GPT Function Calling: Patterns and Pitfalls from Production

Function calling turned AI from a text generator into an action taker. After building a dozen production systems using OpenAI's function calling, I've collected patterns that work and pitfalls to avoid.

First, function descriptions matter enormously. The AI decides when to call functions based on descriptions, not names. "Get the current weather" works; "weather" doesn't give enough context. Include parameter constraints—"location must be a city name, not coordinates."

Keep function counts low. More than 10 functions and the AI starts making poor choices or hallucinating non-existent functions. Group related operations into single functions with mode parameters rather than splitting them.

Always validate function outputs before returning them to the model. AI-generated SQL queries can be malicious; user input passed through functions needs sanitization. Treat function calling as an attack surface, not just a convenience feature.