Text classification API, no labeled data needed

How it works

1. Send the text as the context: a message, an e-mail, a comment or a JSON object.
2. Describe the categories as the options of a pick, 2 to 26 of them, each with a short description of what it covers.
3. Iryx returns the most likely category, its confidence and the spectrum: the probability of each category.

One request holds up to 32 questions about the same text. So one call classifies the topic, says whether it is urgent (check) and rates the mood (scale):

{
  "context": {"message": "I was charged twice on my card for the same order. I need this fixed today."},
  "decisions": {
    "topic":  {"kind": "pick", "prompt": "What is the message about?",
               "options": {"billing": "charges, refunds, cards",
                           "shipping": "delays, tracking, address",
                           "returns": "exchanges or returns",
                           "other": "anything else"}},
    "urgent": {"kind": "check", "prompt": "The customer needs a reply today"},
    "mood":   {"kind": "scale", "prompt": "Customer mood",
               "levels": ["calm", "worried", "annoyed", "furious"]}
  }
}

"results": {
  "topic":  {"kind": "pick", "value": "billing", "confidence": 0.94,
             "spectrum": {"billing": 0.94, "shipping": 0.02, "returns": 0.02, "other": 0.02}},
  "urgent": {"kind": "check", "value": 0.88},
  "mood":   {"kind": "scale", "value": 1.6, "level": "annoyed",
             "spectrum": {"calm": 0.05, "worried": 0.35, "annoyed": 0.55, "furious": 0.05}}
}

Where a classification API helps

• Support: topic, urgency and tone of each ticket, to send it to the right queue.
• E-mail: what needs a person and what can be archived.
• Moderation: "does this comment break the rule?", with a probability.
• Engineering: triage of bugs and alerts by team and severity.
• Systems: JSON records and events, classified as they arrive.

These are example applications; the measured numbers are in the next section.

Published numbers

On 2026-10-19, Iryx 1.0 was measured next to ten other decision models, all called from the same machine in Brasília, at the same time: accuracy on Iryx's own test, in Portuguese, and the time per case, over the internet.

Measured on 2026-10-19: eleven decision models called from the same machine in Brasília, over the internet, at the same time. Accuracy on Iryx's own test and answer key, in Portuguese, the language Iryx is trained in (200 cases, 225 decisions). Time: the same 300 public decision cases (1 to 64 questions each) were sent to all; each model's median uses the cases it answered (the last column; for Iryx, 17 cases above the per-request size limit were left out of the median), including the round trip over the internet. Eight of the other ten were called through an API aggregator; Iryx and the other two, directly. * 10 of the 200 cases hit a service error, and an error counts as wrong. The full table, with the 95% interval and the time p95, is at Iryx 1.0 numbers.

In short: on Iryx's own test, in Portuguese, Iryx had the highest accuracy of the 2026-10-19 round: 83.1% (187 of 225 decisions). On the questions whose answer could be calculated, it got it exactly right. Median of 310 ms per case, from Brasília, over the internet; 95% of cases finish within 3.2 s. The table compares only the 11 decision models measured on 2026-10-19. On the two blind tests from September 2026, on new subjects, Iryx scored 86% and 73% (100 questions each).

No labeled data, or a trained classifier?

Iryx fits when categories change, when there is no labeled data, or when you don't want to train a classifier: the categories go in the request itself.

Limits

• The context fits about 6,000 characters with 1 to 4 questions per request, and shrinks with more questions (see Limits).
• 2 to 26 options per pick and 2 to 26 levels per scale.
• Iryx does not write text and does not know facts outside the context you send.
• Iryx is trained in Portuguese, and its published accuracy numbers are in Portuguese. No number has been published for English yet: for now, use it in Portuguese.
• The API is not open to the public yet: access comes with a reservation.

Questions

Do I need labeled examples to start?

No. The categories go in the request, each with a short description. Clear descriptions help more than short names.

Does it work with JSON and structured data?

Yes. The context can be any JSON object. Clear key names help, because they are part of what Iryx reads.

How much does it cost to classify a text?

Billing is per decision, never per token: each answered question counts as one. Plans start at R$ 110 per month (Brazilian reais) for 1,000,000 decisions. With no subscription, pay-as-you-go credit costs R$ 0.120 per thousand decisions. See the plans.

Does it classify English text?

The API accepts any text, but Iryx is trained in Portuguese, and the published accuracy numbers are in Portuguese. No number has been published for English yet: for now, use it in Portuguese.

Sources

• Iryx, Iryx 1.0 numbers
• Iryx, API documentation