“The application was approved” and “The application was not approved” share a subject and much of their wording. A retrieval system may place them near each other even though the decision they describe is opposite.
Sentence Transformers’ similarity documentation explains how text embeddings and similarity functions are used to compare sentences. A similarity score is a measurement produced by a model and method. It is not, by itself, proof that two statements agree or that either statement is true.
Build a small challenge set
Before using similarity to merge records or answer questions, create examples from your own task. Include pairs that differ by negation, a date, a quantity, an entity and a condition. Keep the expected relationship beside each pair.
- “Delivery is included” versus “Delivery is not included.”
- “The limit is ten” versus “The limit is twenty.”
- “Available to members” versus “Available to all visitors.”
- Two different phrasings that genuinely express the same rule.
These are illustrative test cases, not measured results from a particular model. Run them through the exact model and settings used in your application. Review the errors rather than choosing a threshold solely because it looks comfortably high.
For search, related but disagreeing sentences may still be useful candidates. The next stage must inspect their claims and context. For automatic deduplication, merging them could destroy a meaningful distinction.
Keep the decision tied to the task
Record the model version, similarity function and evaluation examples. If you change the model, rerun the checks. A threshold established for short product questions may not transfer to long policy paragraphs.
Where correctness matters, add suitable validation or human review of the actual statements. Embeddings can help find material worth reading; they should not quietly become the final judge of agreement.