Package dev.dokimos.core.evaluators.agents

Class ToolArgumentHallucinationEvaluator

java.lang.Object

dev.dokimos.core.BaseEvaluator

dev.dokimos.core.evaluators.agents.ToolArgumentHallucinationEvaluator

All Implemented Interfaces:

Evaluator

public class ToolArgumentHallucinationEvaluator
extends BaseEvaluator

Uses a judge LLM to assess whether tool call argument values are factually grounded in the user's input and preceding tool call results.

This is a glass-box evaluator for tool proficiency. For each tool call, the judge evaluates whether argument values can be derived from the user's request or from the results of earlier tool calls in the same execution. This supports multi-step agent workflows where later tool arguments are derived from earlier tool results (e.g., a search returns URLs, then a fetch tool uses one of those URLs).

When ToolCall.result() is populated, the result is included as grounding context for subsequent tool calls. When result is null, only the user input is considered as grounding context.

The score is the fraction of non-hallucinated tool calls (0.0 to 1.0).

Nested Class Summary

Nested Classes

Modifier and Type	Class	Description
static class	ToolArgumentHallucinationEvaluator.Builder	Builder for constructing the evaluator.

Method Summary

Modifier and Type	Method	Description
static ToolArgumentHallucinationEvaluator.Builder	builder()	Creates a new builder for constructing the evaluator.

Methods inherited from class dev.dokimos.core.BaseEvaluator

evaluate, evaluateAsync, evaluateAsync, name, threshold

Methods inherited from class java.lang.Object

equals, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Details

builder

public static ToolArgumentHallucinationEvaluator.Builder builder()

Creates a new builder for constructing the evaluator.

Returns:

a new builder