Trace Function Calls

Overview

Function Calling enables LLMs to interact directly with external tools and APIs, making AI useful for real-world tasks like content generation and system operations. However, debugging is challenging due to the complexity of multiple tools, unclear errors, and incorrect parameter handling.

Arize streamlines this process by logging chat history and function calls to the platform with just a single line of code. Users can easily review function call traces in a human-readable format, load chats and calls into the prompt playground, and refine parameters to optimize workflows — all within a user-friendly UI.

User flow for tracing function calls to Arize and running experiments in the playground to improve LLM function calling performance.

Quickstart Guide: Trace Function Calls to Arize

Refer to the following QuickStart guide to get started with tracing function calls in Open AI using auto instrumentation.

Google Colab

Tutorial notebook for tracing Open AI function calls with auto instrumentation and replaying in the prompt playground.

By adding a single line of code to auto instrument your Open AI chat, the entire chat history (including function calls) will be traced to the Arize platform. You can view your traces to debug the chat history and review whether the LLM called the appropriate functions with the expected parameters.

from openinference.instrumentation.openai import OpenAIInstrumentor
import openai

OpenAIInstrumentor().instrument()
response = openai.chat.completions.create(
    model="gpt-4o",
    messages=messages,
    tools=tools,
)

View traces with function calling in the Arize platform.

The Function Output tab displays the functions and their arguments generated by the LLM, while the Function Definition tab shows the tools available to the LLM, defined in the OpenAI tools parameter.

Pretty Format view of the tools passed in to the Open AI chat completion API.

Load Function Call Traces into Prompt Playground

After tracing your application, you can load the span into the prompt playground to experiment with improving LLM performance. In the playground, you can replay the function call with new LLM parameters, adjust function definitions, modify chat messages, and more!

Coming Soon

Stay tuned—we'll soon be adding function calling support for Gemini, Azure, Anthropic, and other models. We'll also add support for manual instrumentation of OpenAI function calls, complementing the existing auto-instrumentation feature.