AI agent evaluation

You can set this evaluation up using our API configuration.

AI agents are LLM apps designed to autonomously perform tasks, make decisions, and interact with users. AI agents are based on a list of LLM prompts where each LLM call is executed, and the result is fed into the next prompt. Building robust AI agents isn't without its challenges—unreliable reasoning, inconsistent data flow, and frequent hallucinations can all pose significant hurdles. To ensure these systems are reliable, a robust evaluation strategy is essential.

Problems

• Unreliable Reasoning: LLMs often struggle to make accurate and consistent decisions, leading to incorrect interpretations of user intents. For example, an AI customer support agent might misunderstand a user's request for a refund, offering a discount instead.
• Inconsistent Data Flow: Managing data flow in complex applications can be quite challenging. Take, for example, a medical diagnosis AI that, for the same set of symptoms, sometimes provides a recommendation and sometimes doesn't. This inconsistency can undermine trust and reliability.
• Frequent Hallucinations: LLMs sometimes generate false or irrelevant information, known as hallucinations. For example, an AI writing assistant might invent fictitious historical events or facts, misleading users.

Solutions

• Unit Testing: Break down AI agents into individual components to evaluate. Ensure each piece of the AI agent is tested individually to reduce hallucinations and false outputs.
• End-to-End Testing: Assess the entire system's performance with A/B testing and API evaluations.
• Structured Data Flow Testing: Implement structured evaluations to ensure LLMs produce consistent and accurate outputs. See JSON evals for an example.

Example

At its core, evaluating AI agents boils down to two main approaches: unit testing and end-to-end testing.

Unit Testing AI Agents

Unit testing involves breaking down the AI agent into its smallest functional components and testing each one individually.

To test AI agents one unit at a time, you can break the AI agent into individual components, and use a basic configuration to test each unit.

{
  "name": "AI Agent Single Unit Experiment",
  "description": "This experiment tests prompt response on a single step of an AI agent that infers a list of criteria from profile text.",
  "prompts": [
    [
      {
        "role": "system",
        "content": "You are a helpful assistant"
      },
      {
        "role": "user",
        "content": "Infer a list of criteria for this person like works at a startup, is a founder, and HQ location, etc: {{body}}"
      }
    ]
  ],
  "providers": [
    {
      "id": "openai:gpt-3.5-turbo-0125",
      "config": {
        "max_tokens": 1024,
        "temperature": 0.1
      }
    },
    {
      "id": "openai:gpt-4o-mini",
      "config": {
        "organization": "",
        "temperature": 0.1,
        "max_tokens": 1024,
        "top_p": 1,
        "frequency_penalty": 0,
        "presence_penalty": 0
      }
    }
  ],
  "tests": [
    {
      "vars": {
        "body": "Andrew Ng. Founder of DeepLearning.AI; Managing General Partner of AI Fund; Exec Chairman of Landing AI. HQ in New York Area."
      },
      "assert": [
        {
          "type": "factuality",
          "value": "Is a founder"
        },
        {
          "type": "model-graded-closedqa",
          "value": "Mentions location of HQ"
        }
      ]
    },
    {
      "vars": {
        "body": "Andrew Perkins. Chef at Canva. HQ in United States."
      },
      "assert": [
        {
          "type": "factuality",
          "value": "Is not a founder"
        },
        {
          "type": "model-graded-closedqa",
          "value": "Mentions location of HQ"
        }
      ]
    }
  ]
}

See API configurations for additional evaluation examples.

End-to-End Testing for AI Agents

End-to-end testing assesses the AI agent's overall performance from start to finish. This approach is crucial for understanding how well the entire system functions when all components are integrated.

To test the entire AI agent end to end, provide an API that takes an input and outputs the result of your AI agent. Below is an example configuration to test your AI agent by sending questions to an API endpoint.

{  
  "name": "AI Agent Test",
  "prompts": [  
    "{{question}}"  
  ],  
  "providers": [  
    {  
      "id": "https", 
      "config": {  
        "url": "https://example-ai-agent.com/process",  
        "method": "POST",
        "headers": {  
          "Content-Type": "application/json"  
        },
        "body": {  
          "question": "{{question}}"
        }
      }  
    }  
  ], 
  "tests": [
    {  
      "vars": {  
        "question": "Who are the top AI startup founders in United States with a chef?"
      },  
      "assert": [
        {  
          "type": "contains-all",  
          "value": [  
            "Sam Smith",  
            "John Luke"
          ]  
        }  
      ]  
    }  
  ]  
}  

For more on evaluation metric configurations, see the references guide.

Summary

To make AI agents truly reliable, we must address issues like unreliable reasoning, inconsistent data flow, and hallucinations. A robust evaluation strategy is essential. By implementing these tests, we can significantly improve the performance of AI agents, and make them more intelligent.