How does hybrid search combine keyword and vector search?

Exact words + similar meaning = fewer misses.

Picture it first

You lost your blue water bottle at school. Friend A only looks for things with your name written on it (keyword search). Friend B looks for things that look like a blue bottle (vector search). Together they find it faster — and if both point at the same bottle, you're almost sure it's yours.

Now the same idea, in simple QA words

There are two ways to search. Keyword search finds the exact words you typed — perfect for error codes, IDs and names. Meaning search (vector search) finds text that means the same thing, even with different words. Hybrid search runs both and combines the results, so you get exact matches and similar-meaning matches. Results that appear high in both lists go to the top.

How the story connects to the real thing

In the storyIn the AI system
Friend A: looks for your name on the bottleKeyword search (exact words)
Friend B: looks for anything that looks like a blue bottleVector search (similar meaning)
Both friends search togetherHybrid search
Both point at the same bottlesurely yours → Result high in both lists goes first

Real-world example

A support engineer searches the knowledge base for 'ERR_CONN_4012 after login'. Pure vector search returns generic 'login problems' articles. Keyword search finds the one article that mentions ERR_CONN_4012 exactly. Hybrid search puts that exact article first and still shows related login articles below.

Step by step, in plain words

  1. Your question is sent to keyword search and meaning search at the same time.
  2. Each gives its own ranked list.
  3. The two lists are combined.
  4. Items ranked high in both lists come first.
  5. The top results go to the AI.

How a tester tests it

  • Create query categories: exact IDs/codes, natural language, mixed. Measure each with keyword-only, vector- only and hybrid.
  • Test typos and abbreviations — keyword search fails on typos, vectors usually cope.
  • Test the fusion weight: does a change in alpha regress one category?
  • Include names and numbers — vector search alone often misses them.

Try it yourself

1) Search your email for an exact invoice number, then for 'that bill from the electricity company'. Notice which style of search each one needs. 2) Write 5 queries a hybrid search would handle better than either alone.

Interview angle

“When does pure vector search fail?”

On exact tokens with little semantic meaning — order IDs, error codes, SKUs, people's names, version numbers. Embeddings blur these. Hybrid search adds BM25 so the exact match is found, and fusion keeps the semantic benefits for natural-language questions.