Tagging API for developers
Send content and your tag pool. rstags returns only tags from that pool, each with a calibrated score, and it does it without generating a word.
Every tag it returns is one you defined.
{
"object": "tagging",
"tags": ["billing", "churn-risk"],
"scores": [
{ "tag": "billing", "score": 0.94 },
{ "tag": "churn-risk", "score": 0.78 },
{ "tag": "bug", "score": 0.11 },
{ "tag": "feature-request", "score": 0.03 }
],
"model": "typesafe-ai/jev",
"threshold": 0.7,
"usage": { "decisions": 4, "balance": 9996 }
}Tags applied through rstags
5,115,295
Labels it applies, for example
A text model generates tokens until it decides to stop, then you parse what came out and hope it matches your taxonomy. An evaluation model answers a closed question with a probability. That one difference is both promises.
Zero hallucinations
Speed
rstags runs on Jev by TypeSafe AI (typesafe-ai/jev), which publishes its own latency and calibration benchmarks.
You can ask a language model to pick tags. It will usually work, and the times it does not are the times you find out in production, from a report built on a tag nobody defined.
| An LLM prompt | rstags | |
|---|---|---|
| Hallucinated tags | Possible, and silent | Zero, by construction |
| Time to answer | Grows with every token it writes | No output tokens to wait for |
| A score for every tag | No | Yes, for the whole pool |
| Unparseable response | Your problem | Retried, then surfaced as an error |
| Selection rule | Whatever the model felt | Your threshold, ranked, capped |
| Setup | Prompt engineering | One POST |
| Cost | Per token, output included | $50 per 100,000 tags |
No training data. No fine-tuning. Just scores.
Every tag you send comes back with a probability between 0 and 1. Set the threshold yourself, or store the scores and re-rank later without calling us again.
Tags above your threshold, ranked best first, capped at 5 by default. The same input and settings select by the same rule every time.
An evaluation model from TypeSafe AI that returns typed decisions instead of text, which is where both the speed and the zero come from.
POST content and tags. There is no project to create, no index to build, no training set to assemble and no model to fine-tune before the first call works.
Pass rules alongside the pool and they are attached to the tags they name. "Tag billing whenever a payment method is mentioned" is the whole configuration.
Pools up to 500 tags are split into batches and evaluated in parallel under one deadline. A big taxonomy needs no orchestration on your side.
A ticket, a review, an invoice. The pool changes, the request does not.
Content in
My card was declined twice renewing the Pro plan, and now the dashboard says my workspace is read-only. I have a demo on Thursday.
130 characters · 6 tags in the pool
Scores out
3 of 6 cleared the 0.7 threshold · 6 decisions billed
Authenticate with a bearer token, post JSON, read JSON. The response carries the tags, the score for every tag in the pool, and exactly what you were billed.
curl -X POST https://rstags.com/api/v1/tag \
-H "Authorization: Bearer rst_sk_..." \
-H "Content-Type: application/json" \
-d '{
"content": "My card was declined twice renewing the Pro plan.",
"tags": ["billing", "bug", "churn-risk", "feature-request"]
}'Stop shipping a prompt that returns a tag you never defined. Get a label and a number your code can branch on.
A calibrated classifier for routing, filtering and guardrails, without assembling a training set for every new label.
Tag every ticket consistently, including the ones that arrive at 2am, and get a taxonomy your reports can trust.
Classify inbound by intent and push it to the CRM before anyone opens it. Score thresholds decide what is worth a human.
Backfill labels across a table of free text. The same pool, the same threshold, the same result every run.
Ship the categorization feature your customers keep asking for this week, without a model to train or a prompt to babysit.
A fast answer is only useful if a slow or broken one cannot pass for it. Every failure is a clear error with nothing billed, never a half result dressed up as a success.
An unreadable answer is retried. An empty result is a legitimate answer and is returned as one.
Every attempt runs under a timeout you can set per request, inside an overall deadline.
Oversized content is rejected, or truncated only when you ask and flagged in the response.
If the model is unavailable, nothing is charged and the response says so.
One number. No seats, no plans, no minimum, and nothing that expires.
Pay as you go
$50per 100,000 tags
The first 10,000 are free, on every account.
One decision is one tag evaluated against one piece of content. A request with a 50 tag pool counts 50 decisions, because each tag is a separate question.
Duplicate tags collapse and a fallback tag is never billed, so a request always costs at most the number of tags you sent.
One tag evaluated against one piece of content. A request with a 50 tag pool counts 50 decisions, because every tag is a separate question. Duplicates collapse and the fallback tag is free, so you are never charged more than the array you sent.
That a tag outside your pool cannot come back, and not as a matter of policy. The model is asked a yes or no question about each tag you supplied and answers with a probability; there is no step at which a new string could be produced. It does not mean every judgement is right: a score can still be too high or too low, which is why you get the score and set the threshold.
Sub-second for a typical pool, because nothing is generated and the batches run in parallel. We will publish a median once we have measured one over real traffic, not before.
An evaluation model from TypeSafe AI. Instead of generating text it returns typed decisions with calibrated confidence, which is why the response is a set of numbers rather than a sentence to parse.
It is sent to the model to be scored and is not stored. There is no transcript archive on our side, which keeps your retention story short.
The API answers 402 with the number of decisions the request needed and your remaining balance. It never partially evaluates a pool, because half a result is worse than none.
Everything above your threshold, ranked best first, capped at 5 by default. You always get the scores for the rest, so raising or lowering the bar later costs nothing.
Create a key, post some text, read the scores. If it does not fit your taxonomy in ten minutes, nothing was spent.