Function Calling

BRAHMA-30B supports the OpenAI function calling (tools) contract on POST /api/v1/voice-engine/llm/v1/chat/completions. You describe your functions in the tools array; when the model decides a call is needed it answers with finish_reason: "tool_calls" and a message.tool_calls array instead of text. You execute the function yourself, then send the result back in a role: "tool" message for the model’s final answer.

The tools contract

Field	Type	Description
`tools`	array	Function definitions: `{"type": "function", "function": {"name", "description", "parameters"}}` where `parameters` is a JSON Schema object.
`tool_choice`	string or object	`"auto"` (default — the model decides), `"none"` (never call), `"required"` (call some tool), or `{"type": "function", "function": {"name": "..."}}` to force one specific function. `"required"` and forced-function choices are strong steering, not a hard grammar constraint — verify the expected call came back before acting.
`parallel_tool_calls`	boolean	Whether the model may emit several tool calls in one turn. Best-effort on BRAHMA: the model usually emits one call per turn; be prepared to handle multiple entries in `tool_calls` either way.

In the response, each tool_calls[] entry has an id, "type": "function", and a function object whose arguments value is a JSON string (not an object) — parse it before use.

Always validate function.arguments server-side before executing anything: parse the JSON, check required fields and types against your schema, and reject values you would not accept from an untrusted client. The model’s output is a suggestion, not a trusted input.

1. Define tools and let the model call one

curl -X POST https://api.stage.peachdeskai.com/api/v1/voice-engine/llm/v1/chat/completions \
  -H "X-API-Key: bibha_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "brahma-30b",
    "messages": [
      {"role": "system", "content": "You are a support assistant for an appliance store."},
      {"role": "user", "content": "Is order 88412 out for delivery yet?"}
    ],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "get_order_status",
          "description": "Look up the delivery status of a customer order by order id.",
          "parameters": {
            "type": "object",
            "properties": {
              "order_id": {"type": "string", "description": "The numeric order id."}
            },
            "required": ["order_id"],
            "additionalProperties": false
          }
        }
      }
    ],
    "tool_choice": "auto",
    "max_completion_tokens": 256
  }'

The model answers with a tool call — note content: null, the arguments JSON string, and finish_reason: "tool_calls":

{
  "id": "chatcmpl-a1b2c3",
  "object": "chat.completion",
  "model": "brahma-30b",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": null,
        "tool_calls": [
          {
            "id": "call_7f3a",
            "type": "function",
            "function": {
              "name": "get_order_status",
              "arguments": "{\"order_id\": \"88412\"}"
            }
          }
        ],
        "refusal": null,
        "annotations": []
      },
      "logprobs": null,
      "finish_reason": "tool_calls"
    }
  ],
  "usage": {"prompt_tokens": 96, "completion_tokens": 21, "total_tokens": 117}
}

2. Force a specific tool

Use a tool_choice object when the model must call one particular function (for example, an extraction step that always ends in a structured write):

curl -X POST https://api.stage.peachdeskai.com/api/v1/voice-engine/llm/v1/chat/completions \
  -H "X-API-Key: bibha_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "brahma-30b",
    "messages": [
      {"role": "user", "content": "Book me a table for 4 at 7pm tomorrow at the Deccan branch."}
    ],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "create_reservation",
          "description": "Create a restaurant reservation.",
          "parameters": {
            "type": "object",
            "properties": {
              "branch": {"type": "string"},
              "party_size": {"type": "integer"},
              "time": {"type": "string", "description": "ISO 8601 date-time."}
            },
            "required": ["branch", "party_size", "time"],
            "additionalProperties": false
          }
        }
      }
    ],
    "tool_choice": {"type": "function", "function": {"name": "create_reservation"}},
    "max_completion_tokens": 200
  }'

3. Return the tool result and get the final answer

Run the function yourself, then append both the assistant’s tool-call message and a role: "tool" message (with tool_call_id matching the call’s id) and send the whole conversation back:

curl -X POST https://api.stage.peachdeskai.com/api/v1/voice-engine/llm/v1/chat/completions \
  -H "X-API-Key: bibha_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "brahma-30b",
    "messages": [
      {"role": "system", "content": "You are a support assistant for an appliance store."},
      {"role": "user", "content": "Is order 88412 out for delivery yet?"},
      {
        "role": "assistant",
        "content": null,
        "tool_calls": [
          {
            "id": "call_7f3a",
            "type": "function",
            "function": {"name": "get_order_status", "arguments": "{\"order_id\": \"88412\"}"}
          }
        ]
      },
      {
        "role": "tool",
        "tool_call_id": "call_7f3a",
        "content": "{\"status\": \"out_for_delivery\", \"eta\": \"today 17:30\"}"
      }
    ],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "get_order_status",
          "description": "Look up the delivery status of a customer order by order id.",
          "parameters": {
            "type": "object",
            "properties": {"order_id": {"type": "string"}},
            "required": ["order_id"],
            "additionalProperties": false
          }
        }
      }
    ],
    "max_completion_tokens": 256
  }'

The model now answers in plain text (finish_reason: "stop"), e.g. “Yes — order 88412 is out for delivery and should arrive around 5:30 pm today.”

Keep sending the tools array on follow-up turns. The API is stateless — every request must carry the full conversation and the full tool definitions.

Streaming tool calls

With stream: true, tool calls arrive incrementally in delta.tool_calls. Each fragment carries an index identifying which tool call it belongs to; the call id, type, and function.name arrive on the first fragment, and function.arguments arrives as string fragments you concatenate per index:

data: {"id":"chatcmpl-a1b2c3","object":"chat.completion.chunk","model":"brahma-30b","choices":[{"index":0,"delta":{"role":"assistant","tool_calls":[{"index":0,"id":"call_7f3a","type":"function","function":{"name":"get_order_status","arguments":""}}]},"finish_reason":null}]}

data: {"id":"chatcmpl-a1b2c3","object":"chat.completion.chunk","model":"brahma-30b","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"{\"order_id\": "}}]},"finish_reason":null}]}

data: {"id":"chatcmpl-a1b2c3","object":"chat.completion.chunk","model":"brahma-30b","choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\"88412\"}"}}]},"finish_reason":"tool_calls"}]}

data: [DONE]

Accumulation rule: group fragments by tool_calls[].index, assign id/type/name when they appear, and append every arguments fragment in order. When finish_reason: "tool_calls" arrives, each accumulated arguments string is a complete JSON document. Official OpenAI SDKs do this accumulation for you.

Resources

Authentication & Errors

Voice Engine API

Bibha BRAHMA LLM

Bibha VAANI TTS

Bibha LIPI STT

The tools contract

1. Define tools and let the model call one

2. Force a specific tool

3. Return the tool result and get the final answer

Streaming tool calls

​The tools contract

​1. Define tools and let the model call one

​2. Force a specific tool

​3. Return the tool result and get the final answer

​Streaming tool calls

The tools contract

1. Define tools and let the model call one

2. Force a specific tool

3. Return the tool result and get the final answer

Streaming tool calls