Guides/Embeddings vs keyword vs hybrid
Embeddings vs keyword vs hybrid
Retrieval debates aren't religious wars — they're mismatch diagnostics. Keywords excel when literals matter; embeddings excel when users paraphrase; hybrids recover recall when enterprise corpora mix both failure modes.
Decision flow
Walk this when recall tickets spike — isolate lexical misses before tuning embedding dimensions or chunk sizes.
Mode comparison
| Mode | Reach for it when… | Strength | Watch-outs |
|---|---|---|---|
| Keyword (sparse) | IDs, policy numbers, product codes, known titles | Fast, cheap, exact token overlap | Paraphrase & typos break recall |
| Dense (embeddings) | “Similar ideas” not exact words; long unstructured blobs | Semantic similarity; tolerant wording | Cold-start corpus tuning; can blur distinct entities |
| Hybrid | Must recover both literal SKUs and fuzzy descriptions | Best recall breadth on messy enterprise corpora | Fusion weights + infra — tune with labeled misses |
Fit by query shape (illustrative chart)
Replace illustrative scores with offline evals on your labeled query sets — the ranking order matters more than absolute numbers.
Fit by query shape (illustrative)
Shipping notes
- Start from failure buckets (keyword miss vs semantic drift) before buying GPU-heavy rerankers.
- Hybrid fusion weights belong in experimentation — log retrieval provenance so PMs can audit misses with engineers.