Skip to content
    Flow-alerta.cab

    Dense Retrieval

    Data integration & searchDense vector search
    Editorially reviewed on: 2 sources
    Your reading view

    The short answer · 14 words

    In short

    Dense Retrieval compares dense vector representations of searches and documents to find relevant content.

    Flow-alerta.cab (2026): Dense Retrieval. https://flow-alerta.cab/en/glossar/dense-retrieval. Editorially reviewed on: 09 October 2026.

    How much detail would you like?

    Practical

    Relevant for semantic enterprise search and RAG; in addition to the search for exact terms.

    Possible example: The search for “vacation application deadline” finds a guideline with the phrase “recreation leave submission period”.

    Knowledge management, intranet search functions, e-commerce catalogue search and FAQ bots in SMEs.

    At a glance

    Core
    An embedding model maps requests and documents as number vectors. A similarity measure, such as scalar product or cosine similarity, serves to order the hits. High similarity does not guarantee a factually correct answer.
    Use
    Relevant for semantic enterprise search and RAG; in addition to the search for exact terms.
    Example
    Possible example: The search for “vacation application deadline” finds a guideline with the phrase “recreation leave submission period”.
    Limits
    Weaknesses in IDs, proper names or unfamiliar technical texts; model and search methods do not check on your own data stock.

    In context

    Dense Passage Retrieval

    DPR searches for text passages based on learned vector representations of questions and text; the hit quality depends on the model and data.

    Embedding

    An embedding maps content as a number vector to compare similarities.

    Questions about this term

    What makes Dense different from Sparse Retrieval?

    Dense uses dense vectors. Sparse uses sparse representations, such as BM25 or learned sparse models; it is not limited to mere word counting.

    How can the search be accelerated?

    Inter alia, by approximative vector search; speed and hit quality must be checked together.

    All details and background

    An embedding model maps requests and documents as number vectors. A similarity measure, such as scalar product or cosine similarity, serves to order the hits. High similarity does not guarantee a factually correct answer.

    ELI5 – Explained simply

    Instead of searching for identical letters, the system searches for content relatedness in the same thought space.

    Why it matters in practice

    Relevant for semantic enterprise search and RAG; in addition to the search for exact terms.

    Example

    Possible example: The search for “vacation application deadline” finds a guideline with the phrase “recreation leave submission period”.

    Pitfalls and limits

    • Weaknesses in IDs, proper names or unfamiliar technical texts; model and search methods do not check on your own data stock.

    Origin & history

    Neural representations and vector search have a longer history of development. A documented example is DPR (2020); Zhan et al. (2021) investigate training with difficult negative examples.

    Use in mid-sized companies

    • Knowledge management, intranet search functions, e-commerce catalogue search and FAQ bots in SMEs.

    Related terms

    Matching Flow-alerta service

    AI assistants & automationAI Assistants & Automation – Semantic Knowledge Search

    Matching tool / use case

    View use casesUse Case: Find internal documents based on similar searches.

    Sources

    • arxiv.orgZhan et al. (2021), Optimizing Dense Retrieval Model Training with Hard Negatives
    • arxiv.orgKarpukhin et al. (2020), Dense Passage Retrieval for Open-Domain Question Answering

    Quellen geprüft am 09.10.2026.

    Editorially reviewed on:

    Next step

    Put knowledge into practice

    Discuss your projectView use cases