LLM API for function calling

Function calling in a production application

A sales assistant can look up an account, calculate a quote, and create a draft order. The user's request should call lookup_account first, then calculate_quote. create_order is unavailable until a human confirms the exact draft.

The model proposes tool calls; the application validates schemas, authorization, and state before execution. Tool errors are returned as structured results, and identical invalid calls are not retried.

How the function calling workflow operates

• Expose only tools allowed in the current state.
• Let the model select a tool and propose arguments.
• Validate every argument and permission in application code.
• Execute with an idempotency key where writes are possible.
• Return a structured result and continue until a bounded stop condition.

Call gpt-5.6-terra through LLMFly AI

Send the request from your server. Replace the example content and placeholder tool schema with data and tools from your application.

curl https://app.llmfly.ai/v1/chat/completions \
  -H "Authorization: Bearer $LLMFLY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-terra",
    "messages": [
      {"role": "system", "content": "Use the available tools to prepare a quote. Never create an order until the user confirms the exact draft. Do not invent account IDs or prices."},
      {"role": "user", "content": "Prepare a quote for 25 seats for Acme's current plan."}
    ],
    "tools": [{"type": "function", "function": {"name": "YOUR_TOOL", "description": "Replace with the tool contract for this workflow", "parameters": {"type": "object", "properties": {}, "additionalProperties": false}}}]
  }'

Why gpt-5.6-terra is the primary model

Terra is the default for reliable tool selection and multi-step argument construction without using the flagship on every call.

Use gpt-5.4 for one-tool read-only requests. Escalate to Sol only for ambiguous plans involving several dependent tools.

Acceptance checks for function calling

Failures to handle before deployment

• Executing unvalidated arguments
• Retrying the same invalid call
• Using overlapping tool descriptions
• Letting text bypass permissions

Returned output and run records

Validated tool calls followed by a draft quote; no external order creation before confirmation.

Record the model ID, request ID, token usage, retries, validation result, and final disposition for every production run.