Function Call¶
A function call is a request from the model to the calling system (the app using the LLM API) to run a function.
1. Single function call¶
The user's query requires exactly one function call.
"What is the weather like in Tokyo?"
2. Multiple parallel function calls¶
The user's query requires several independent function calls, made at once.
"What is the weather like in Tokyo and Paris?"
- Same tool, many args: "Convert 10, 20, 30 USD to EUR" means 3 parallel calls, or 1 call if the parameter takes an array.
3. Multiple consecutive function calls¶
The user's query requires functions to be called in sequence, because a later call depends on an earlier call's output.
"Create an appointment for XYZ on day A or B."
(First, the model needs to check XYZ's availability on day A and day B.)
- IDs and values must come from the earlier output; never invent them.
4. Parallel then consecutive¶
The model fans out lookups, compares the results, then acts.
"Find the cheapest of hotels A, B and C, and book it."
5. Conditional call¶
The next call depends on a tool result.
"If it rains tomorrow, cancel my picnic booking."
6. Loop / pagination¶
The model follows next_page (or a cursor) until done, and knows when to stop. It must not call again when the output hasn't changed.
7. Retry and fallback¶
- Retry: on a validation error, fix the args and retry once. Never loop forever.
- Fallback: if the primary tool fails, use an alternative tool or explain the failure.
8. Zero-argument call¶
A tool with no parameters is called with {}, e.g. get_current_time().
9. No function call¶
The user's query does not require a function call. Either the query is answered from knowledge, or no available tool supports the action.
"What are the signs of the rain?"
"Book me a flight." (no flight tool)
The model says it can't do this, and never invents a tool.