Decisions
API

Models and price

The first 1,000 requests a day are free. The quota belongs to the account, is shared by both models and resets at 00:00 UTC. A request costs the same whatever the size of its state and however many questions it carries. Requests that fail are not charged and do not use the quota. Create a key under settings → API keys.

Endpoints

Authenticate with Authorization: Bearer skrm-.... Decision models are not chat models: they do not appear in /v1/models and /v1/chat/completions rejects them.

Request

curl https://www.rawmodels.ai/api/alpha/decisions \
  -H "Authorization: Bearer $RAWMODELS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "decide",
    "state": "My checkout page shows a blank screen after I click Pay.",
    "questions": {
      "is_bug": {
        "type": "noul",
        "instructions": "Is the customer reporting a software defect?",
        "criteria": {
          "true": "The customer describes broken or unexpected product behavior.",
          "false": "The customer is asking a question or requesting a feature."
        }
      },
      "team": {
        "type": "choice",
        "instructions": "Which team should own this ticket?",
        "criteria": {
          "payments": "Checkout, billing, or payment processing issues.",
          "frontend": "Rendering, layout, or browser compatibility issues."
        }
      },
      "urgency": {
        "type": "score",
        "instructions": "How urgent is this ticket?",
        "criteria": ["Can wait for the next release", "Should be fixed this week", "Blocking revenue right now"]
      }
    }
  }'

Question types

Every question also takes instructions: the question itself, in plain words. Write criteria as full sentences; each one is read against the state on its own.

Response

{
  "id": "gen-dec-1790677555-3f9a1c0b7d2e4a6f8b1c",
  "model": "decide",
  "provider": "rawmodels",
  "answers": {
    "is_bug": { "type": "noul", "noul": 0.9685 },
    "team": {
      "type": "choice", "choice": "payments", "confidence": 0.3881,
      "probabilities": { "payments": 0.6136, "frontend": 0.3864 }
    },
    "urgency": {
      "type": "score", "score": 1.94, "confidence": 0.8,
      "probabilities": { "0": 0.0017, "1": 0.0529, "2": 0.9454 },
      "legend": { "0": "Can wait for the next release", "1": "Should be fixed this week", "2": "Blocking revenue right now" }
    }
  },
  "probability_method": "normalized_entailment_v1",
  "usage": { "input_tokens": 487, "output_tokens": 0, "cost": 0 }
}

probabilities are the model's entailment scores normalised over the options of one question. They are a forecast, not a fitted calibration. confidence is one minus the normalised entropy of that distribution. usage.cost is what the request was charged, in USD: zero inside the free quota. The X-Free-Requests-Remaining header says how much of today's quota is left.

Images

{
  "model": "decide",
  "state": "An image: <<IMG>>",
  "image_data": "<base64 or data:image/png;base64,...>",
  "questions": {
    "damaged": {
      "type": "noul",
      "instructions": "Is the parcel visibly damaged?",
      "criteria": { "true": "The packaging is torn, crushed or wet.", "false": "The packaging is intact." }
    }
  }
}

One image per request, shared by all questions. The model receives the pixels, not a caption. URLs are not fetched. <<IMG>> marks where the image sits in the state and is appended when absent. Image requests take longer than text, and the first request at a new resolution longer still.

Video

Send one clip in video_data in place of image_data: MP4, WebM or Ogg, base64 or a data URI, up to 5.5 MiB and ten minutes. Six evenly spaced frames are decided on separately and merged: an option is as likely as its best frame makes it. The response carries answers for the clip and frames, the answers of every frame with its seconds. A video request costs the same as any other and usually takes 2 to 12 seconds.

Rate limit

5 requests a second per API key, counted per model. Every response carries X-RateLimit-Limit and X-RateLimit-Remaining. Over the limit you get 429 with Retry-After: 1; the refused request is not charged and does not use the quota. The per-minute limit of the chat API does not apply here.

Errors

{ "error": { "code": 429, "message": "Rate limit exceeded", "metadata": {} } }

TypeSafe SDK

The SDK takes a base URL, so it works unchanged:

import { TypeSafeClient } from '@typesafe-ai/sdk';

const client = new TypeSafeClient({
  apiKey: process.env.RAWMODELS_API_KEY,
  baseURL: 'https://www.rawmodels.ai/api',
});

Limits of the models

They are not hardened against prompt injection and they are weak at date arithmetic. Do not make one the only check in front of a destructive action. Benchmarks against Jev, the training recipe and the measured failures are in the launch post.