dense retrieval

**Dense retrieval** is the **semantic search approach that represents queries and documents as dense vectors and ranks by embedding similarity** - it excels at conceptual matching beyond exact keyword overlap. **What Is Dense retrieval?** - **Definition**: Neural retrieval method using learned embeddings for both query and document representations. - **Scoring Function**: Uses cosine similarity or dot-product distance in vector space. - **Strength Profile**: Captures paraphrases, synonyms, and semantic relations. - **Infrastructure Need**: Requires vector indexing and ANN search for large-scale performance. **Why Dense retrieval Matters** - **Semantic Recall**: Finds relevant content even when wording differs from query terms. - **Modern RAG Core**: Common baseline for knowledge retrieval in LLM pipelines. - **Cross-Domain Utility**: Works well for natural-language questions and conceptual topics. - **Scalability**: Embedding precomputation plus ANN supports large corpus search. - **Quality Tradeoff**: Can miss rare exact tokens like IDs, codes, and uncommon names. **How It Is Used in Practice** - **Encoder Selection**: Choose domain-tuned embedding models for better relevance. - **Index Optimization**: Tune ANN parameters for latency-recall balance. - **Hybrid Fusion**: Combine with sparse retrieval to recover exact-term precision. Dense retrieval is **a central semantic-search primitive in RAG systems** - vector similarity enables broad conceptual coverage that lexical-only methods often miss.