Comparisons
Vector search vs entity lookup
Vector search finds semantically similar material; entity lookup tests which specific subject a record denotes.
The distinction
Vector search compares encoded representations and returns nearby items under a chosen similarity measure. It helps discover relevant passages when wording differs, the question is exploratory, or a user cannot supply an exact identifier. The nearest result is a candidate for inspection, not proof that two names designate the same entity. Similar organizations can occupy nearby semantic space, while one organization can be described in very different language. Index coverage and the representation model also shape what can be found.
Entity lookup starts from an identifier, an established mapping, or an identity decision about the subject being requested. It is suited to retrieving records for a specific legal entity, person, product, or other defined object. Ambiguous names still need disambiguation, and an identifier is only useful within its issuing system and scope. When the right entity is unknown, semantic search can propose candidates; once identity matters, the lookup should rest on explicit identity evidence rather than similarity rank alone.
By dimension
| Dimension | Vector search | Entity lookup |
|---|---|---|
| Question answered | Which items resemble this query? | Which record refers to this subject? |
| Input | Text or another encodable representation. | An identifier or disambiguated identity. |
| Result | Ranked candidates with similarity scores. | A record tied to a specific subject. |
| Ambiguity | Similar wording may conflate distinct subjects. | Name ambiguity requires explicit resolution. |
| Coverage | Depends on the indexed content and encoder. | Depends on identifier coverage and maintained mappings. |
| Audit question | Why was this item ranked nearby? | What evidence supports this identity mapping? |
| Appropriate use | Discovery and candidate generation. | Subject-specific retrieval and joins. |
Choose Vector search when
- The user has a descriptive question rather than an identifier.
- Relevant material uses varied terminology.
- Returning candidates for human inspection is acceptable.
Choose Entity lookup when
- An action must attach to one specific entity.
- Two similarly named subjects must remain separate.
- Records are joined across sources using governed identities.
Using both
Use vector search to surface possible matches, then use identifiers and identity evidence to decide which subject each result actually describes.
A worked example
A fictional analyst types "Halden" and a drug name into a vector search over a document collection. The results include a press release about Halden Therapeutics AG, a paper whose author is named Halden, and a filing by Halden Biologics Inc., a different company. All are near the query in meaning space. The ranking carries no verdict on which is the intended entity.
An entity lookup starts elsewhere. The analyst supplies a registry number or an LEI, or picks a resolved entity from a list that disambiguates by country and registration. Every document or fact attached to that entity then comes back, and those attached to the namesake do not. The search was useful for finding candidates, and the lookup was needed to say which one was meant.
Where each one fails
Vector search fails on identity. Embeddings place names, descriptions and contexts near each other by similarity, so near neighbours include related but different entities, and different descriptions of one entity can be far apart. Changes of the embedding model change the neighbours. Index coverage also limits what can be found, and nothing signals that something is missing.
Entity lookup fails when the identifier is missing, wrong or shared. Not every subject has a registry identifier, and one identifier may stand for a group while another stands for a location. A lookup that works only for entities already resolved says nothing about those not yet resolved. Record linkage research has long treated candidate generation and decision as separate steps for this reason.
Questions to settle before choosing
Does the task end in an action on a named entity? If so, identity must be decided by evidence, with a reviewable basis. Is the query descriptive, with no identifier available? Then a candidate search is the right first step. How costly is a false match compared with a missed one? What identifiers exist for the domain, and how reliable are they? Where the answer is a list of candidates, who chooses, and is the choice recorded?
How the two meet in a review
The usual pipeline uses similarity to propose and rules or reviewers to decide. A reviewer should see for each resolved entity how it was selected, which candidates were rejected and why. Evaluation should separate the two stages: how often the right entity appears among the candidates, and how often the decision picks it. A system that reports only the final accuracy cannot tell which stage to improve.
Evaluating the two stages
Candidate generation and decision deserve separate metrics. For the first, ask how often the right entity appears in the candidate list at all, and at what rank. For the second, ask how often the chosen candidate is correct given that it was present, and how often the system should have declined to choose. A single end-to-end accuracy figure averages these together and can hide a retrieval problem behind a good decision step or the reverse. Samples should include namesakes, abbreviations, former names and entities that appear in the collection under only one description, since those are where similarity and identity diverge most.
Found an error on this page? See corrections for what counts as one, how to report it and the dated log.
Enquiry
Which do you run today: Vector search or Entity lookup?
Describe how your organization handles this now and one entity where it falls short. We reply with where a first record would sit alongside it and what it would need from you.
Or write to [email protected] in your own words.