# What is Dense Retrieval?

Dense retrieval finds relevant text by comparing embeddings, so a query matches passages whose meaning is close even when no words overlap. Both query and documents are converted into fixed length vectors by a neural model, and similarity is computed in that shared space. It is the retrieval style that most vector database workflows are built around.

The defining property is that matching happens in a learned space rather than on surface tokens. A query about lowering staff turnover can retrieve a passage about employee retention because the model placed both near each other during training. That tolerance for paraphrase, synonyms, and different phrasing is what makes dense retrieval useful for natural language questions typed by people who do not know the corpus vocabulary.

The same property is also its weakness. Dense models generalize, so they are unreliable on exact identifiers such as part numbers, error codes, rare proper nouns, and version strings, where a single character changes the meaning but barely moves the vector. They also inherit the training distribution of the embedding model, which is why performance can drop sharply on specialized jargon the model rarely saw.

Quality depends heavily on how documents are prepared. Passages that are too long blur several topics into one vector; passages that are too short lose the context that made them meaningful. Asymmetric setups, where short questions are compared against longer passages, often use models trained specifically for that shape, and some models expect a task instruction prefix on the query.

In practice dense retrieval is rarely deployed alone. It is combined with keyword scoring, filtered by metadata, and followed by a reranking pass, because each layer covers a different failure mode. Treating dense similarity as the whole system, rather than as the recall-oriented first stage, is a common source of disappointing results.

## Key points

- Matches meaning rather than shared words
- Handles paraphrase and synonyms well
- Unreliable for exact codes, identifiers, and rare terms
- Quality is sensitive to passage length and preparation
- Usually paired with keyword scoring and reranking

## In practice

A user asks an internal assistant how to get reimbursed for a train ticket. The policy document never uses the word reimbursed; it says expense claims for rail travel are settled monthly. Dense retrieval still returns that paragraph, because the query and passage vectors sit close together. A pure keyword search over the same corpus returns nothing useful, since no query word appears in the paragraph.

## Related terms

- [Embedding](/en/glossary/embedding)
- [Semantic Search](/en/glossary/semantic-search)
- [Sparse Retrieval](/en/glossary/sparse-retrieval)
- [Hybrid Search](/en/glossary/hybrid-search)
- [Retrieval Pipeline](/en/glossary/retrieval-pipeline)

[Back to the AI Glossary](/en/glossary)