Using Tools in Playground

As enterprises increasingly adopt agentic workflows, tool calling has become a crucial component of many production LLM applications. However, debugging tool calls can be complex, as errors may occur at various stages of the pipeline. Key questions to consider include:

• Did the LLM have access to the necessary function to respond to the input message?

• If the correct function was available, did the LLM choose to invoke it?

• Was the function called with the correct parameter values?

• Were there ambiguities in the function description or parameter definitions?

These challenges highlight the need for a user-friendly interface for debugging. Once tracing is set up for function calls, the playground automatically populates function definitions and outputs when loading in a span or a dataset, facilitating precise iteration and refinement. Additionally, you can define and modify functions directly in the playground interface, enabling rapid prototyping and iteration before making changes to the codebase.

The playground allows for testing assistant tool calls and reviewing JSON outputs for streamlined debugging and refinement.

Multiple functions can be defined in JSON format based on the OpenAI API. The function selection dropdown allows you to choose between 'auto,' 'required,' 'none,' or specific functions for tailored execution.

The original and updated outputs can be compared side-by-side in a full-screen view, displayed in a structured JSON format for easy analysis.