Seven structured text tools

Clef, Clef Flash and Clef Omni

Use questions for probability-based yes/no judgments (type noul) or a choice between two to eight named options. This first integration supports text-only state and those two question types.

{
  "state": "Order T927 has been delivered. The package is blue.",
  "questions": {
    "delivered": {
      "type": "noul",
      "instructions": "Has the order been delivered?"
    },
    "color": {
      "type": "choice",
      "instructions": "Which color is stated?",
      "criteria": {
        "blue": "Blue",
        "red": "Red"
      }
    }
  }
}

The reply is JSON with an answer for each question, its reported probabilities, model identity and measured usage.

Clairic

Use fields to extract text, dates or numbers. Missing values can be null; confidence is a model estimate, not a guarantee.

{
  "state": "Invoice T927: 4 items at 7 dollars each and 2 dollars shipping.",
  "fields": {
    "invoice": {
      "type": "text",
      "instructions": "Invoice reference"
    },
    "total": {
      "type": "number",
      "decimals": 2,
      "instructions": "Total including shipping"
    }
  }
}

Limits and processing

Keep the whole message within 3,000 UTF-8 bytes, state within 2,200 bytes, and use no more than eight fields or questions. Images, URLs as attachments, list fields, estimates and score questions are not enabled in this integration. No extra inference is performed to rewrite invalid input. These tools are never selected by Auto.

All seven use the existing Cloudflare operator Free allowance. Choosing another model does not refill shared capacity. Provider-reported zero output tokens are preserved rather than replaced with made-up usage. Clairic is a preview model and availability is not guaranteed.

Clef API documentation · Clairic documentation and preview terms · Publisher notices · Privacy

M2M100 — translation

Provide source_lang and target_lang as language codes, such as en, es, fr, de, nl or el. Text must be at most 1,200 UTF-8 bytes. Machine translations can be wrong; review important content.

{
  "text": "Hello, how are you?",
  "source_lang": "en",
  "target_lang": "es"
}

DistilBERT — English sentiment

English text only, at most 1,200 UTF-8 bytes. This binary model returns positive/negative scores; it has no neutral class. Sarcasm and demographic bias can produce misleading results. Do not use these scores to make decisions about people. Token usage is not supplied by this provider response; KageMux records one completed request without inventing token counts.

{
  "text": "The service was excellent. I am very happy with the result."
}

BGE Reranker — rank supplied passages

Provide a query up to 240 UTF-8 bytes and 2–8 text passages. The JSON must stay within 2,800 bytes. Returned indices refer to your supplied passages. Scores are relative relevance signals, not calibrated probabilities. No URL is fetched and no search is performed.

{
  "query": "Which document describes refunds?",
  "contexts": [
    {
      "text": "Returns are accepted within 30 days with a receipt."
    },
    {
      "text": "The garden has roses and apple trees."
    }
  ],
  "top_k": 2
}