# Hybrid search

> Hybrid search combines word matching and embedding search so an exact term and a paraphrase can both hit.

Combining word matching and embedding search so exact terms and paraphrases can both hit.

Track: Retrieval & Grounding
Access: free
Card: https://buildsparkai.com/hybrid-search
Plain text: https://buildsparkai.com/hybrid-search.md
Last checked: 2026-10-18

## What it is

Hybrid search runs a word match and an embedding search, then merges the two lists. Azure AI Search's hybrid search overview says a hybrid query is one request configured for both full-text and vector queries. It runs those searches in parallel and merges them with reciprocal rank fusion (RRF). The same page says full-text ranking uses BM25, and vector ranking uses a nearest-neighbor search over embeddings. Keyword search can hit an exact term, such as a product code, a date, or a name. Vector search can hit text that is conceptually similar even when the words do not match. Elasticsearch's retrievers reference describes an RRF retriever that produces the top documents from reciprocal rank fusion, and a linear retriever that combines lexical fields and semantic fields.

## Why it matters

**Cost.** Azure's hybrid search overview describes one request that runs full-text search and vector search in parallel. This card does not add a price for that request. If you add semantic ranking on top, Azure's semantic ranking overview says that feature is billed by usage. The reranking card covers that second pass.

**Quality.** Azure's overview says keyword search is precise on exact terms such as product codes, dates, and names, and vector search can find conceptually similar text when the words do not match. A merge can keep a hit that either method alone would have dropped. This card does not claim the merge is always better.

**Risk.** A merged list can still put the wrong chunk first. The chunks that enter the prompt are still untrusted text. Hybrid search does not check that a chunk is true, or that this user may see it.

## Try today

Take one query that uses an exact name, and one that paraphrases the same fact. Run a keyword search and an embedding search. Note which query each method misses.

## Common mistake

Treating embedding search as a replacement for word matching. Azure's overview says some queries, such as product codes and people's names, perform better with keyword search because it can identify exact matches.

## Self-check

A query is the product code NX-14. Which search is built to hit that exact string?

Keyword, or word, matching

## Practice and deep

Coming soon. One exact-name query and one paraphrase, with the chunk each method missed written down.

Coming soon. How Azure and Elasticsearch each merge a keyword list and a vector list with reciprocal rank fusion.

## Sources

- [Azure AI Search, Hybrid search overview](https://learn.microsoft.com/en-us/azure/search/hybrid-search-overview)
- [Azure AI Search, Semantic ranking overview](https://learn.microsoft.com/en-us/azure/search/semantic-search-overview)
- [Elasticsearch, Retrievers](https://www.elastic.co/docs/reference/elasticsearch/rest-apis/retrievers)

Opened 2026-10-18. Vendor docs. This card does not add a rate, a usage count, or a benchmark score. Azure's page mentions benchmark testing. This card does not repeat that claim.