Chat completions endpoint

Super Decider also answers through the OpenAI-compatible chat completions endpoint, so you can call it from an existing OpenAI client. It makes the same decisions as /v1/systemone; only the request shape differs.

POST https://inference.asicloud.cudos.org/v1/chat/completions

Messages

Send exactly two messages:

Role	Content
system	The question schema, as a JSON string.
user	The state, as a JSON string. A plain string must be JSON-encoded too, for example with json.dumps. Raw text returns 400. For images, see below.

Question schema

{"questions": [
  {"id": "next_step", "type": "choice", "instructions": "What should support do next?",
   "options": [{"name": "refund_now", "description": "Refund without a return"},
               {"name": "start_return", "description": "Start a return"},
               {"name": "escalate", "description": "Hand over to a human"}]},
  {"id": "allowed", "type": "noul", "instructions": "Is cancelling this order allowed by the policy?"},
  {"id": "urgency", "type": "score", "instructions": "How urgent is this ticket?",
   "levels": ["low", "medium", "high"]}
]}

Compared with /v1/systemone:

• questions is an array, and each question carries its own id.
• A choice lists its options in options, each with a name and an optional description.
• A score lists its levels in levels.
• A span puts max_tokens and max_items directly on the question.
• The request's thinking field is ignored. Control thinking with schema keys: "adaptive": false for no thought, or "think": 512 for a thought of up to 512 tokens on every question. See Thinking.

{"think": 512, "questions": [...]}

Example

Python

import json
import os
from openai import OpenAI

client = OpenAI(
    base_url="https://inference.asicloud.cudos.org/v1",
    api_key=os.environ["ASI_API_KEY"],
)

schema = {"questions": [
    {"id": "next_step", "type": "choice", "instructions": "What should support do next?",
     "options": [{"name": "refund_now", "description": "Refund without a return"},
                 {"name": "start_return", "description": "Start a return"},
                 {"name": "escalate", "description": "Hand over to a human"}]},
    {"id": "allowed", "type": "noul", "instructions": "Is cancelling this order allowed by the policy?"},
    {"id": "urgency", "type": "score", "instructions": "How urgent is this ticket?",
     "levels": ["low", "medium", "high"]},
]}
state = (
    "Order 8812 shipped on Oct 1. The customer writes: "
    '"I want to cancel and get my money back." '
    "Policy: shipped orders cannot be cancelled; they can be returned "
    "within 30 days of delivery for a full refund."
)

r = client.chat.completions.create(
    model="super-decider",
    messages=[
        {"role": "system", "content": json.dumps(schema)},
        {"role": "user", "content": json.dumps(state)},
    ],
)
answers = json.loads(r.choices[0].message.content)["answers"]

cURL

curl https://inference.asicloud.cudos.org/v1/chat/completions \
  -H "Authorization: Bearer $ASI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "super-decider",
    "messages": [
      {"role": "system", "content": "{\"questions\": [{\"id\": \"allowed\", \"type\": \"noul\", \"instructions\": \"Is cancelling this order allowed by the policy?\"}]}"},
      {"role": "user", "content": "\"Order 8812 shipped on Oct 1. Policy: shipped orders cannot be cancelled.\""}
    ]
  }'

Note the escaped quotes: both message contents are JSON strings, and the state is a JSON-encoded string inside them.

Response

choices[0].message.content is compact JSON with the answers only. Each answer has the same fields as on /v1/systemone, plus a ready label:

{"answers": {
  "next_step": {"type": "choice", "choice": "start_return", "probabilities": {...}, "confidence": 0.96, "label": "start_return"},
  "allowed": {"type": "noul", "noul": 0.04, "label": "no", "confidence": 0.96, "probabilities": {"yes": 0.04, "no": 0.96}},
  "urgency": {"type": "score", "score": 0.66, "legend": {...}, "probabilities": {...}, "confidence": 0.61, "label": "medium"}
}}

probabilities and legend are shortened to {...} here; they have the same shape as on /v1/systemone.

Type	label
choice	The chosen option.
score	The most likely level.
noul	yes when noul is at least 0.5, otherwise no. Noul answers also carry confidence and probabilities here.

The response's usage has the token counts. completion_tokens includes the thought when the request thinks.

Images

Put images in the user message as a list of image_url parts, with an optional text part for context:

{"role": "user", "content": [
    {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{img}"}},
    {"type": "text", "text": "Scanned document from the accounts payable inbox."},
]}

See Images for limits and a full example.