When teams build retrieval, they often treat it as a choice: old-fashioned keyword search or modern vector (semantic) search. The better answer is usually both. Hybrid search combines them and consistently outperforms either alone.
Two kinds of "relevant"
Keyword search matches exact terms — great for names, codes, acronyms, and rare specific words. It's literal: search "GPT-4o" and it finds documents containing exactly that. But it misses meaning — search "car" and it won't find "automobile."
Vector search matches meaning — "car" finds "automobile," and it grasps paraphrases and concepts. But it can miss exact matches: a specific product code or an unusual name might not have a strong semantic signal, and vector search can overlook it.
Vector search knows what you mean. Keyword search knows what you said. You usually want both.
How hybrid works
Hybrid search runs both and combines the results — often by blending their relevance scores so a document that scores well on either (or both) ranks highly. This catches the exact-match cases keyword search excels at and the meaning-based matches vector search excels at. A re-ranker on top then sorts the combined shortlist by true relevance.
Why it matters
In real applications, queries are mixed: some are conceptual ("how do I reset my password?"), some are literal ("error code E-4021"). A pure vector system fumbles the literal ones; a pure keyword system fumbles the conceptual ones. Hybrid handles both, which is why it's become the default for serious retrieval and RAG systems. If your RAG "can't find things it obviously should," adding keyword search alongside your vectors is often the fix.
