Precision / recall / F1
Three scores per pair — precision (coverage of the candidate in the reference), recall (coverage of the reference in the candidate), and F1 (the balanced combination).
CatalogLatency Typically 0.3–1 second per pair Async false Rate Limit Per API key Max Input A few hundred tokens per text
cn8/text-similarity
Send two aligned lists — candidates and references — and the service compares them
Learn more about Text Similarity Text NLPsyncInstant previewtextanalysisnlp
Try it
Text A (candidate)Text B (reference)
Result
Results show up here
The sample input is ready — just hit Run.
Overview
Send two aligned lists — candidates and references — and the service compares them pair by pair. For each pair you get precision (how much of the candidate appears in the reference), recall (how much of the reference appears in the candidate), and F1 (the balanced score combining the two). Comparison is at the token level, so it captures meaning rather than exact wording. Responses are synchronous — one request, one set of scores per pair.
Key capabilities
Token-level matching
Compares texts token by token in meaning space, for a more nuanced score than whole-sentence matching.
Pairwise list input
Send aligned candidates and references arrays; you get one result per pair, in the same order.
Multilingual
Works across many languages out of the box.
When to use it
Duplicate detection
Find near-duplicate content using a threshold on F1.
Paraphrase detection
Check whether two sentences convey the same meaning (high F1).
Generation eval
Compare a generated text against a reference to measure how well meaning is preserved.
Input & output
input
A candidates array and a references array of strings, aligned by index.
JSON body
output
One {precision, recall, f1} result per (candidate, reference) pair.
JSON
Guides & tips
How it works
- You send aligned
candidatesandreferencesarrays, and each pair is scored on how closely the two texts match in meaning. precisionmeasures how much of the candidate appears in the reference;recallmeasures how much of the reference appears in the candidate;f1balances the two.- Results come back in the same order as your input:
results[i]is the score for(candidates[i], references[i]). Both arrays must be the same length.
Choosing F1 vs precision/recall
- Use
f1for general "are these the same?" questions — it's the balanced, all-round metric. - Use
precisionto ask "is the candidate a faithful subset of the reference?" (e.g. translation or paraphrase fidelity). - Use
recallto ask "does the candidate cover everything in the reference?" (e.g. summary completeness).
Specs
Schema
Request body
candidatesrequiredarray<string>Candidate texts (e.g. generated outputs or queries).
referencesrequiredarray<string>Reference texts. Must be the same length as candidates — pairs are matched by index.
Response
results[].precisionnumberHow much of the candidate is present in the reference, from 0 to 1.
results[].recallnumberHow much of the reference is present in the candidate, from 0 to 1.
results[].f1numberThe balanced combination of precision and recall, from 0 to 1. The main similarity score.
Pricing
Billed per request.
| Service | Unit | Price |
|---|---|---|
| Text Similarity | item | $0.004/request |
Cost is per request, not per pair — batch as many comparisons as you can per call.