First, the key terms.

Vector
An ordered list of numerical coordinates.
Similarity
A score defined by a comparison rule, not a truth judgment.

The idea, explained

The geometry of an embedding space depends on how the representation was trained. Nearby vectors can help identify related text, but proximity does not establish truth. A retrieval pipeline typically represents documents and queries, finds candidates, and checks whether those candidates actually answer the question. Retrieved text remains evidence to inspect.

STEP 1

Represent comparable items

An embedding model turns queries and documents into vectors. The model and preprocessing define what patterns the coordinates capture. Vectors from incompatible models should not be compared as though their coordinates share meaning.

STEP 2

Retrieve, then inspect

Search selects candidates using a distance or similarity score. Check their date, permissions and actual relevance before using them. A later reranking step can compare candidates more carefully, but is another fallible model or rule.

Follow a worked example.

A query about resetting an account could retrieve a help article titled “Regain access” even though the wording differs. It could also retrieve an outdated policy. Showing the document date and opening the original can reveal that limitation.

Walk it through, one step at a time

  1. A query asks about this year’s school refund policy. The nearest passage discusses refunds but is dated last year.
  2. A slightly lower-ranked passage contains the current policy and the specific exception. Choose evidence based on applicability, not score alone.
  3. Preserve the passage’s original URL and date so the reader can verify the answer.

Where people get stuck

Semantic similarity is neither factual verification nor permission to execute instructions found in a retrieved document.

Now try it yourself.

Pause before opening each answer. Explain your reasoning in your own words.

1. Check the core idea

Why might the closest document still be the wrong answer?

Reveal the answer & reasoning

Its topic may be similar while its date, scope, exceptions or actual instructions do not fit the question.

2. Apply what you learned

Can the highest similarity score be interpreted as a percentage probability that the document is correct?

Reveal the answer & reasoning

No. Similarity measures a relationship in the representation space. Its scale is not automatically a calibrated probability of factual correctness or relevance.