Custom Targets

If you want to test an agent that is not natively supported, you can bring your own Target by defining a Python module containing a subclass of BaseTarget. The name of this module must contain the suffix _target (e.g. my_custom_target).

The subclass should implement the invoke method to invoke your agent and return a TargetResponse.

my_custom_target.py

from agenteval.targets import BaseTarget, TargetResponse
from my_agent import MyAgent

class MyCustomTarget(BaseTarget):

    def __init__(self, **kwargs):
        self.agent = MyAgent(**kwargs)

    def invoke(self, prompt: str) -> TargetResponse:

        response = self.agent.invoke(prompt)

        return TargetResponse(response=response)

Once you have created your subclass, specify the module path to the Target.

agenteval.yml

target:
  type: path.to.my_custom_target.MyCustomTarget`
  my_agent_parameter: "value" # (1)

1. This will be passed as kwargs when initializing the Target.

Warning

During a run, an instance of the Target will be created for each test in the test plan. We recommend avoiding testing Targets that load large models or vector stores into memory, as this can lead to a memory error. Consider deploying your agent and exposing it as a RESTful service.

Examples

REST API

We will implement a custom Target that invokes an agent exposed as a REST API.

my_api_target.py

import json

import requests

from agenteval.targets import BaseTarget, TargetResponse


class MyAPITarget(BaseTarget):
    def __init__(self, **kwargs):
        self.url = kwargs.get("url")

    def invoke(self, prompt: str) -> TargetResponse:
        data = {"message": prompt}

        response = requests.post(
            self.url, json=data, headers={"Content-Type": "application/json"}
        )

        return TargetResponse(response=json.loads(response.content)["agentResponse"])

Create a test plan that references MyAPITarget.

agenteval.yml

evaluator:
  model: claude-3
target:
  type: my_api_target.MyAPITarget
  url: https://api.example.com/invoke
tests:
  get_backlog_tickets:
    steps:
    - Ask agent how many tickets are left in the backlog
    expected_results:
    - Agent responds with 15 tickets

LangChain agent

We will create a simple LangChain agent which calculates the length of a given piece of text.

my_langchain_target.py

from langchain_community.llms import Bedrock
from langchain.agents import tool
from langchain import hub
from langchain.agents import AgentExecutor, create_xml_agent

from agenteval.targets import BaseTarget, TargetResponse

llm = Bedrock(model_id="anthropic.claude-v2:1")

@tool
def calculate_text_length(text: str) -> int:
    """Returns the length of a given text."""
    return len(text)


tools = [calculate_text_length]
prompt = hub.pull("hwchase17/xml-agent-convo")

agent = create_xml_agent(llm, tools, prompt)


class MyLangChainTarget(BaseTarget):
    def __init__(self, **kwargs):
        self.agent = AgentExecutor(agent=agent, tools=tools, verbose=False)

    def invoke(self, prompt: str) -> TargetResponse:

        response = self.agent.invoke({"input": prompt})["output"]

        return TargetResponse(response=response)

Create a test plan that references MyLangChainTarget.

agenteval.yml

evaluator:
  model: claude-3
target:
  type: my_langchain_target.MyLangChainTarget
tests:
  calculate_text_length:
    steps:
    - "Ask agent to calculate the length of this text: Hello world!"
    expected_results:
    - The agent responds that the length is 12.