Chat completions

POST https://api.ai.itsincom.org/v1/chat/completions

The OpenAI chat completions request and response. Two additions: the X-Siati-Tier header chooses the tier of the request, and the X-Zona header in the response says in which zone it was processed.

Request#

curl -i https://api.ai.itsincom.org/v1/chat/completions \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma-4-26b",
    "messages": [
      {"role": "system", "content": "Answer in three sentences."},
      {"role": "user",   "content": "What is data sovereignty?"}
    ],
    "temperature": 0.7,
    "max_completion_tokens": 400
  }'

Required: model, and messages with at least one message. The roles are system, user, assistant and tool. Everything else is optional.

Every parameter we accept is in Every parameter: honoured or refused, with what it does. A parameter that is not on that list is refused with 400, and the error names it.

Python#

import os

from openai import OpenAI

client = OpenAI(
    base_url="https://api.ai.itsincom.org/v1",
    api_key=os.environ["API_KEY"],
)

resp = client.chat.completions.create(
    model="gemma-4-26b",
    messages=[
        {"role": "system", "content": "Answer in three sentences."},
        {"role": "user", "content": "What is data sovereignty?"},
    ],
    max_completion_tokens=400,
    extra_headers={"X-Siati-Tier": "fast"},  # optional
)
print(resp.choices[0].message.content)

JavaScript#

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.ai.itsincom.org/v1",
  apiKey: process.env.API_KEY,
});

const resp = await client.chat.completions.create(
  {
    model: "gemma-4-26b",
    messages: [{ role: "user", content: "What is data sovereignty?" }],
  },
  { headers: { "X-Siati-Tier": "fast" } }, // optional
);
console.log(resp.choices[0].message.content);

Request headers#

Response#

{
  "id": "req_288648c0f938503f1236f4d1",
  "object": "chat.completion",
  "created": 1791028778,
  "model": "gemma-4-26b",
  "choices": [
    {
      "index": 0,
      "message": { "role": "assistant", "content": "…" },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 31,
    "completion_tokens": 67,
    "total_tokens": 98
  }
}

Where it differs from OpenAI:

• id is the identifier under which we record the request, req_…, not a chatcmpl-….
• model is the model_id you asked for.
• There is no system_fingerprint.

When the model calls a function, message.tool_calls carries the calls and content is null if there is no text: see Guaranteed JSON and function calling. With n greater than 1, choices carries every answer; with logprobs, each choice carries its token probabilities.

Response headers#

Zones#

The request goes only to machines in the zones of your key. If a machine fails before answering, the request moves to the next one, still inside your zones. If none is available there, you get 503 with type zone_unavailable and the zone named in the message: the request is never moved to another zone. More in Zones.

Streaming#

Add "stream": true to receive the answer while it is written. The format, what the stream does not carry and how errors arrive are in Streaming.

Images#

The models with the vision badge in the catalogue accept images: today gemma-4-26b. The content of a message becomes a list of parts, text and image_url:

curl https://api.ai.itsincom.org/v1/chat/completions \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gemma-4-26b",
    "messages": [{
      "role": "user",
      "content": [
        {"type": "text", "text": "What document is this? Extract the date and the amount."},
        {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,…"}}
      ]
    }]
  }'

In Python, the data URI of a file:

import base64

with open("invoice.jpg", "rb") as f:
    data_uri = "data:image/jpeg;base64," + base64.b64encode(f.read()).decode()

image_url.url takes two forms:

• a base64 data URI, which we recommend: there is nothing to download;
• a public http or https URL: we download the image ourselves and pass it to the model. It must be reachable from the internet. Private and local addresses are refused with 400, and redirects are not followed, so give the final address. If remote URLs are turned off on ITS INCOM AI, the 400 says so: send a data URI.

The format is read from the bytes, not from the extension or the declared type: valid base64 that does not contain an image returns 400. A part of another type than text or image_url returns 400 too, and so does an image sent to a model that does not accept images; that error lists the models that do.

Cost: the image becomes tokens, counted in prompt_tokens, so you pay for it at the input price of the model.

Errors#

The most common ones. All of them, with what to do, are in Errors.

Price#

Each model has a price for input tokens and one for output tokens, and the tier applies a multiplier. The values are in Pricing.