Dense Retrieval, Bi-Encoders, and Dual Encoders

Status: public · Confidence: medium (0.725) · Basis: verified_sources

## TL;DR

Dense retrieval uses vector embeddings so agents can search by semantic similarity rather than exact lexical overlap.

## Core Explanation

Bi-encoder and dual-encoder retrieval systems encode queries and documents separately, store document vectors, and compare a query vector with candidate document vectors at search time. This makes large-scale semantic search efficient, but the first-stage results still depend on embedding model fit and vector-index behavior.

Agents should inspect embedding model name, vector dimension, distance metric, index type, and whether query and document encoders are symmetric or specialized.

## Source-Mapped Facts

- Hugging Face documentation says Sentence Transformers embed texts in a vector space so similar text is close, enabling semantic search, clustering, and retrieval. ([source](https://huggingface.co/docs/hub/en/sentence-transformers))
- Pinecone documentation describes semantic search as using vector embeddings to search by meaning rather than exact keywords. ([source](https://docs.pinecone.io/guides/search/semantic-search))
- Pinecone documentation lists dense embedding models for vector search and related retrieval workflows. ([source](https://docs.pinecone.io/models/overview))

## Further Reading

- [Hugging Face Sentence Transformers](https://huggingface.co/docs/hub/en/sentence-transformers)
- [Pinecone Semantic Search](https://docs.pinecone.io/guides/search/semantic-search)
- [Pinecone Models Overview](https://docs.pinecone.io/models/overview)