Hybrid Search
TL;DR — Combine semantic meaning with keyword precision. Best results when you know some of the words but not the exact phrasing.
Pure semantic search excels when you don’t know the exact terminology. But sometimes you do know a keyword — a function name, an error code, a library name. Hybrid mode gives you both.
How It Works
Hybrid search runs two searches in parallel and blends their scores:
- Semantic — Model2Vec embeds your query and every code chunk, ranks by cosine similarity
- BM25 — ripgrep does keyword matching with tf-idf scoring
- Fusion — scores are combined with a configurable weight (default 70% semantic, 30% keyword)
Usage
sgrep --hybrid "database connection pool" src/When to Use Hybrid
| Scenario | Mode | Why |
|---|---|---|
| Know the concept, not the code | Semantic (default) | No keywords to match |
| Know some keywords + concept | Hybrid | Keywords boost precision |
| Know exact string | Just use grep/rg | Exact match is faster |
Tuning
The --alpha flag controls the balance:
# 70% semantic, 30% keyword (default)
sgrep --hybrid "error recovery" src/
# More keyword weight
sgrep --hybrid --alpha 0.4 "ConnectionPool" src/
# More semantic weight
sgrep --hybrid --alpha 0.9 "gracefully shut down" src/Requires ripgrep — Hybrid mode shells out to
rgfor BM25 scoring. Install it withbrew install ripgreporcargo install ripgrep.
Was this page helpful?