Tool Selection

Evaluate whether LLMs select the correct tools for given tasks

Overview

The Tool Selection evaluator determines whether an LLM selected the most appropriate tool (or tools) for a given task. This evaluator focuses on the what of tool calling - validating that the right tool was chosen - rather than whether the invocation arguments were correct.

When to Use

Use the Tool Selection evaluator when you need to:

• Validate tool choice decisions - Ensure the LLM picks the most appropriate tool for the task
• Detect hallucinated tools - Identify when the LLM tries to use tools that don’t exist
• Evaluate tool necessity - Check if the LLM correctly determines when tools are (or aren’t) needed
• Assess multi-tool selection - Validate when the LLM needs to select multiple tools for complex tasks

This evaluator validates tool selection correctness, not invocation correctness. For evaluating whether tool arguments are properly formatted, use the Tool Invocation evaluator instead.

Supported Levels

The level of an evaluator determines the scope of the evaluation in OpenTelemetry terms. Some evaluations are applicable to individual spans, some to full traces or sessions, and some are applicable at multiple levels.

Level	Supported	Notes
Span	Yes	Best for LLM spans that contain tool calls. Evaluate individual tool selection decisions.

Relevant span kinds: LLM spans with tool calls, particularly in agentic applications.

Input Requirements

The Tool Selection evaluator requires three inputs:

Field	Type	Description
input	string	The conversation context or user query
available_tools	string	List of available tools and their descriptions
tool_selection	string	The tool(s) selected by the LLM

In TypeScript, the fields use camelCase: availableTools and toolSelection.

Formatting Tips

While you can pass full JSON representations for each field, human-readable formats typically produce more accurate evaluations.

input (conversation context adapted from input messages):

User: I need to book a flight from New York to Los Angeles
Assistant: I'd be happy to help you book a flight. When would you like to travel?
User: Tomorrow morning, the earliest available

available_tools (tool descriptions adapted by JSON schemas):

book_flight: Book a flight between two cities. Requires origin, destination, and date.
search_hotels: Search for hotel accommodations by city and dates.
get_weather: Get current weather conditions for a location.
cancel_booking: Cancel an existing flight or hotel reservation.

Tool argument descriptions are optional; the focus is on the selection itself so tool names and descriptions are sufficient.

tool_selection (the LLM’s tool selection adapted from tool_calls in the output):

book_flight

If the LLM did not produce any tool calls, you can put “No tools called” as the tool_selection input.

Output Interpretation

The evaluator returns a Score object with the following properties:

Property	Value	Description
label	"correct" or "incorrect"	Classification result
score	1.0 or 0.0	Numeric score (1.0 = correct, 0.0 = incorrect)
explanation	string	LLM-generated reasoning for the classification
direction	"maximize"	Higher scores are better
metadata	object	Additional information such as the model name. When tracing is enabled, includes the trace_id for the evaluation.

Criteria for Correct (1.0):

• The LLM chose the best available tool for the user query
• The tool name exists in the available tools list
• The tool selection is safe and appropriate
• The correct number of tools were selected for the task

Criteria for Incorrect (0.0):

• The LLM used a hallucinated or nonexistent tool
• The LLM selected a tool when none was needed
• The LLM did not use a tool when one was required
• The LLM chose a suboptimal or irrelevant tool

Usage Examples

Python

from phoenix.evals import LLM
from phoenix.evals.metrics import ToolSelectionEvaluator

# Initialize the LLM client
llm = LLM(provider="openai", model="gpt-4o")

# Create the evaluator
tool_selection_eval = ToolSelectionEvaluator(llm=llm)

# Inspect the evaluator's requirements
print(tool_selection_eval.describe())

# Evaluate a tool selection using human-readable format
eval_input = {
    "input": """User: I need to book a flight from New York to Los Angeles
Assistant: I'd be happy to help you book a flight. When would you like to travel?
User: Tomorrow morning, the earliest available""",
    "available_tools": """book_flight: Book a flight between two cities. Requires origin, destination, and date.
search_hotels: Search for hotel accommodations by city and dates.
get_weather: Get current weather conditions for a location.
cancel_booking: Cancel an existing flight or hotel reservation.""",
    "tool_selection": "book_flight"
}

scores = tool_selection_eval.evaluate(eval_input)
print(scores[0])
# Score(name='tool_selection', score=1.0, label='correct', ...)

