Agentic applications give an LLM freedom over control flow in order to solve problems. While this freedom can be extremely powerful, the black box nature of LLMs can make it difficult to understand how changes in one part of your agent will affect others downstream. This makes evaluating your agents especially important.
This package contains a collection of evaluators and utilities for evaluating the performance of your agents, with a focus on agent trajectory, or the intermediate steps an agent takes as it runs. It is intended to provide a good conceptual starting point for your agent's evals.
If you are looking for more general evaluation tools, please check out the companion package openevals.
To get started, install agentevals:
Python
pip install agentevalsTypeScript
npm install agentevals @langchain/coreThis quickstart will use an evaluator powered by OpenAI's o3-mini model to judge your results, so you'll need to set your OpenAI API key as an environment variable:
export OPENAI_API_KEY="your_openai_api_key"Once you've done this, you can run your first trajectory evaluator. We represent the agent's trajectory as a list of OpenAI-style messages:
Python
from agentevals.trajectory.llm import create_trajectory_llm_as_judge, TRAJECTORY_ACCURACY_PROMPT_WITH_REFERENCE
trajectory_evaluator = create_trajectory_llm_as_judge(
prompt=TRAJECTORY_ACCURACY_PROMPT,
model="openai:o3-mini",
)
# This is a fake trajectory, in reality you would run your agent to get a real trajectory
outputs = [
{"role": "user", "content": "What is the weather in SF?"},
{
"role": "assistant",
"tool_calls": [
{
"function": {
"name": "get_weather",
"arguments": json.dumps({"city": "SF"}),
}
}
],
},
{"role": "tool", "content": "It's 80 degrees and sunny in SF."},
{"role": "assistant", "content": "The weather in SF is 80 degrees and sunny."},
]
eval_result = trajectory_evaluator(
outputs=outputs,
)
print(eval_result){
'key': 'trajectory_accuracy',
'reasoning': 'The trajectory accurately follows the user's request for weather information in SF. Initially, the assistant recognizes the goal (providing weather details), then it efficiently makes a tool call to get the weather, and finally it communicates the result clearly. All steps demonstrate logical progression and efficiency. Thus, the score should be: true.',
'score': true
}
TypeScript
import {
createTrajectoryLLMAsJudge,
TRAJECTORY_ACCURACY_PROMPT,
} from "agentevals";
const trajectoryEvaluator = createTrajectoryLLMAsJudge({
prompt: TRAJECTORY_ACCURACY_PROMPT,
model: "openai:o3-mini",
});
const outputs = [
{ role: "user", content: "What is the weather in SF?" },
{
role: "assistant",
content: "",
tool_calls: [
{
function: {
name: "get_weather",
arguments: JSON.stringify({ city: "SF" }),
},
},
],
},
{ role: "tool", content: "It's 80 degrees and sunny in SF." },
{
role: "assistant",
content: "The weather in SF is 80 degrees and sunny.",
},
];
const evalResult = await trajectoryEvaluator({
outputs,
});
console.log(evalResult);{
key: 'trajectory_accuracy',
score: true,
comment: '...'
}
You can see that despite the small difference in the final response and tool calls, the evaluator still returns a score of true since the overall trajectory is the same between the output and reference!
You can install agentevals like this:
Python
pip install agentevalsTypeScript
npm install agentevals @langchain/coreFor LLM-as-judge evaluators, you will also need an LLM client. By default, agentevals will use LangChain chat model integrations and comes with langchain_openai installed by default. However, if you prefer, you may use the OpenAI client directly:
Python
pip install openaiTypeScript
npm install openaiIt is also helpful to be familiar with some evaluation concepts and LangSmith's pytest integration for running evals, which is documented here.
Agent trajectory evaluators are used to judge the trajectory of an agent's execution either against an expected trajectory or using an LLM.
These evaluators expect you to format your agent's trajectory as a list of OpenAI format dicts or as a list of LangChain BaseMessage classes, and handle message formatting
under the hood.
The trajectory_strict_match evaluator, compares two trajectories and ensures that they contain the same messages
in the same order with the same tool calls. It allows for differences in message content and tool call arguments,
but requires that the selected tools at each step are the same.
