CORE GUIDE
Vector similarity is a retrieval signal
Vector similarity estimates closeness in an embedding space; it is useful for candidate retrieval but should not be confused with truth, authority or task relevance.
Mental model
An embedding maps content to vectors and a similarity function ranks geometric closeness. That score says 'these representations are nearby under this model', not 'this source is correct, current or sufficient'.
Why it matters
Vector search feels semantic, which makes teams over-trust the score. Similar passages can be outdated, duplicate the same source, omit the required identifier or be relevant to the general topic but wrong for the user's exact situation. Reliable retrieval therefore treats similarity as one feature in a pipeline that can also use lexical signals, metadata, authority and reranking.
01
Use similarity for recall, then apply stronger selection rules
Embed the query and candidate chunks with a compatible model, compute a similarity measure such as cosine similarity, and retrieve a broad candidate set. Then apply filters and ranking signals for scope, metadata, source authority, recency and exact identifiers. Evaluate recall on labeled queries so threshold and top-k choices are based on evidence rather than intuition.
02
Example: two refund policies with similar wording
An old and current refund policy embed very close together because most wording is shared. Pure vector similarity may rank the old policy first. Adding effective-date metadata and source priority lets the system keep semantic recall while ensuring the current controlling policy wins final selection.
Common failure modes
- Interpreting a high similarity score as factual correctness.
- Using one global similarity threshold without evaluating different query types.
- Ignoring exact IDs, dates and authority because embeddings feel more intelligent.
Engineering heuristics
- Measure retrieval recall before tuning generation prompts.
- Combine vector similarity with metadata, lexical and authority signals when the task needs them.
- Inspect false positives and false negatives by query slice.
Takeaways
- 01Similarity is geometric evidence, not epistemic authority.
- 02Retrieval quality depends on the full candidate-selection pipeline.
- 03A model cannot use evidence that retrieval failed to surface.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.