Dense Retrieval

Dense retrieval relies on a bi-encoder design where queries and documents are encoded independently by separate transformer models into dense vectors. This independence allows document embeddings to be pre-computed and indexed offline, dramatically reducing online latency. During inference, only the query must be encoded, and its vector is compared against the pre-built index using Approximate Nearest Neighbor (ANN) search.

• Query encoder: Processes the user's natural language input into a fixed-length vector
• Document encoder: Maps each passage to a vector in the same embedding space
• Shared latent space: Ensures relevant query-document pairs have high cosine similarity
• Offline indexing: Document vectors are computed once and stored, not recomputed per query

Relevance in dense retrieval is measured through vector proximity rather than term overlap. The most common metric is cosine similarity, which measures the angle between two embedding vectors regardless of their magnitude. This allows the model to recognize that 'automobile' and 'car' are conceptually close even when they share zero lexical tokens.

• Cosine similarity formula: cos(θ) = (A · B) / (||A|| × ||B||)
• Range: Scores fall between -1 and 1, with 1 indicating identical direction
• Dot product: Often used as a faster proxy when vectors are L2-normalized
• Euclidean distance: An alternative metric, sensitive to vector magnitude differences

Dense retrievers are trained using contrastive learning objectives that pull relevant query-document pairs together in the embedding space while pushing irrelevant pairs apart. The quality of training depends critically on hard negative mining—selecting documents that are superficially similar to the query but actually irrelevant. Without hard negatives, the model learns trivial distinctions and fails on nuanced queries.

• In-batch negatives: Other documents in the same training batch serve as negative examples
• Hard negatives: BM25 top results that don't contain the answer force the model to learn semantic distinctions beyond keyword overlap
• InfoNCE loss: A common objective that maximizes mutual information between positive pairs
• Knowledge distillation: Student bi-encoders often learn from more powerful cross-encoder teacher models

Exhaustive comparison against millions of document vectors is computationally prohibitive. Dense retrieval systems use ANN algorithms that trade a small amount of recall for orders-of-magnitude speed improvements. These algorithms organize vectors into graph or tree structures that enable sub-linear search time.

• HNSW (Hierarchical Navigable Small World): A multi-layer graph structure that enables logarithmic search complexity
• IVF (Inverted File Index): Clusters vectors using k-means and searches only the nearest clusters
• PQ (Product Quantization): Compresses vectors to reduce memory footprint and accelerate distance computations
• Recall vs. latency tradeoff: ANN parameters like ef_search and nprobe directly control this balance

Off-the-shelf embedding models trained on general web data often underperform on specialized enterprise corpora. Domain-adaptive fine-tuning uses proprietary query-document pairs to realign the embedding space toward the organization's specific terminology, acronyms, and entity relationships. This is essential for regulated industries like healthcare, law, and finance.

• Synthetic query generation: LLMs create plausible queries for existing documents to build training pairs
• Hard negative mining from production logs: Real user queries with low click-through rates provide authentic negative examples
• Parameter-efficient methods: LoRA adapters allow fine-tuning without full model retraining
• Evaluation benchmarks: Domain-specific retrieval metrics like NDCG@10 measure fine-tuning impact

Pure dense retrieval struggles with exact term matching for rare identifiers like product codes, legal citations, or error numbers. Production systems combine dense vector scores with sparse lexical scores from BM25 through reciprocal rank fusion or learned weighting. This hybrid approach captures both semantic intent and precise keyword matching.

• Reciprocal Rank Fusion (RRF): Combines ranked lists without requiring score calibration
• Linear interpolation: Weighted sum of normalized dense and sparse scores
• Learned fusion: A small neural model trained to predict final relevance from both signal types
• Metadata filtering: Pre-filtering by date, author, or category before vector search ensures precision

The foundational dual-tower model for dense retrieval. A query encoder and a document encoder independently map text to a shared vector space. This separation allows documents to be pre-encoded and indexed offline, enabling sub-linear search times via ANN. The key trade-off is speed over interaction fidelity—unlike cross-encoders, the query and document do not attend to each other during encoding.

The algorithmic backbone enabling fast vector similarity. Exact k-NN is computationally prohibitive at scale. ANN libraries like FAISS, ScaNN, and Annoy use techniques such as product quantization (PQ) and hierarchical navigable small world (HNSW) graphs to trade a small amount of recall for massive speed gains. This is what makes searching billion-scale vector indexes feasible in milliseconds.

The dominant learning paradigm for dense retrievers. Models like DPR and Contriever are trained on pairs of related texts (e.g., query-document, title-body). The objective is to pull embeddings of positive pairs together while pushing negative pairs apart in the vector space, typically using InfoNCE loss. In-batch negatives—treating other items in the mini-batch as negative examples—are a crucial efficiency trick that dramatically improves model quality.

Choosing the right encoder is critical and domain-dependent. General-purpose models like text-embedding-3-large or E5 work well out-of-the-box. For specialized domains (legal, medical, code), fine-tuned models like Voyage-law-2 or proprietary in-house encoders are necessary. Key evaluation metrics include MTEB benchmark scores and domain-specific retrieval accuracy. The embedding dimension (e.g., 768, 1024, 3072) directly impacts storage cost and search latency.

A critical design distinction. Asymmetric search uses different models for queries and documents (e.g., a lightweight query encoder and a powerful document encoder), optimizing for speed. Symmetric search uses the same encoder for both, which is required when searching for similar documents (e.g., 'find papers like this one'). Understanding this prevents a common architectural mistake where a system designed for short queries fails on long-document similarity tasks.