← AI Terminology
Dense Retrieval
Dense retrieval encodes queries and documents into continuous embedding vectors and retrieves by nearest-neighbour similarity in that space.
DPR-style dual encoders made it standard for open-domain QA and RAG.
DPR-style dual encoders made it standard for open-domain QA and RAG.
Why It Matters in AI
Sparse methods miss paraphrases; dense dual encoders learn semantic match from data. They enable multilingual and fuzzy retrieval central to modern QA and RAG — with ANN indexes for scale.
Key Points
| Aspect | Description |
|---|---|
| Serve | Offline embed docs; online embed query + ANN |
| Limits | Exact strings; need hybrid/rerank |
| Classic | DPR (Karpukhin et al.), Sentence-BERT lineage |
| Related | Vector DB, ColBERT, cross-encoders |
| Training | Contrastive losses on relevant pairs |
| Architecture | Dual encoder / bi-encoder for query and doc |
Simple Analogy
Mapping every book and every question onto a globe of meaning, then walking to the nearest book-points to your question-point.
Common Usage Examples
- DPR retriever training
- embed documents at index time
all-mpnet-base-v2style sentence embeddings- Top-k from FAISS/HNSW
Summary
In short: Dense retrieval finds documents via embedding similarity — semantic nearest neighbours instead of only keyword overlap.