Chat Completions

Send a multi-turn conversation to a chat model and receive an assistant message. This is the primary endpoint for LLMs and is fully compatible with the OpenAI Chat Completions API.

POST/v1/chat/completions

Request body

📋Request

{
  "model": "pkrd/cosmosrp-3.5",
  "messages": [
    { "role": "system", "content": "You are a helpful assistant." },
    { "role": "user", "content": "Hello!" }
  ],
  "temperature": 0.7,
  "max_tokens": 1024,
  "stream": false
}

Parameters

model

stringrequired

Model ID. Optional if the model is in the URL path.

messages

arrayrequired

Conversation history. Each item has role and content.

stream

boolean

If true, response is sent as SSE chunks. Default: false.

temperature

number

Sampling temperature. Range and default depend on the model.

top_p

number

Nucleus sampling probability mass.

top_k

integer

Top-K sampling (when supported).

min_p

number

Minimum probability cutoff (when supported).

max_tokens

integer

Maximum tokens to generate.

frequency_penalty

number

Penalize repeated tokens (-2 to 2).

presence_penalty

number

Penalize tokens already in the conversation (-2 to 2).

repetition_penalty

number

Alternative repetition control (1.0 = no penalty).

stop

string | array

Up to 4 stop sequences.

seed

integer

Deterministic sampling seed.

n

integer

Number of completions per request.

logprobs

boolean | integer

Return per-token log probabilities.

tools

array

List of tool/function definitions for tool-calling models.

tool_choice

string | object

"auto", "none", "required", or a forced tool selection.

response_format

object

{ type: "json_object" } to force JSON output.

reasoning_effort

string

"low" | "medium" | "high" for reasoning models.

user

string

End-user identifier passed through for analytics.

Message roles

system

Sets behavior, persona and constraints. Usually first.

user

Input from the end-user.

assistant

Previous model replies, for multi-turn context.

tool

Result of a tool call, paired with the original tool_call_id.

Vision input

Models with the vision capability accept image parts in user messages. Pass an HTTPS URL or a data: base64 URL.

📋Vision request

{
  "model": "pkrd/cosmosrp-3.5",
  "messages": [
    {
      "role": "user",
      "content": [
        { "type": "text", "text": "What is in this image?" },
        {
          "type": "image_url",
          "image_url": { "url": "https://example.com/cat.jpg" }
        }
      ]
    }
  ]
}

Tool / function calling

Models with the toolCalling capability can choose to call a tool you declare. The model returns tool_calls in the assistant message; you execute the tool and feed the result back as a tool message.

📋Tools request

{
  "model": "openai/gpt-oss-20b",
  "messages": [
    { "role": "user", "content": "What is the weather in Paris?" }
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Get current weather for a city",
        "parameters": {
          "type": "object",
          "properties": {
            "city": { "type": "string" }
          },
          "required": ["city"]
        }
      }
    }
  ],
  "tool_choice": "auto"
}

Reasoning

Models with thinking support a reasoning_effort parameter. The hidden chain-of-thought may be exposed via reasoning field on the response message.

📋Reasoning request

{
  "model": "pkrd/cosmosrp-3.5:msr",
  "messages": [
    { "role": "user", "content": "Solve: if 3x + 5 = 20, what is x?" }
  ],
  "reasoning_effort": "medium"
}

Streaming

Set stream: true to receive Server-Sent Events. See the Response Format page for chunk structure.

💻Streaming curl

curl https://api.pawan.krd/v1/chat/completions \
  -H "Authorization: Bearer pk-your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "pkrd/cosmosrp-3.5",
    "messages": [{ "role": "user", "content": "Tell me a story" }],
    "stream": true
  }'

✓

Per-model parameter ranges

Each model publishes its own valid ranges for these parameters. Check the model on the Models page or via the List Models endpoint to discover supported values.

Rate Limits

Text Completions