# What is Similarity Score?

Also called Relevance Score, Match Score.

A similarity score is the number a retrieval system assigns to a candidate result indicating how closely it matches the query, used to rank results and sometimes to filter weak ones. Scores are relative to the scoring method and the corpus, so a value that looks high in one system may be unremarkable in another.

Different retrieval methods produce different score scales. Cosine similarity ranges from negative one to one, though embeddings from many models cluster in a narrow positive band. Inner product is unbounded. Euclidean distance is a dissimilarity where smaller is better. Lexical scores such as BM25 are unbounded and depend on corpus statistics.

Because scales differ, absolute thresholds transfer poorly. A cutoff tuned on one embedding model or one corpus can silently discard everything or admit everything after a change. Where a threshold is needed, it should be calibrated on labeled examples from the actual corpus and rechecked whenever the model or the content distribution changes.

Scores measure similarity, not correctness or relevance in a human sense. A passage discussing the same topic in a contradictory or outdated way can score highly, and a passage that answers the question in unusual wording can score modestly. This gap is why reranking and grounded verification exist as separate stages.

Score distributions still carry useful signal in aggregate. A query whose top results all score near each other suggests ambiguity, while a large gap between first and second suggests a confident match. Some systems use that shape to decide whether to rerank, ask a clarifying question, or abstain from answering at all.

## Key points

- Ranks candidates and can filter weak matches.
- Scales differ by method, so thresholds rarely transfer.
- High similarity does not mean correct or current.
- Calibrate cutoffs on labeled data from the real corpus.
- Score gaps signal confidence or ambiguity.

## In practice

A query returns passages scoring 0.81, 0.79, and 0.78 under cosine similarity, all close together, which suggests several plausible matches rather than one clear answer. The system sends all three to a reranker instead of trusting the top hit. Another query returns 0.88 followed by 0.52, a wide gap indicating a single confident match.

## Related terms

- [Cosine Similarity](/en/glossary/cosine-similarity)
- [Semantic Search](/en/glossary/semantic-search)
- [Reranking](/en/glossary/reranking)
- [Embedding](/en/glossary/embedding)
- [Recall and Precision](/en/glossary/recall-and-precision)

[Back to the AI Glossary](/en/glossary)