---
title: Tuatara Topics — topics from your documents, no training
url: https://cymetica.com/topics
api: https://cymetica.com/api/v1/topics/v1
---

# Tuatara Topics

A topic model served as an OpenAI-compatible provider: POST up to 2,000 documents to
`https://cymetica.com/api/v1/topics/v1/topics` with `k` topics (2 to 100) and get every topic's words, a label and each
document's topic mixture back in one call, or one document (`document`, optional `topic_set`: `auto`, `general`,
`finance`) to place on the served topic sets — the answer names the set, each document's `confidence` (its top
topic's weight) and `on_topic_set` (false when nothing served covers the text). Nothing is stored (zero data
retention). Model id `tuatara-topics-1.0`; `https://cymetica.com/api/v1/topics/v1/models` describes it,
`https://cymetica.com/api/v1/topics/v1/health` reports uptime, `https://cymetica.com/api/v1/topics/v1/chat/completions` takes the same request as a chat message.

## Measured on public data (50 topics, TopMost preprocessing, mean ± spread over seeds)

| Dataset | Model | Seeds | C_V | TD | Purity | NMI | Published C_V / TD / purity / NMI |
|---|---|---|---|---|---|---|---|
| 20NG | FASTopic | 5 | 0.597 ± 0.011 | 0.992 ± 0.004 | 0.594 ± 0.010 | 0.531 ± 0.004 | 0.426 / 0.983 / 0.577 / 0.525 |
| 20NG | LDA (ours; published: LDA-Mallet) | 5 | 0.597 ± 0.006 | 0.683 ± 0.009 | 0.523 ± 0.014 | 0.442 ± 0.008 | 0.371 / 0.763 / 0.514 / 0.451 |
| 20NG | Tuatara Topics 1.0 (served configuration) | 5 | 0.616 ± 0.006 | 0.780 ± 0.014 | 0.546 ± 0.007 | 0.445 ± 0.009 | — / — / — / — |
| 20NG | Tuatara Topics (research configuration: runs=9, embed=0.6, iters=1000, 2026-09-29) | 5 | 0.643 ± 0.010 | 1.000 ± 0.000 | 0.646 ± 0.008 | 0.579 ± 0.004 | — / — / — / — |

Published figures are the FASTopic paper's (arXiv 2405.17978) for the same model on the same data. Every
run's topic words and document mixtures are CSVs under https://cymetica.com/static/research/tuatara-topics/ and
https://cymetica.com/static/research/tuatara-topics/score.py recomputes diversity, purity and NMI from them with the standard library alone.