TypeScript

import { createToolSelectionEvaluator } from "@arizeai/phoenix-evals";
import { openai } from "@ai-sdk/openai";

// Create the evaluator
const toolSelectionEvaluator = createToolSelectionEvaluator({
  model: openai("gpt-4o"),
});

// Evaluate a tool selection using human-readable format
const result = await toolSelectionEvaluator.evaluate({
  input: `User: I need to book a flight from New York to Los Angeles
Assistant: I'd be happy to help you book a flight. When would you like to travel?
User: Tomorrow morning, the earliest available`,
  availableTools: `book_flight: Book a flight between two cities. Requires origin, destination, and date.
search_hotels: Search for hotel accommodations by city and dates.
get_weather: Get current weather conditions for a location.
cancel_booking: Cancel an existing flight or hotel reservation.`,
  toolSelection: "book_flight",
});

console.log(result);
// { score: 1, label: "correct", explanation: "..." }

Using Input Mapping

When your data has different field names, use input mapping.

Python

from phoenix.evals import LLM
from phoenix.evals.metrics import ToolSelectionEvaluator

llm = LLM(provider="openai", model="gpt-4o")
tool_selection_eval = ToolSelectionEvaluator(llm=llm)

eval_input = {
    "conversation": """User: I want to search for flights to Paris
Assistant: Sure, I can help with that. When are you planning to travel?
User: Next weekend""",
    "tools_available": """flight_search: Search for available flights by destination and date.
hotel_search: Search for hotel accommodations.
car_rental: Search for rental car options.""",
    "selected_tool": "flight_search"
}

input_mapping = {
    "input": "conversation",
    "available_tools": "tools_available",
    "tool_selection": "selected_tool"
}

scores = tool_selection_eval.evaluate(eval_input, input_mapping)

For more details on input mapping options, see Input Mapping.

TypeScript

import { bindEvaluator, createToolSelectionEvaluator } from "@arizeai/phoenix-evals";
import { openai } from "@ai-sdk/openai";

const toolSelectionEvaluator = createToolSelectionEvaluator({
  model: openai("gpt-4o"),
});

const boundEvaluator = bindEvaluator(toolSelectionEvaluator, {
  inputMapping: {
    input: "conversation",
    availableTools: "toolsAvailable",
    toolSelection: "selectedTool",
  },
});

const result = await boundEvaluator.evaluate({
  conversation: `User: I want to search for flights to Paris
Assistant: Sure, I can help with that. When are you planning to travel?
User: Next weekend`,
  toolsAvailable: `flight_search: Search for available flights by destination and date.
hotel_search: Search for hotel accommodations.
car_rental: Search for rental car options.`,
  selectedTool: "flight_search",
});

For more details on input mapping options, see Input Mapping.

Configuration

For LLM client configuration options, see Configuring the LLM.

Viewing and Modifying the Prompt

You can view the latest versions of our prompt templates on GitHub. The evaluators are designed to work well in a variety of contexts, but we highly recommend modifying the prompt to be more specific to your use case. Feel free to adapt them.

Python

from phoenix.evals.metrics import ToolSelectionEvaluator
from phoenix.evals import LLM, ClassificationEvaluator

llm = LLM(provider="openai", model="gpt-4o")
evaluator = ToolSelectionEvaluator(llm=llm)

# View the prompt template
print(evaluator.prompt_template)

# Create a custom evaluator based on the built-in template
custom_evaluator = ClassificationEvaluator(
    name="tool_selection",
    prompt_template=evaluator.prompt_template,  # Modify as needed
    llm=llm,
    choices={"correct": 1.0, "incorrect": 0.0},
    direction="maximize",
)

TypeScript

import { TOOL_SELECTION_CLASSIFICATION_EVALUATOR_CONFIG, createToolSelectionEvaluator } from "@arizeai/phoenix-evals";
import { openai } from "@ai-sdk/openai";

// View the prompt template
console.log(TOOL_SELECTION_CLASSIFICATION_EVALUATOR_CONFIG.template);

// Create a custom evaluator with a modified template
const customEvaluator = createToolSelectionEvaluator({
  model: openai("gpt-4o"),
  promptTemplate: TOOL_SELECTION_CLASSIFICATION_EVALUATOR_CONFIG.template, // Modify as needed
});

Using with Phoenix

Evaluating Traces

Run evaluations on traces collected in Phoenix and log results as annotations:

• Evaluating Phoenix Traces
• Logging LLM Evaluations

Running Experiments

Use the Tool Selection evaluator in Phoenix experiments:

• Using Evaluators in Experiments

API Reference

• Python: ToolSelectionEvaluator
• TypeScript: createToolSelectionEvaluator

Related

• Tool Invocation Evaluator - For evaluating tool invocation correctness
• Correctness Evaluator - For evaluating factual accuracy