Dense retrieval · Dense passage retrieval
Dense retrieval is document retrieval based on dense embeddings, as opposed to sparse retrieval based on term frequency (BM25, TF-IDF). It finds passages by closeness in meaning, even with no word in common with the query.
Standalone definition · 36 words
How does dense retrieval work?
An encoder turns the query and each passage into vectors. The system returns the passages whose vector is closest, usually through approximate nearest-neighbor search. The Dense Passage Retrieval paper (Karpukhin et al., 2020) showed this approach beats BM25 on several open-domain question answering datasets.
Dense or sparse: what is the difference?
Sparse retrieval relies on exact matching of weighted terms and excels on proper nouns and references. Dense retrieval captures paraphrases and synonyms. Hybrid systems combine both and merge the rankings, for example with Reciprocal Rank Fusion.
Consequences for content
Exact vocabulary still matters (product names, references, acronyms), but a passage’s semantic clarity weighs more. Writing passages focused on one subject, with the technical term and its common rephrasing, serves both types of retrieval.
Frequently asked questions
Do AI engines use dense retrieval?
Details are not public. The US antitrust ruling mentions embedding-based retrieval at Google.
Does dense retrieval replace BM25?
Not always: hybrid systems are common.
Should I still work on keywords?
Yes for exact terms. Complement with semantic coverage.
Related terms
Further reading
- Checklist GEO and LLM / AI 2026
- 2026 GEO Tools Comparison: The Decision Guide
- ← Back to the GEO glossary (114 terms)