Agent Evaluation Toolkit(智能体评估工具)
知识库位置:
FDE Knowledge Base/11 Tools/04-Engineering/03-Agent-Evaluation-Toolkit.md
Agent(智能体)不能只测 Final Answer(最终答案)。
必须测试完整轨迹(Trajectory,轨迹)。
Evaluation Dimensions(评估维度)
Section titled “Evaluation Dimensions(评估维度)”Goal Completion(目标完成度)
Section titled “Goal Completion(目标完成度)”任务最终完成了吗?
Tool Selection(工具选择)
Section titled “Tool Selection(工具选择)”选对 Tool(工具)吗?
Arguments(参数)
Section titled “Arguments(参数)”参数正确吗?
Policy Compliance(策略合规)
Section titled “Policy Compliance(策略合规)”有没有越权?
Efficiency(效率)
Section titled “Efficiency(效率)”用了多少 Step(步骤)?
Recovery(恢复)
Section titled “Recovery(恢复)”Tool 出错后能恢复吗?
Termination(终止)
Section titled “Termination(终止)”能正确停止吗?
Test Case(测试用例)
Section titled “Test Case(测试用例)”User Goal(用户目标):
Available Tools(可用工具):
Expected Tool Sequence(预期工具序列):
Forbidden Actions(禁止动作):
Expected Final State(预期最终状态):
Max Steps(最大步数):Agent Trace Evaluation(智能体轨迹评估)
Section titled “Agent Trace Evaluation(智能体轨迹评估)”User(用户)↓Decision(决策)↓Tool(工具)↓Arguments(参数)↓Result(结果)↓Next Decision(下一个决策)逐步评分。
Failure Types(失败类型)
Section titled “Failure Types(失败类型)”Wrong Tool(选错工具)Wrong Argument(参数错误)Unnecessary Tool(多余调用)Loop(循环)Premature Stop(过早停止)Permission Violation(越权)Hallucinated Success(幻觉式成功)Failure to Escalate(未升级)Critical Tests(必测场景)
Section titled “Critical Tests(必测场景)”必须加入:
- Missing input(输入缺失)
- Tool timeout(工具超时)
- Tool 403(工具拒绝访问)
- Tool 500(工具服务端错误)
- Duplicate result(重复结果)
- Dangerous request(危险请求)
- Ambiguous instruction(模糊指令)