Python
import json
from agentevals.trajectory.strict import trajectory_strict_match
outputs = [
{"role": "user", "content": "What is the weather in SF?"},
{
"role": "assistant",
"tool_calls": [
{
"function": {
"name": "get_weather",
"arguments": json.dumps({"city": "SF"}),
}
}
],
},
{"role": "tool", "content": "It's 80 degrees and sunny in SF."},
{"role": "assistant", "content": "The weather in SF is 80 degrees and sunny."},
]
reference_outputs = [
{"role": "user", "content": "What is the weather in San Francisco?"},
{
"role": "assistant",
"tool_calls": [
{
"function": {
"name": "get_weather",
"arguments": json.dumps({"city": "San Francisco"}),
}
}
],
},
{"role": "tool", "content": "It's 80 degrees and sunny in San Francisco."},
{"role": "assistant", "content": "The weather in SF is 80˚ and sunny."},
]
result = trajectory_strict_match(
outputs=outputs, reference_outputs=reference_outputs
)
print(result){
'key': 'trajectory_accuracy',
'score': True,
'comment': None,
}
TypeScript
import { trajectoryStrictMatch } from "agentevals";
const outputs = [
{ role: "user", content: "What is the weather in SF?" },
{
role: "assistant",
tool_calls: [{
function: { name: "get_weather", arguments: JSON.stringify({ city: "SF" }) }
}]
},
{ role: "tool", content: "It's 80 degrees and sunny in SF." },
{ role: "assistant", content: "The weather in SF is 80 degrees and sunny." },
];
const referenceOutputs = [
{ role: "user", content: "What is the weather in San Francisco?" },
{ role: "assistant", tool_calls: [{ function: { name: "get_weather", arguments: JSON.stringify({ city: "San Francisco" }) } }] },
{ role: "tool", content: "It's 80 degrees and sunny in San Francisco." },
];
const result = await trajectoryStrictMatch({
outputs,
referenceOutputs,
});
console.log(result);{
'key': 'trajectory_accuracy',
'score': true,
}
The trajectory_unordered_match evaluator, compares two trajectories and ensures that they contain the same number of tool calls in any order. This is useful if you want to allow flexibility in how an agent obtains the proper information, but still do care that all information was retrieved.
Python
import json
from agentevals.trajectory.unordered import trajectory_unordered_match
inputs = {}
outputs = [
{"role": "user", "content": "What is the weather in SF and is there anything fun happening?"},
{
"role": "assistant",
"tool_calls": [{
"function": {
"name": "get_weather",
"arguments": json.dumps({"city": "SF"}),
}
}],
},
{"role": "tool", "content": "It's 80 degrees and sunny in SF."},
{
"role": "assistant",
"tool_calls": [{
"function": {
"name": "get_fun_activities",
"arguments": json.dumps({"city": "SF"}),
}
}],
},
{"role": "tool", "content": "Nothing fun is happening, you should stay indoors and read!"},
{"role": "assistant", "content": "The weather in SF is 80 degrees and sunny, but there is nothing fun happening."},
]
reference_outputs = [
{"role": "user", "content": "What is the weather in SF and is there anything fun happening?"},
{
"role": "assistant",
"tool_calls": [
{
"function": {
"name": "get_fun_activities",
"arguments": json.dumps({"city": "San Francisco"}),
}
},
{
"function": {
"name": "get_weather",
"arguments": json.dumps({"city": "San Francisco"}),
}
},
],
},
{"role": "tool", "content": "Nothing fun is happening, you should stay indoors and read!"},
{"role": "tool", "content": "It's 80 degrees and sunny in SF."},
{"role": "assistant", "content": "In SF, it's 80˚ and sunny, but there is nothing fun happening."},
]
result = trajectory_unordered_match(
outputs=outputs, reference_outputs=reference_outputs
)
print(result){
'key': 'trajectory_unordered_match',
'score': True,
'comment': None,
}
TypeScript
import { trajectoryUnorderedMatch } from "agentevals";
const outputs = [
{ role: "user", content: "What is the weather in SF and is there anything fun happening?" },
{
role: "assistant",
tool_calls: [{
function: {
name: "get_weather",
arguments: JSON.stringify({ city: "SF" }),
}
}],
},
{ role: "tool", content: "It's 80 degrees and sunny in SF." },
{
role: "assistant",
tool_calls: [{
function: {
name: "get_fun_activities",
arguments: JSON.stringify({ city: "SF" }),
}
}],
},
{ role: "tool", content: "Nothing fun is happening, you should stay indoors and read!" },
{ role: "assistant", content: "The weather in SF is 80 degrees and sunny, but there is nothing fun happening." },
];
const referenceOutputs = [
{ role: "user", content: "What is the weather in SF and is there anything fun happening?" },
{
role: "assistant",
tool_calls: [
{
function: {
name: "get_fun_activities",
arguments: JSON.stringify({ city: "San Francisco" }),
}
},
{
function: {
name: "get_weather",
arguments: JSON.stringify({ city: "San Francisco" }),
}
},
],
},
{ role: "tool", content: "Nothing fun is happening, you should stay indoors and read!" },
{ role: "tool", content: "It's 80 degrees and sunny in SF." },
{ role: "assistant", content: "In SF, it's 80˚ and sunny, but there is nothing fun happening." },
];
const result = await trajectoryUnorderedMatch({
outputs,
referenceOutputs,
});
console.log(result){
'key': 'trajectory_unordered_match',
'score': true,
}
There are other evaluators for checking partial trajectory matches (ensuring that a trajectory contains a subset and superset of tool calls compared to a reference trajectory).
