tech-studies

Gemini Function Calling

2 min read

Gemini Function Calling

Google Gemini’s take on llm-native-function-calling: you declare function_declarations, Gemini replies with a functionCall part, and you answer with a functionResponse part.

Why it matters

It’s the dialect for Gemini-based agents (AI Studio and Vertex AI), and it bakes in two things the others bolt on: a built-in automatic function calling mode in the Python SDK (pass real Python functions; the SDK runs the loop for you), and first-class modes that force or suppress calling. Schemas use a subset of the OpenAPI 3.0 spec rather than raw JSON-Schema.

How it works

Content is a list of parts. When Gemini wants a tool, a part contains a functionCall with name and args (a parsed object). You reply by sending a functionResponse part carrying the result back into the conversation.

  • function_calling_config.mode — AUTO (model decides), ANY (must call; optionally restrict to allowed_function_names), or NONE (text only).
  • Automatic mode (Python SDK) — give the SDK callable functions and it intercepts functionCall, executes, and resends the functionResponse without you writing the loop. Disable it for manual control.
  • Parallel & compositional — Gemini supports multiple functionCall parts at once, and compositional calling where one call’s output feeds the next.
  • Schema — OpenAPI 3.0 subset: type, properties, required, enum; some JSON-Schema keywords are unsupported.
OpenAIGemini
tool_callsfunctionCall part
arguments (string)args (parsed object)
tool_choice: requiredmode: ANY

Example

# manual loop
resp = model.generate_content(prompt, tools=[weather_tool])
part = resp.candidates[0].content.parts[0]
if part.function_call:
    out = run_tool(part.function_call.name, dict(part.function_call.args))
    resp = model.generate_content([
        prompt, part,
        {"function_response": {"name": part.function_call.name,
                               "response": {"result": out}}}])

Pitfalls

  • Unsupported schema keywords — it’s OpenAPI 3.0, not full JSON-Schema; anyOf/$ref-heavy schemas may be rejected or ignored.
  • Automatic mode hides failures — convenient, but exceptions in your functions and silent retries are easy to miss; disable it when debugging or in production loops you must control.
  • Sending results wrong — the functionResponse must echo the name and wrap the payload (commonly under a result key); a bare value confuses the model.
  • Mode confusion — leaving AUTO when you need a guaranteed call lets Gemini answer from memory; use ANY with allowed_function_names.

See also

Graph View

Backlinks