What is Dense Retrieval?
Dense retrieval uses bi-encoders to map queries and documents into the same vector space. Similarity (dot product/cosine) retrieves top candidates efficiently.
How It Works (quick)
- Train: Contrastive learning on query–doc pairs.
- Encode: Precompute doc vectors; encode queries at runtime.
- ANN search: Return nearest neighbors; filter; pass to ranker.
- Cold-start: Use synthetic pairs if labels are scarce.
Why It Matters in E-commerce
- Great for long-tail and language variance.
- Low latency with precomputed vectors.
Best Practices
- Fine-tune on catalog + query logs.
- Refresh embeddings when inventory changes.
- Pair with neural re-rankers for quality.
Summary
Dense retrieval is fast semantic recall. Tune on your domain and combine with re-ranking.
