Decision Models

POST /v1/systemone answers typed questions about a piece of text: which of these options applies, or how far along a scale something sits. It is compatible with TypeSafe's System One API, and it runs the same decision model behind reranking, tuned for agent memory.

Use it where you would otherwise prompt an LLM for a one-word judgement and parse the reply: classifying a message, picking the best of several candidates, or grading how much of a list is relevant. You get a structured answer with probabilities, at a fraction of the cost and latency.

Model: hs-decide-1-flash

Quick Start

curl https://inference.hindsight.vectorize.io/v1/systemone \
  -H "Authorization: Bearer $HINDSIGHT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "hs-decide-1-flash",
    "state": "Question: When is the offsite?",
    "questions": {
      "answer": {
        "type": "choice",
        "instructions": "Which note answers the question?",
        "criteria": {
          "c0": "The offsite moved to April 17.",
          "c1": "Priya prefers window seats."
        }
      }
    }
  }'

{
  "model": "hs-decide-1-flash",
  "answers": {
    "answer": {
      "type": "choice",
      "choice": "c0",
      "confidence": 0.9,
      "probabilities": { "c0": 0.95, "c1": 0.05 }
    }
  },
  "usage": { "input_tokens": 120, "output_tokens": 0 }
}

A client already written against TypeSafe's System One API works by changing its base URL to https://inference.hindsight.vectorize.io and using a Hindsight API key.

Request

Field	Type	Required	Description
state	string	Yes	The text the questions are about: a conversation, a document, a question with candidate answers
questions	object	Yes	Question name → question. Answers come back under the same names
model	string	No	hs-decide-1-flash (the default)

Each question:

Field	Type	Description
type	string	choice or score
instructions	string	What to decide, in plain words
criteria	object or array	For choice, an object of label → option text (at least two). For score, an array of levels, lowest first

The request body can be up to 2 MB.

Question Types

choice: which option applies

The model picks one of the labelled options and reports a probability for each. The probabilities sum to 1, so they also rank the options against each other.

"intent": {
  "type": "choice",
  "instructions": "What does the user want?",
  "criteria": {
    "book": "To book a meeting",
    "cancel": "To cancel a meeting",
    "other": "Something else"
  }
}

Answer:

Field	Description
choice	The label picked
confidence	How sure the model is of that pick
probabilities	Label → probability, for every option

score: where on a scale

The model places the state on an ordered scale you describe, lowest level first.

"depth": {
  "type": "score",
  "instructions": "How far down this ranked list does genuine relevance to the question extend?",
  "criteria": [
    "Only the first candidate is relevant",
    "The first two are relevant",
    "The first five are relevant",
    "All of the listed candidates are relevant"
  ]
}

Answer: score, the position on the scale counting from 0 for the first level. It can fall between two levels, so round it to pick one.

Response

Field	Description
model	The model that answered
answers	Question name → answer, as described above. Each answer also repeats its type
usage.input_tokens	The tokens billed
usage.output_tokens	Not billed

Errors

Besides the common errors:

Status	Meaning
400	The model can't answer this request as asked. The body gives its reason
413	The body is over 2 MB
422	The request is malformed: a missing or mistyped field. The body says which
429	The model is busy. Retry after the Retry-After header (seconds)

Billing

Decision requests cost $0.02 per 1M input tokens, the same rate as reranking. Output is free. usage.input_tokens reports exactly what was billed. See Pricing & billing.