Decide

Decide is in Beta. It answers typed questions about a text with probabilities instead of prose: which category a letter belongs to, how urgent it is, whether it contains personal data. A decision model writes no text, so there is nothing to parse. Every answer comes with the full probability distribution behind it and a confidence.

Decide runs on komplai's local infrastructure in Denmark, with no AI charge. The text is not stored.

Before you start

Your organisation has Decide in its agreement and has approved its processing, and an administrator has given you access to Decide under Organisation › People › Tool access in the hub.
You have a personal API key (kp_live_...) with the scope decide:process, created on the hub's API keys page at hub.komplai.dk/api-access. An organisation integration key (kp_org_...) gets 403 integration_not_permitted.
Beta operations may still change, with notice. Without an approved data processing agreement, use test data only.

Your first call

Send one text in state and your questions in questions, each under an id you choose.

Shell

curl https://api.komplai.dk/v1/decide \
  -H "Authorization: Bearer $KOMPLAI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "state": "Hej. Jeg vil gerne have en kopi af min ansøgning om byggetilladelse fra marts. Venlig hilsen Mette Kjærgaard Holm",
    "questions": {
      "category": {
        "type": "choice",
        "instructions": "Hvad handler henvendelsen om?",
        "criteria": {
          "aktindsigt": "Borgeren beder om kopi af dokumenter eller aktindsigt",
          "klage": "Borgeren klager over en afgørelse",
          "andet": "Noget andet"
        }
      },
      "urgency": {
        "type": "score",
        "instructions": "Hvor hurtigt skal henvendelsen besvares?",
        "criteria": ["Inden for den normale frist", "Inden for en uge", "Samme dag"]
      },
      "personal_data": {
        "type": "noul",
        "instructions": "Indeholder teksten personoplysninger?",
        "criteria": {
          "true": "Ja, for eksempel et navn eller et CPR-nummer",
          "false": "Nej"
        }
      }
    }
  }'

JSON

{
  "model": "convaiinnovations/laya/multilingual",
  "answers": {
    "category": {
      "type": "choice",
      "choice": "aktindsigt",
      "probabilities": { "aktindsigt": 0.91, "klage": 0.03, "andet": 0.06 },
      "confidence": 0.672,
      "action": { "act_probability": 0.88 }
    },
    "urgency": {
      "type": "score",
      "score": 0.32,
      "max": 2,
      "probabilities": [0.72, 0.24, 0.04],
      "confidence": 0.356
    },
    "personal_data": {
      "type": "noul",
      "noul": 0.97,
      "probabilities": { "true": 0.97, "false": 0.03 },
      "confidence": 0.806
    }
  },
  "usage": { "input_tokens": 118, "output_tokens": 0 },
  "latency_ms": 38
}

The answers come back under the same ids. usage.output_tokens is always 0, and latency_ms is the model's own time without the network.

Question types

Each question has type, instructions (the question in your own words) and criteria, whose shape depends on the type.

Type	`criteria`	The answer
`choice`	An object of 2 to 32 options: a short key and what that option means	`choice` is the key with the highest probability, and `probabilities` gives every key its probability
`score`	A list of 2 to 32 levels, lowest first	`score` is the expected level (each level times its probability, added up), `max` the highest level, and `probabilities` one value per level in the order you sent them
`noul`	`true` and `false`, and what each means	`noul` is the probability that the statement is true

The probabilities in an answer add up to 1. Keep option keys short and machine-friendly, because the answer names the key.

Every answer also has:

confidence: how decided the answer is, from 0 (spread evenly) to 1 (everything on one option). It says nothing about whether the answer is right.
action.act_probability, when the model gives it: how likely the answer is safe to act on without a person checking it.

A request takes from 1 to 64 questions. They share one reading of the text, so several questions cost little more than one. state is a text or a JSON object, which komplai turns into text; arrays and other values are refused. The request may have at most 80,000 characters as JSON and the body at most 320,000 bytes. Fields that are not listed are ignored.

Models

model chooses the decision model:

`model`	Model
`komplai-decide`	Laya
`komplai-decide-claes`	Claes

Leave it out for the default: the model you chose in the hub, else your organisation's default, else komplai-decide. model in the answer names the model that answered by its checkpoint. A model Decide does not offer your organisation answers 409 model_not_enabled, and 404 when local infrastructure does not offer it right now. The API never picks another model for you.

Many texts in one call

POST /v1/decide/batch asks the same questions about up to 100 texts in states, in one pass of the model. The answers come back in results in the order of states, and the batch is answered whole or refused whole. It counts as one request against your request limit.

Shell

curl https://api.komplai.dk/v1/decide/batch \
  -H "Authorization: Bearer $KOMPLAI_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "states": [
      "Hej. Jeg vil gerne have en kopi af min ansøgning om byggetilladelse fra marts.",
      "Jeg klager over afslaget på min ansøgning om boligstøtte. Afgørelsen er forkert, og jeg har brug for svar hurtigt."
    ],
    "questions": {
      "category": {
        "type": "choice",
        "instructions": "Hvad handler henvendelsen om?",
        "criteria": {
          "aktindsigt": "Borgeren beder om kopi af dokumenter eller aktindsigt",
          "klage": "Borgeren klager over en afgørelse",
          "andet": "Noget andet"
        }
      }
    }
  }'

JSON

{
  "model": "convaiinnovations/laya/multilingual",
  "results": [
    {
      "answers": {
        "category": { "type": "choice", "choice": "aktindsigt", "probabilities": { "aktindsigt": 0.91, "klage": 0.03, "andet": 0.06 }, "confidence": 0.672 }
      },
      "usage": { "input_tokens": 61, "output_tokens": 0 }
    },
    {
      "answers": {
        "category": { "type": "choice", "choice": "klage", "probabilities": { "aktindsigt": 0.08, "klage": 0.86, "andet": 0.06 }, "confidence": 0.544 }
      },
      "usage": { "input_tokens": 66, "output_tokens": 0 }
    }
  ],
  "usage": { "input_tokens": 127, "output_tokens": 0 },
  "latency_ms": 52
}

A batch request may have at most 200,000 characters as JSON and the body at most 800,000 bytes.

Time limits and failures

The model has 30 seconds to answer. When it fails, does not answer in time or cannot be reached, the call answers service_unavailable. An answer that fails komplai's checks, such as a missing answer or probabilities that do not add up, is never passed on: the call answers invalid_response. A request that does not have the documented form answers invalid_input.

Where data is processed

Decide runs only on komplai's validated decision models, on komplai's local infrastructure in Denmark. Data does not leave the EU, and your organisation's own AI connections are never used for Decide. The text is not stored; komplai records the call's metadata, such as who called, when and which model.

Reference and errors

Decide reference: every field of POST /v1/decide and POST /v1/decide/batch, with code samples and a console to try each call.
Errors: every code with what to do. Request limits are on Limits.