eval-tool-use (Function and Tool Calling)

This example demonstrates how to evaluate LLM function/tool calling capabilities using promptfoo.

You can run this example with:

bash

npx promptfoo@latest init --example eval-tool-use
cd eval-tool-use

Overview

This example shows how to configure and test function/tool calling capabilities across multiple LLM providers:

OpenAI (with native function calling)
Anthropic (with Claude's tool use)
AWS Bedrock models
Groq (with function calling)

Each provider has slightly different syntax and requirements for implementing function/tool calling.

Environment Variables

This example requires the following environment variables:

OPENAI_API_KEY - Your OpenAI API key
ANTHROPIC_API_KEY - Your Anthropic API key
AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY - For AWS Bedrock (if using the Bedrock example)
GROQ_API_KEY - If using Groq's LLaMA models

You can set these in a .env file or directly in your environment.

Provider Documentation

Each provider implements tool use with different syntax:

Running the Example

The configuration for this example is in:

promptfooconfig.yaml - Main example with OpenAI, Anthropic, and Groq
promptfooconfig.bedrock.yaml - Example specifically for AWS Bedrock models

To run the main example:

bash

promptfoo eval

To run the Bedrock example:

bash

promptfoo eval -c promptfooconfig.bedrock.yaml

After running the evaluation, view the results with:

bash

promptfoo view

Example Tool: Weather Function

This example uses a simple weather lookup function that takes a location and optionally a temperature unit. The example illustrates how different providers handle the same function definition with different syntaxes.

External tools can also be loaded from separate files, as demonstrated with external_tools.yaml.

Anthropic Strict Mode

The Anthropic provider includes an example with strict: true enabled, which uses Anthropic's structured outputs feature to guarantee that tool parameters exactly match your schema. This is useful for:

Building reliable agentic workflows
Ensuring type-safe function calls
Production systems that require guaranteed schema conformance

When strict: true is enabled, Claude will always return tool inputs that strictly follow your input_schema, with no type mismatches or missing required fields. See the Anthropic structured outputs example for more details.

Finish Reason Assertions

This example also demonstrates the use of finish-reason assertions to validate why a model stopped generating:

tool_calls: Verifies the model stopped to make a function/tool call (e.g., weather lookup for cities)

The example shows that when models are asked about weather in real cities (Boston, New York, Paris), they correctly stop generation to make tool calls, resulting in a tool_calls finish reason. This helps ensure your models are using tools appropriately when they should be.