In short. Dense retrieval uses learned vector representations to select passages that match a query, including passages with different wording.

Definition: Dense retrieval selects text passages by comparing numerical representations of a query and candidate passages. These representations, called embeddings, encode learned relationships between texts so that retrieval can match related meanings across different wording.

Key facts

Karpukhin and colleagues' November 2020 paper, Dense Passage Retrieval for Open-Domain Question Answering, describes separate question and passage encoders with dot-product scoring.
As of October 2026, Microsoft's vector search overview illustrates conceptual similarity with “dog” and “canine”.
The DPR repository, consulted in October 2026, documents exhaustive search as its default, with optional approximate indexes.
The BEIR benchmark paper evaluates retrieval across 18 datasets and finds that dense models often underperform other approaches when transferred beyond their training setting.

How does Dense Passage Retrieval work?

An encoder is a model that maps text into a numerical vector. In the DPR paper, separate encoders process the question and each passage. This arrangement is called a dual encoder. The authors use independent BERT language models and produce vectors with 768 dimensions.

The following notation matches the paper's scoring rule. E_Q is the question encoder, E_P is the passage encoder, and dot multiplies corresponding vector components and sums the results.

similarity(question, passage) = dot(E_Q(question), E_P(passage))

Passage vectors are computed in advance and indexed with FAISS, a library for vector similarity search. At query time, the system encodes the question and retrieves passages with high similarity scores. Training teaches the encoders to score relevant question-passage pairs above irrelevant pairs.

The paper splits Wikipedia articles into passages of 100 words and prepends each passage's article title. This is a specific experimental setup; passage boundaries remain a retrieval design choice.

What does a website retrieval system need to store?

Microsoft's vector search documentation describes an indexing workflow and a query workflow. Source text becomes embeddings stored in vector fields. A query is also encoded, and the search service returns documents associated with nearby vectors.

For a business building search over its own website, this means preparing both searchable content and its vector representation. Azure AI Search supports splitting documents into chunks and generating embeddings within its indexing pipeline, or accepting embeddings generated externally.

The same documentation defines hybrid search as running vector search and keyword search together, then merging their results. This gives a concrete comparison option when evaluating retrieval for a website's content.

How can retrieval quality be checked?

The BEIR paper examines zero-shot evaluation: applying a retrieval model to tasks without training it on examples from those tasks. Its findings make performance on the intended content collection an important evaluation question. The benchmark finds BM25, a term-based retrieval method, to be a strong baseline across its varied tasks.

For passage-level inspection, the DPR repository documents result records containing passage text, a retrieval score and a has_answer field. Results are ordered by similarity. These fields support checking whether highly ranked passages actually contain the information needed to answer a question.

The repository also offers approximate indexes alongside exhaustive search. Evaluation therefore needs to consider the index configuration as well as the encoder.

FAQ

Does dense retrieval write the answer?

DPR's paper separates the retriever from a reader that extracts an answer from retrieved passages. Dense retrieval supplies candidate evidence within that system.

Does dense retrieval always use approximate search?

The DPR repository documents exhaustive search as the default. Approximate search is an optional configuration, with a different tradeoff between searching and indexing costs.

Can keyword search run alongside dense retrieval?

Yes. Microsoft's vector search overview describes hybrid requests that execute keyword and vector queries in parallel and return merged results.

Sources

Check how a page reads to answer engines. Open the GEO score analyzer