Measured
On the production service, 28 October 2026, against the public Wisesight Sentiment test set.
| Measure | Result |
|---|
| Accuracy, 2,614 Thai social-media posts | 75.8% |
| Macro-F1 | 70.2 |
| Recall: positive / neutral / negative | 42.3% / 87.3% / 74.7% |
| Positive and negative confused | 1.5% of posts |
| Accuracy when the score is 0.9 or more (43% of posts) | 89.4% |
| Agreement with star ratings, 1,019 restaurant reviews | 84.4% |
| Median time through the API | 0.33 s |
The service is cautious with praise: a positive post it is unsure of comes back as neutral. Act automatically on labels with a score of 0.9 or more and review the rest. Method, confusion matrix and errors: the benchmark (PDF).
Limits
- Thai text.
- Up to 964 Thai characters per request, because the text travels in the query string. Send longer text as separate paragraphs.
Data handling
Text is processed in memory and not kept after the response. Consent and notice under PDPA and GDPR remain with you as the data controller.