Draft preview. Not yet published. Items marked TODO are unconfirmed.

The SauerkrautLM-decision API gives hosted access to VAGO's decision models at api.sauerkrautlm.de. You send an input (the state), optional images and one or more typed questions; the API returns one structured answer per question as probabilities, with zero generated tokens. To run the open-weight model yourself instead, see SauerkrautLM-decision.

  • No GPU to manage

    VAGO hosts and scales the models. You call one endpoint.

  • Text and images

    Evaluate text, JSON and images together in the same request.

  • Pay for input only

    Billed per million input tokens. Output is always zero tokens and free.

Models

Choose the hosted model with the model field. Both models share the same endpoint, question types and response format.

ModelInputsDescription
sauerkrautlm-decision-smallText, imagesDefault hosted model. Fast and low-cost for high-volume routing and classification. TODO: placeholder ID
sauerkrautlm-decision-largeText, imagesLarger model for harder or more nuanced decisions. TODO: placeholder ID

The examples below use sauerkrautlm-decision-small. Switch to sauerkrautlm-decision-large for harder decisions.

Pricing

You pay only for input tokens. Output is always zero tokens and is not billed. Each question is billed for its own text plus all images in the request.

ModelPrice per 1M input tokensOutput
sauerkrautlm-decision-small€0.10 (TODO: placeholder price)Free
sauerkrautlm-decision-large€0.40 (TODO: placeholder price)Free

TODO: link to the VAGO pricing page and any free credits.

Setup

  1. Sign in to the VAGO console at console.sauerkrautlm.de.
  2. Create an API key. TODO: menu path. Keys start with vago_.
  3. Store the key in an environment variable:
bash
export VAGO_API_KEY="vago_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"

Your first call

A request has these fields:

FieldDescription
modelWhich hosted decision model to use.
stateThe input to evaluate: plain text, or a JSON object with structured fields.
questionsNamed questions. Each has a type (noul, choice or score), instructions, and for Choice and Score, criteria.
imagesOptional. Images to evaluate with the state. See Image inputs.

Endpoint: POST https://api.sauerkrautlm.de/v1/systemone

bash
curl -s https://api.sauerkrautlm.de/v1/systemone \
  -H "Authorization: Bearer $VAGO_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sauerkrautlm-decision-small",
    "state": "Ich warte seit drei Wochen auf meine Bestellung und niemand antwortet auf meine E-Mails.",
    "questions": {
      "is_complaint": {
        "type": "noul",
        "instructions": "Is this message a complaint?"
      }
    }
  }'

Response:

json
{
  "model": "sauerkrautlm-decision-small",
  "answers": {
    "is_complaint": {
      "type": "noul",
      "noul": 0.97
    }
  },
  "usage": {
    "input_tokens": 61,
    "output_tokens": 0
  }
}

The numbers above are illustrative, not measured.

  • answers.<name>: one answer per question, keyed by the question name.
  • usage.input_tokens: total input tokens for all questions, including image tokens.
  • usage.output_tokens: always 0. Decision models do not generate tokens.

Question types

Type noul. Returns one probability between 0 and 1 that the answer is "yes". Threshold it, use it as a weight, or pass it on.

json
"is_harmful": {
  "type": "noul",
  "instructions": "Does this message contain threatening or abusive content?"
}

Answer:

json
"is_harmful": { "type": "noul", "noul": 0.98 }

All answer values on this page are illustrative.

Combining questions

Ask a Yes/No, a Choice and a Score question in the same request to classify, route and prioritise one input in a single round trip:

python
answers = result["answers"]

is_bug = answers["is_bug"]["noul"] > 0.7
team = answers["team"]["choice"]
urgency = answers["urgency"]["score"]

if is_bug and urgency > 1.5:
    print(f"Escalate to {team} on-call")
elif answers["team"]["confidence"] < 0.5:
    print("Routing unclear: send to triage")
else:
    print(f"Queue in {team}")

Thresholds here are examples. Tune them on your own data.

Image inputs

Both hosted models accept images. Add an images array next to model, state and questions; question types and the response format stay the same. Every question sees every image, in the order given, so refer to images by position ("the first image", "the second image").

Each image is either a Base64 data URL or an object with content_type and base64. Remote image URLs are not accepted: download and encode the file first.

bash
IMAGE_BASE64=$(base64 < board.jpg | tr -d '\r\n')

curl -s https://api.sauerkrautlm.de/v1/systemone \
  -H "Authorization: Bearer $VAGO_API_KEY" \
  -H "Content-Type: application/json" \
  --data-binary @- <<JSON
{
  "model": "sauerkrautlm-decision-small",
  "state": "Kamerabild einer Leiterplatte aus der Fertigungslinie.",
  "images": ["data:image/jpeg;base64,$IMAGE_BASE64"],
  "questions": {
    "defect": {
      "type": "noul",
      "instructions": "Does this circuit board have a visible defect?"
    }
  }
}
JSON

Image tokens count toward usage.input_tokens and are billed at the model's input rate, once per question. TODO: image token formula, supported formats, maximum images per request and maximum request size.

Errors and limits

TODO: error response format and status codes, rate limits, maximum request size and maximum number of questions per request.