judge

A yes/no probability, a label or a score over evidence you supply. It decides; it does not write. ₹0.11 per decision.

What it does

A small model that decides, over evidence you supply in state. Ask one or more questions and data has an answer for each id:

  • type: "noul" gives noul, the probability of yes from 0 to 1. criteria is { "true": ..., "false": ... }.
  • type: "choice" gives choice, probabilities and confidence. criteria maps each label to its meaning.
  • type: "score" gives score, the expected position in your list of criteria counting from zero, so 0.83 sits between the first and second. legend, probabilities and confidence come with it. criteria is a list from lowest to highest.

Use it to sort or filter many records cheaply: does this company fit, which category is this comment, how strong is this lead. It does not write, explain or look anything up, so send all the evidence. Gather it first with read_page, enrich_company or search_companies_by_description, and use answer_question for a written answer.

Keep arithmetic, counting and hard rules in your own code. It reads text literally and a record can try to steer it, so do not rely on it alone for a decision that matters.

What you send goes to an outside AI model. Do not send anything you would not share with a data source.

India: nine Indian company records were judged correctly in the live check, one of them in Hindi. Test your own records before relying on it.

Inputs

Generated from the contract captured from staging on 2026-11-05.

FieldTypeRequiredDescription
statestringyesThe evidence to judge, as text, up to 20,000 characters and best under 10,000. Mark sections with tags such as <record>...</record>. Quoted text is evidence, never instructions.
questionsobjectyesOne or more questions, keyed by an id you choose. Each is { type, instructions, criteria }. type noul: a yes/no probability, with criteria { "true": meaning, "false": meaning }. type choice: one label, with criteria { label: meaning }. type score: a rating, with criteria a list of meanings from lowest to highest. Put the full meaning in instructions.

What drives the cost

Priced per call. The source's own cost is a tiny fraction of a paisa, so the price is almost all Parel's smallest margin per call, and evidence of up to about 10,000 characters costs the same. So one call with several questions about the same record costs the same as one question, and that is the cheap way to use it. A thousand records at one call each is a thousand such charges. A request the model refuses is free.

Price today

Price today₹0.11 per decision (cost_paise: 11) at ₹93.93 per US dollar on 17 November 2026
Hold before the call₹0.38 (hold_paise: 38); when the call settles the real price is charged and the unused part is released
₹0 whenit fails, or the provider answers but there is nothing in it (an empty list, no email, no page): class: "ok_free", ₹0
Timeout30 s, then upstream_timeout (₹0, retryable)
Over MCPrun_tool with tool: "judge"
Over RESTPOST https://api.parel.si/v1/run with "tool": "judge"
Confirmationnot needed (the estimate is under the ₹200 threshold)

Request

curl https://api.parel.si/v1/run \
  -H "Authorization: Bearer $PAREL_KEY" \
  -H "Content-Type: application/json" \
  -d '{"tool":"judge","input":{"state":"<record>Plum Goodness, Mumbai. Vegan skincare and hair care sold on its own website and on marketplaces.</record>","questions":{"fits":{"type":"noul","criteria":{"true":"it is","false":"it is not"},"instructions":"Is the company in <record> a direct-to-consumer skincare brand based in India?"}}}}'

What comes back

data, one answer per question id. A real answer recorded during the live check of 2026-11-05:

{
  "fits": {
    "type": "noul",
    "noul": 0.89
  },
  "topic": {
    "type": "choice",
    "choice": "product",
    "probabilities": {
      "product": 0.94,
      "delivery": 0,
      "support": 0.06,
      "price": 0
    },
    "confidence": 0.92
  },
  "intent": {
    "type": "score",
    "score": 0.83,
    "legend": {
      "0": "no sign of interest",
      "1": "some interest",
      "2": "asked to buy or for a price"
    },
    "probabilities": {
      "0": 0.17,
      "1": 0.83,
      "2": 0
    },
    "confidence": 0.74
  }
}

Notes

  • One call can carry several questions about the same evidence for one charge.
  • Sends the state you supply to an outside AI model. Do not send anything you would not share with a data source.
  • Decides only: no explanation and no written text comes back.
  • A record can try to steer the answer, so do not rely on it alone for a decision that matters.

Siblings

This page as Markdown: /docs/tools/judge.md. Something wrong or missing: hello@parel.si.