How to configure Sarvam Chat Completion API endpoint?

[singhal-shagun]

Hi!

I wanted to configure Sarvam AI API to use their Sarvam-105B model with pi coding agent.

As per the model's documentation here, it supports OpenAI-compatible chat completions format. For example, a multi-turn chat completion request would look something like this:

curl -X POST https://api.sarvam.ai/v1/chat/completions \
  -H "Authorization: Bearer $SARVAM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {"role": "system", "content": "You are a senior legal advisor specializing in Indian corporate law."},
      {"role": "user", "content": "What are the key compliance requirements for a startup in India?"},
      {"role": "assistant", "content": "Key compliance requirements include company registration under the Companies Act 2013, GST registration, PF/ESI registration for employees, and annual filings with MCA."},
      {"role": "user", "content": "Can you elaborate on the annual filing requirements and deadlines?"}
    ],
    "model": "sarvam-105b",
    "temperature": 0.3,
    "top_p": 1,           
    "max_tokens": 2000
  }'

However, I observed that pi's requests were being sent in the following format: :

{
    "model": "sarvam-105b",
    "input": [
        {
            "role": "system",
            "content": "PI_SYSTEM_PROMPT"
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "input_text",
                    "text": "hi"
                }
            ]
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "input_text",
                    "text": "hi"
                }
            ]
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "input_text",
                    "text": "hi"
                }
            ]
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "input_text",
                    "text": "hi"
                }
            ]
        }
    ],
    .
    .
    .

This was resulting in the API returning Error: 400 body.messages.1.user.content : Input should be a valid string.

To make the chat completions request work, I had to intercept the requests using Burp Proxy and change them to the following format:

{
    "model": "sarvam-105b",
    "input": [
        {
            "role": "system",
            "content": "PI_SYSTEM_PROMPT"
        },
        {
            "role": "user",
            "content":"hi"
        },
        {
            "role": "user",
            "content": "hi"
        },
        {
            "role": "user",
            "content": "input_text",
        },
        {
            "role": "user",
            "content": "input_text",
        }
    ],
    .
    .
    .

What configuration changes do I need to make in my Pi setup so that its requests are sent in this format?

Right now, my settings.json is as follows:

{
  "defaultModel": "sarvam-105b",
  "defaultProvider": "sarvam",
  "enableSkillCommands": true,
  "lastChangelogVersion": "0.67.68",
  "packages": [
    "npm:@ollama/pi-web-search"
  ],
  "theme": "dark",
  "retry": {
    "enabled": true
  },
  "showHardwareCursor": false,
  "doubleEscapeAction": "none",
  "treeFilterMode": "default",
  "steeringMode": "one-at-a-time"
}

and my models.json is as follows:

{
  "providers": {
      "sarvam": {
      "baseUrl": "https://api.sarvam.ai/v1",
      "api": "openai-completions",
      "apiKey": "API_KEY_SARVAM",
      "compat": {
        "supportsDeveloperRole": false,
        "supportsReasoningEffort": false
      },
      "models": [
        {
          "id": "sarvam-105b",
          "name": "Sarvam-105B",
          "input": ["text"],
          "contextWindow": 128000,
          "reasoning": false
        }
      ]
    }
  }
}

A more detailed documentation of Sarvam's chat completions endpoint is here.