embedding model dense retrieval

**Embedding Models for Dense Retrieval** are the **neural encoder architectures (typically transformer-based bi-encoders) that map queries and documents into a shared high-dimensional vector space where semantic similarity is measured by dot product or cosine distance — replacing traditional sparse keyword matching (BM25) with continuous, meaning-aware search**. **Why Dense Retrieval Replaced Keyword Search** BM25 counts exact token overlaps — it cannot match "automobile" to a document about "cars" or understand that "how to fix a leaking faucet" is relevant to a plumbing repair guide that never uses the word "fix." Dense retrieval encodes meaning into geometry: semantically related texts cluster together in vector space regardless of lexical overlap. **Architecture: The Bi-Encoder** - **Query Encoder**: A transformer (e.g., BERT, MiniLM, or a specialized model like E5/GTE) encodes the user query into a single fixed-dimensional vector (typically 768 or 1024 dimensions) via mean pooling or [CLS] token extraction. - **Document Encoder**: The same or a separate transformer independently encodes each document/passage into a vector of the same dimensionality. - **Similarity Score**: At search time, the system computes score = dot(query_vec, doc_vec) for every indexed document. Because both encodings are precomputed, this reduces to a Maximum Inner Product Search (MIPS) over the vector index. **Training Methodology** - **Contrastive Loss**: The model is trained on (query, positive_passage, hard_negative_passages) triplets. The loss pulls the query embedding toward its relevant passage and pushes it away from hard negatives — passages that are lexically similar but semantically irrelevant. - **Hard Negative Mining**: The quality of negatives determines model quality. BM25-retrieved negatives (high lexical overlap but wrong answer) and in-batch negatives (random passages from the same batch) provide complementary training signal. - **Distillation from Cross-Encoders**: A cross-encoder (which reads query and document jointly) produces soft relevance scores used to supervise the bi-encoder, transferring cross-attention quality into the fast bi-encoder architecture. **Deployment Stack** Document vectors are pre-indexed in approximate nearest-neighbor (ANN) systems like FAISS, ScaNN, or Pinecone. A query is encoded in real-time (5-20ms on GPU), and the ANN index returns the top-k most similar documents in sub-millisecond time even over 100M+ vectors. Embedding Models for Dense Retrieval are **the backbone of modern RAG (Retrieval-Augmented Generation) pipelines** — converting the entire knowledge base into a searchable geometric structure that LLMs can query for grounded, factual answers.