Extras
Inference and tool use
Overview
ToolUse
class automates the process of using tools by LLM, i.e.:
- calling LLM with provided context (message sequence),
- extracting tool calls requested by LLM from the response,
- calling the requested tool and storing its results,
- constructing message sequence with the result of call,
- sending updated message sequence back to LLM.
This cycle is repeated until one of the exit criteria is met:
- LLM no longer requests any tool calls,
- specified maximum number of iterations is reached,
- specified token usage limit is reached
- there are any errors during the process (e.g. LLM requested a tool that is not available).
ToolUse
class provides 3 ways to iterate through the process:
- manual control - code is responsible for checking
hasNextStep()
and callingnextStep()
in a loop, - using iterator - code uses foreach loop to iterate through the steps (internally it checks
hasNextStep()
and callsnextStep()
), - just get final step - you only get the final step, iteration process is done internally.
Example
This example demonstrates 3 ways to use ToolUse
class to allow LLM call functions
if needed to answer simple math question. We provide 2 functions (add_numbers
and
subtract_numbers
) as tools available to LLM and specify the task in plain language.
The LLM is expected to call the functions in the correct order to get the final result.