← Volver a las guías

Tool Calling Troubleshooting: Fix the Full API Loop

Understanding Tool Use and Function Calling

Function calling lets a model request an operation through a structured interface, such as checking inventory, fetching an order, or querying a database. Tool use is the broader concept; function calling is one common implementation. The critical distinction: the model proposes a call, but your application validates the arguments, executes the operation, and returns the result.

Most integration failures happen in that handoff, not in the model’s reasoning. Use the checks below to troubleshoot the complete request-and-response loop.

Start with a Minimal, Complete Tool Loop

• Define one read-only function. Start with get_order_status and a JSON Schema object containing a required string field named order_id. Explain what the function returns and when to use it.
• Send the tool definition with the conversation. In OpenAI Chat Completions, tools contains entries with type set to function and a function object containing name, description, and parameters.
• Inspect the response for tool calls. Do not assume the answer lives in message.content. A tool-requesting assistant message can have no text content.
• Validate and execute. Parse the function’s arguments as JSON, check their types and permitted values, and dispatch only to an explicitly registered handler.
• Return the result and continue. Preserve the assistant message containing tool_calls. Then append a role: tool message with the matching tool_call_id and a string result, commonly serialized JSON. Send the updated history to obtain the final answer.

These message shapes describe Chat Completions. OpenAI Responses and native Claude interfaces use different structures. Do not mix protocols within one conversation.

Why Is the Model Not Calling My Function?

First, confirm that the selected model and endpoint support tools. Then inspect the actual outgoing request: SDK wrappers sometimes omit tools during retries or conversation updates.

Make descriptions actionable. “Gets data” is vague; “Returns the shipping status for an existing order ID” gives the model a clear selection rule. Ask a test question that genuinely requires external information. Where supported, use tool_choice to require a tool or select a specific function during debugging. Restore automatic selection when a direct answer is appropriate.

Why Are Function Arguments Invalid?

Treat arguments as untrusted input, even when they look plausible. Check required fields, enums, numeric limits, and unknown properties before execution. Never pass generated strings directly into shell commands or SQL.

Where supported, strict schema mode improves structural compliance. For OpenAI strict function schemas, objects require additionalProperties: false, and all declared properties must be required; represent optional values with nullable types. Strict mode does not establish authorization or factual correctness.

Why Does the API Reject My Tool Result?

Check conversation ordering and call identifiers. Each tool result must correspond to an actual pending call. If the assistant requests several tools, return a result for each before requesting its next answer.

For Chat Completions, serialize objects into the tool message’s string content. Keep errors structured and useful, for example with an error category and retryable flag, without exposing credentials. Return explicit failures rather than inventing successful results.

Why Does Streaming Break Argument Parsing?

Streaming can deliver function names, identifiers, and arguments across multiple chunks. Accumulate fragments by tool-call index, then parse the completed arguments. A partial JSON fragment is not a malformed final request. Log call IDs, selected function names, validation failures, latency, and execution status; redact secrets and personal data.

FAQ: Safety, Costs, and 59API Setup

Can a model execute tools by itself? Not through ordinary function calling. Your application supplies execution. Require authorization and confirmation for payments, deletions, or other consequential actions.

How can I control costs? Limit tool-loop iterations, cap result sizes, and avoid repeatedly sending large schemas or database dumps. Treat retries carefully: use idempotency controls so a repeated request cannot duplicate a purchase.

Where does 59API fit? 59API provides low-cost, pay-as-you-go access to Claude and GPT models using native, official-quality models rather than downgraded substitutes. It supports Claude Code, Codex, and OpenAI SDKs. Use https://api.59api.com as the relay host; confirm the required endpoint prefix and model identifier before testing.

Is it worth trying for development? Its inexpensive pricing and referral rebate make it attractive for iterative debugging. Sign up for 59API and start with one read-only tool to verify your integration before expanding it.

Precios · Docs