Definition
Hybrid search combines more than one retrieval signal, usually lexical search and semantic vector search.
The plain-English version:
use exact word matching and meaning-based matching together
This is useful because neither keyword search nor embedding search is perfect on its own.
Why This Concept Exists
Vector search is good at meaning-like similarity.
Keyword search is good at exact terms.
RAG products often need both.
Consider these queries:
ERR_CONN_RESET
invoice INV-48291
SOC 2 retention policy
how do I stop monthly billing?
The first three contain exact tokens that matter. The last one may need semantic matching to find "cancel subscription."
Hybrid search exists because real user queries mix fuzzy intent with exact strings, identifiers, acronyms, dates, names, and product-specific vocabulary.
The Beginner Mental Model
A beginner may think:
Embeddings replace keyword search.
That is a trap.
Embeddings help when words differ but meanings align. Keyword search helps when exact terms carry the answer.
A better mental model:
lexical search catches exact language
vector search catches semantic intent
hybrid search blends both candidate sets
How It Works At A Practical Level
A hybrid system may run two searches:
query -> keyword retriever -> lexical candidates
query -> embedding retriever -> vector candidates
Then it combines them:
merge candidates
normalize or fuse scores
deduplicate
rerank
return final results
Some systems blend scores directly. Others use one search path for candidate generation and another for reranking or filtering.
The right design depends on the corpus, query types, latency budget, and evaluation data.
A Concrete Example
A user asks:
Why does API error E1049 happen during checkout?
Vector search may find general checkout troubleshooting docs.
Keyword search may find the exact error-code page.
Hybrid search can include both:
- exact error-code match
- semantically related checkout flow docs
- maybe recent incident notes filtered by service
Then reranking can decide which candidates best answer the question.
Fusion And Ranking
Hybrid search needs a way to combine different score types.
Keyword scores and vector similarity scores may live on different scales.
Common strategies include:
- retrieve from both paths, then rerank
- normalize scores before blending
- use reciprocal-rank-style fusion
- weight one path more for certain query types
- route exact-ID queries to lexical-first retrieval
Score fusion is not only math. It is product judgment about what should matter for users.
Failure Modes
Hybrid search can fail when:
- one signal dominates every query
- scores are combined without calibration
- exact matches drown out better semantic evidence
- semantic matches hide required identifiers
- duplicate chunks fill the top results
- query classification routes to the wrong path
- evaluation only tests easy semantic questions
Hybrid search adds power, but also adds tuning surface.
Common Confusions
Hybrid search is not always better by default.
It helps when both signals add value. It can hurt if poorly fused.
Keyword search is not obsolete.
Exact terms, identifiers, error codes, names, and dates often matter.
Semantic search is not enough for all RAG.
Meaning similarity can miss literal constraints.
Hybrid search is not the same thing as reranking.
Hybrid search combines retrieval signals. Reranking reorders candidates, often after retrieval.
What This Does Not Mean
Hybrid search does not remove the need for evals.
You still need to measure whether the final retrieved context supports the answer. A blended candidate list can look impressive and still miss the evidence users need.