Python
import json
from openevals.trajectory.subset import trajectory_subset
# from openevals.trajectory.superset import trajectory_superset
outputs = [
{"role": "user", "content": "What is the weather in SF and London?"},
{
"role": "assistant",
"tool_calls": [{
"function": {
"name": "get_weather",
"arguments": json.dumps({"city": "SF and London"}),
}
}],
},
{"role": "tool", "content": "It's 80 degrees and sunny in SF, and 90 degrees and rainy in London."},
{"role": "assistant", "content": "The weather in SF is 80 degrees and sunny. In London, it's 90 degrees and rainy."},
]
reference_outputs = [
{"role": "user", "content": "What is the weather in SF and London?"},
{
"role": "assistant",
"tool_calls": [
{
"function": {
"name": "get_weather",
"arguments": json.dumps({"city": "San Francisco"}),
}
},
{
"function": {
"name": "get_weather",
"arguments": json.dumps({"city": "London"}),
}
},
],
},
{"role": "tool", "content": "It's 80 degrees and sunny in San Francisco."},
{"role": "tool", "content": "It's 90 degrees and rainy in London."},
{"role": "assistant", "content": "The weather in SF is 80˚ and sunny. In London, it's 90˚ and rainy."},
]
result = trajectory_subset(
outputs=outputs, reference_outputs=reference_outputs
)
print(result){
'key': 'trajectory_subset',
'score': True,
'comment': None,
}
TypeScript
import { trajectorySubset } from "agentevals";
// import { trajectorySuperset } from "agentevals";
const outputs = [
{ role: "user", content: "What is the weather in SF and London?" },
{
role: "assistant",
tool_calls: [{
function: {
name: "get_weather",
arguments: JSON.stringify({ city: "SF and London" }),
}
}],
},
{ role: "tool", content: "It's 80 degrees and sunny in SF, and 90 degrees and rainy in London." },
{ role: "assistant", content: "The weather in SF is 80 degrees and sunny. In London, it's 90 degrees and rainy."},
];
const referenceOutputs = [
{ role: "user", content: "What is the weather in SF and London?" },
{
role: "assistant",
tool_calls: [
{
function: {
name: "get_weather",
arguments: JSON.stringify({ city: "San Francisco" }),
}
},
{
function: {
name: "get_weather",
arguments: JSON.stringify({ city: "London" }),
}
},
],
},
{ role: "tool", content: "It's 80 degrees and sunny in San Francisco." },
{ role: "tool", content: "It's 90 degrees and rainy in London." },
{ role: "assistant", content: "The weather in SF is 80˚ and sunny. In London, it's 90˚ and rainy." },
];
const result = await trajectorySubset({
outputs,
referenceOutputs,
});
console.log(result){
'key': 'trajectory_subset',
'score': true,
}
The LLM-as-judge trajectory evaluator that uses an LLM to evaluate the trajectory. Unlike the other trajectory evaluators, it doesn't require a reference trajectory, and supports This allows for more flexibility in the trajectory comparison:
Python
import json
from openevals.trajectory.llm import create_trajectory_llm_as_judge, TRAJECTORY_ACCURACY_PROMPT_WITH_REFERENCE
evaluator = create_trajectory_llm_as_judge(
prompt=TRAJECTORY_ACCURACY_PROMPT_WITH_REFERENCE,
model="openai:o3-mini"
)
outputs = [
{"role": "user", "content": "What is the weather in SF?"},
{
"role": "assistant",
"tool_calls": [
{
"function": {
"name": "get_weather",
"arguments": json.dumps({"city": "SF"}),
}
}
],
},
{"role": "tool", "content": "It's 80 degrees and sunny in SF."},
{"role": "assistant", "content": "The weather in SF is 80 degrees and sunny."},
]
reference_outputs = [
{"role": "user", "content": "What is the weather in SF?"},
{
"role": "assistant",
"tool_calls": [
{
"function": {
"name": "get_weather",
"arguments": json.dumps({"city": "San Francisco"}),
}
}
],
},
{"role": "tool", "content": "It's 80 degrees and sunny in San Francisco."},
{"role": "assistant", "content": "The weather in SF is 80˚ and sunny."},
]
eval_result = evaluator(
outputs=outputs,
reference_outputs=reference_outputs,
)
print(eval_result){
'key': 'trajectory_accuracy',
'score': True,
'comment': 'The provided agent trajectory is consistent with the reference. Both trajectories start with the same user query and then correctly invoke a weather lookup through a tool call. Although the reference uses "San Francisco" while the provided trajectory uses "SF" and there is a minor formatting difference (degrees vs. ˚), these differences do not affect the correctness or essential steps of the process. Thus, the score should be: true.'
