Hybrid Retrieval

What is hybrid retrieval?

Keyword search works like a dictionary — you type "apple", it only matches the word "apple". Vector search works like meaning — say "a juicy fruit" and it can still think of an apple. Hybrid retrieval combines both: it recognizes exact words and understands intent.

Why isn't one method enough?

Keyword search: precise but rigid
Search "AI large model" and it misses a paragraph that says "large language model". Synonyms and casual phrasing slip through.
Vector search: flexible but floaty
It's great at finding "similar meaning", but it can miss exact matches — a model number, a legal article number.

How does the mixing work?

Search twice, then merge
Keyword search returns one batch, vector search another, and the two get fused by some rule.
Weighted scoring
Give each method a weight and rank by combined score, so results that are both relevant and exact float to the top.

What problem does it solve?

Hybrid retrieval is standard kit for RAG systems. It covers the gaps of any single method, handling jargon, code and serial numbers alongside casual questions — recall and precision both go up.

Bottom line: hybrid retrieval runs the dictionary and the meaning-detector side by side, so results come back both accurate and complete.