In plain terms
Keyword search is the colleague who remembers exact wording; semantic search is the one who remembers the gist. Each misses what the other catches. Hybrid search asks both and combines their shortlists, so a query like “error E-4012 when exporting to PDF” matches the exact code and also the articles about export failures in general.
Why it matters
For business content, hybrid search is the sensible default. Real queries mix exact things (product codes, customer names, legal references) with loose descriptions, and neither method alone handles both. Teams that build RAG on semantic search only tend to rediscover this when users search for a contract number and get nothing.
Example
A manufacturer's support assistant uses semantic search alone and fails on “torque spec for M8-1.25 bolt”, returning general articles on fasteners. After adding keyword search and merging results, the exact specification sheet ranks first, and loosely worded questions still work.
Most often confused with
Hybrid Search vs. Semantic Search
Semantic search is one of the two ingredients. Hybrid search adds a keyword engine next to it and a fusion step that reconciles the two rankings. The added complexity is modest, and the gain on queries containing names, codes and rare terms is usually large.
Under the hood
The lexical side is typically BM25 over an inverted index; the dense side is nearest-neighbour search over embeddings. Results are merged with reciprocal rank fusion, which needs only the rank positions, or with a weighted sum of normalised scores, which needs tuning. A reranker is often applied to the merged candidates. Learned sparse models such as SPLADE offer a middle path. Most search engines and vector databases now support hybrid queries natively, so the remaining work is tuning weights and evaluating on real queries.