Embedding
Gömme
D3
An embedding is a learned numerical vector representation that a system can compare with other vectors or use as model input.
Review status: 2027-08-28
Technical explanation
The cited evidence covers word vectors and a dense-passage-retrieval design that scores question-passage candidates with learned representations and a similarity function.
Conceptual boundaries
An embedding is a numerical representation rather than a retrieved document, and the cited evidence does not establish semantic truth.
Provider-neutral example
A search system can encode a question and policy passages as vectors, then present the highest-ranked passages for additional checks.
Limitations
The cited evidence covers word embeddings and one dense-passage-retrieval design; it does not establish that one representation or similarity function works for every item or task.
Related concepts
Atomic claims and evidence
1.1Mikolov et al., Efficient Estimation of Word Representations in Vector Space
- Source
- Mikolov et al., Efficient Estimation of Word Representations in Vector Space
- Source role
- Authoritative source
- Exact locator
- Abstract: continuous vector representations and word similarity evaluation
- Supported claim
- Learned continuous vector representations can be computed from large datasets and evaluated with word-similarity tasks.
- Last verification
- Review due
- Scope limitation
- This evidence concerns word representations and does not claim that every embedding represents meaning correctly.
2.1Karpukhin et al., Dense Passage Retrieval for Open-Domain Question Answering
- Source
- Karpukhin et al., Dense Passage Retrieval for Open-Domain Question Answering
- Source role
- Authoritative source
- Exact locator
- Abstract; Sections 1 and 3: dense representations, candidate passages, and similarity scoring
- Supported claim
- Dense retrieval uses learned representations and a selected similarity function to score and rank question-passage candidates.
- Last verification
- Review due
- Scope limitation
- This describes the cited dense-passage-retrieval design and does not require every retrieval system to use embeddings.