Dense Retrieval

A dense retriever commonly uses a query encoder and a document encoder to produce vectors in a compatible space. Document vectors are computed during indexing, and a query vector is compared with them at search time. Training can use relevant query–passage pairs and negative examples to shape that space. Approximate nearest-neighbor indexes make searching large collections more efficient with a recall trade-off. Dense retrieval differs from a cross-encoder, which jointly processes each query–document pair, and from lexical retrieval, which ranks primarily through term matches. These methods can be combined rather than treated as exclusive choices.

The practitioner evaluates the encoder on representative relevance judgments, including languages, identifiers and difficult near matches. Preprocessing, truncation and query instructions are recorded with the model version. An exact vector search baseline separates embedding errors from approximate-index errors. Useful artifacts include the encoding pipeline, relevance benchmark and index parameters. Fine-tuning requires defensible positives and negatives with a held-out test. Changing encoders generally means rebuilding document vectors, because equal dimensions do not establish that two models produce comparable representations.

Dense models can blur rare terms, numerical differences or negation, and long documents may be truncated before the relevant material is encoded. Similarity scores are not calibrated across every query. Public benchmark performance may not transfer to specialist collections. Quality should be assessed with lexical and exact-search baselines, source-level relevance judgments and end-to-end answer checks, rather than assuming semantic representation makes every retrieval result meaningfully relevant.