Dense Passage Retrieval
The short answer · 23 words
In short
DPR searches for text passages based on learned vector representations of questions and text; the hit quality depends on the model and data.
Flow-alerta.cab (2026): Dense Passage Retrieval. https://flow-alerta.cab/en/glossar/dense-passage-retrieval. Editorially reviewed on: 09 October 2026.
How much detail would you like?
Practical
Relevant for semantic document search and RAG; checking utility and limits with your own search queries.
Possible example: A support assistant searches for suitable sections in a repair manual for a customer question.
Semantic search in internal knowledge bases, RAG systems for document analysis and support automation.
At a glance
- Core
- DPR trains two separate encoders (question and text section) in the same vector space. In contrast to pure keyword search, DPR also captures meaning relationships with different word choices.
- Use
- Relevant for semantic document search and RAG; checking utility and limits with your own search queries.
- Example
- Possible example: A support assistant searches for suitable sections in a repair manual for a customer question.
- Limits
- Inappropriate training data; overlooked article numbers or technical terms; similar but factually wrong hits. A combination with keyword search can help and should be tested.
In context
Vector Database
A vector database stores number vectors and supports their similarity search.
BM25 Ranking
Classic ranking function in Information Retrieval, which evaluates documents based on search terms, term frequency, document length and rarity of terms.
Bi-Encoder
Model architecture that encodes two texts or objects independently and then compares their resulting vectors.
Questions about this term
Does DPR replace keyword search?
Not mandatory. A hybrid search can offer advantages, but must be checked on your own data stock.
Is a vector database mandatory?
No. A local vector search library or, in the case of small stocks, a direct similarity calculation is also possible.
All details and background
DPR trains two separate encoders (question and text section) in the same vector space. In contrast to pure keyword search, DPR also captures meaning relationships with different word choices.
ELI5 – Explained simply
Like a search that can also connect differently formulated questions with suitable text passages. Whether the reference is correct must still be checked.
Why it matters in practice
Relevant for semantic document search and RAG; checking utility and limits with your own search queries.
Example
Possible example: A support assistant searches for suitable sections in a repair manual for a customer question.
Pitfalls and limits
- Inappropriate training data; overlooked article numbers or technical terms; similar but factually wrong hits. A combination with keyword search can help and should be tested.
Origin & history
Launched in 2020 by Karpukhin et al. (Facebook AI Research) for Open-Domain Question Answering.
Use in mid-sized companies
- Semantic search in internal knowledge bases, RAG systems for document analysis and support automation.
Related terms
- RAG
- Vector database
- Embeddings
- BM25
- Bi-encoder
- Cross encoders
Matching Flow-alerta service
AI assistants & automationAI Assistants & Automation – Search in Corporate KnowledgeMatching tool / use case
View use casesUse Case: Retrieve suitable manual sections for an AI assistant.Sources
- arxiv.orgKarpukhin et al. (2020), Dense Passage Retrieval for Open-Domain Question Answering
Quellen geprüft am 09.10.2026.
Editorially reviewed on: