A lexical method such as BM25 finds documents that contain the query’s actual terms, ranked by how well those terms match.
  2. Semantic (dense) retrieval: An embedding model converts the query and documents into vectors and finds the closest matches by meaning, even when no words overlap.
  3. Parallel execution: Both retrieval methods analyze the same content, each returning its own ranked list of candidates.
  4. Fusion: A fusion step merges the two lists into one. A common method is Reciprocal Rank Fusion (RRF), which combines documents by their rank position rather than their raw scores, sidestepping the fact that the two methods score on different scales.
  5. Reranking: Optionally, the fused list is passed to a reranking model for a final, more precise ordering before results are returned.

The defining idea is that keyword and semantic search have different blind spots: one can miss synonyms and paraphrasing, while the other can miss exact terms like product codes or names. Fusing results with hybrid search can reduce misses caused by either weakness.

  • Keyword search matches exact terms and excels at precise strings such as codes, names, and rare technical words, but misses synonyms and paraphrase; hybrid search keeps that precision while adding semantic recall.
  • Semantic search matches meaning and handles paraphrasing well, but can undervalue (or “under-weight”) exact, rare-term matches; hybrid search keeps that understanding while restoring exact-match reliability.
  • Reranking reorders an existing candidate list for precision; hybrid search is about generating that candidate list from two retrieval methods in the first place, and the two are often used together.

Where hybrid search is used

  • RAG pipelines: Retrieving the right chunks for an LLM benefits from both exact and semantic matching, so hybrid retrieval is commonly used.
  • Enterprise and product search: Catalogs mix exact identifiers such as SKUs and model numbers with natural-language queries, which hybrid search handles in one pass.
  • Documentation and question answering: Users phrase questions loosely but sometimes need a specific term, and hybrid retrieval serves both.
  • Mixed-query workloads: Any system where some queries are keyword-like and others are conversational, since neither method wins on every query.

Hybrid search matters most when a corpus and its queries are varied enough that no single retrieval method is reliably best, which describes many real-world search workloads.

Semantic search, keyword search, vector database, embeddings, BM25, reciprocal rank fusion (RRF), reranking, retrieval-augmented generation (RAG), chunking.