Why pure vector search isn't enough
Basic RAG embeds the query and finds the most similar chunks — which assumes the query is clear and well-formed. In reality (10:15):
- Users ask vague questions.
- Users use different words than what's in your documents.
- Sometimes one query needs information from multiple places.
The hybrid approach
| Method | Strength |
|---|---|
| Vector search | Great for meaning / semantic similarity (10:33) |
| Keyword search | Catches exact matches vectors miss: product names, error codes, specific terms (10:33) |
- Most production systems today use both, then re-rank the combined results to surface the most relevant chunks (
10:33).
flowchart LR
Q[Query] --> VS[Vector search<br/>meaning]
Q --> KS[Keyword search<br/>exact terms]
VS --> M[Merge]
KS --> M
M --> RR[Re-rank] --> TOP[Most relevant chunks]