}
TypeScript
import {
createTrajectoryLLMAsJudge,
TRAJECTORY_ACCURACY_PROMPT_WITH_REFERENCE
} from "agentevals";
const evaluator = createTrajectoryLLMAsJudge({
prompt: TRAJECTORY_ACCURACY_PROMPT_WITH_REFERENCE,
model: "openai:o3-mini",
});
const outputs = [
{role: "user", content: "What is the weather in SF?"},
{
role: "assistant",
tool_calls: [
{
function: {
name: "get_weather",
arguments: JSON.stringify({ city: "SF" }),
}
}
],
},
{role: "tool", content: "It's 80 degrees and sunny in SF."},
{role: "assistant", content: "The weather in SF is 80 degrees and sunny."},
]
const referenceOutputs = [
{role: "user", content: "What is the weather in SF?"},
{
role: "assistant",
tool_calls: [
{
function: {
name: "get_weather",
arguments: JSON.stringify({ city: "San Francisco" }),
}
}
],
},
{role: "tool", content: "It's 80 degrees and sunny in San Francisco."},
{role: "assistant", content: "The weather in SF is 80˚ and sunny."},
]
const result = await evaluator({
outputs,
referenceOutputs,
});
console.log(result){
'key': 'trajectory_accuracy',
'score': true,
'comment': 'The provided agent trajectory is consistent with the reference. Both trajectories start with the same user query and then correctly invoke a weather lookup through a tool call. Although the reference uses "San Francisco" while the provided trajectory uses "SF" and there is a minor formatting difference (degrees vs. ˚), these differences do not affect the correctness or essential steps of the process. Thus, the score should be: true.'
}
create_trajectory_llm_as_judge takes the same parameters as create_llm_as_judge in openevals, so you can customize the prompt and scoring output as needed.
In addition to prompt and model, the following parameters are also available:
continuous: a boolean that sets whether the evaluator should return a float score somewhere between 0 and 1 instead of a binary score. Defaults toFalse.choices: a list of floats that sets the possible scores for the evaluator.system: a string that sets a system prompt for the judge model by adding a system message before other parts of the prompt.few_shot_examples: a list of example dicts that are appended to the end of the prompt. This is useful for providing the judge model with examples of good and bad outputs. The required structure looks like this:
Python
few_shot_examples = [
{
"inputs": "What color is the sky?",
"outputs": "The sky is red.",
"reasoning": "The sky is red because it is early evening.",
"score": 1,
}
]TypeScript
const fewShotExamples = [
{
inputs: "What color is the sky?",
outputs: "The sky is red.",
reasoning: "The sky is red because it is early evening.",
score: 1,
}
];See the openevals repo for a fully up to date list of parameters.
For frameworks like LangGraph that model agents as graphs, it can be more convenient to represent trajectories in terms of nodes visited rather than messages. agentevals includes a category of evaluators called graph trajectory evaluators that are designed to work with this format, as well as convenient utilities for extracting trajectories from a LangGraph thread, including different conversation turns and interrupts.
The below examples will use LangGraph with the built-in formatting utility, but graph evaluators accept input in the following general format:
Python
class GraphTrajectory(TypedDict):
# Only set when specifying reference_outputs
inputs: Optional[list[dict]]
results: list[dict]
steps: list[list[str]]
def evaluator(
*,
inputs: Optional[Union[dict, list]] = None,
outputs: GraphTrajectory,
reference_outputs: Optional[GraphTrajectory] = None,
) -> ...TypeScript
export type GraphTrajectory = {
inputs?: (Record<string, unknown> | null)[];
results: Record<string, unknown>[];
steps: string[][];
};
const evaluator: ({ inputs, outputs, referenceOutputs, ...extra }: {
inputs: (string | Record<string, unknown> | null)[] | {
inputs: (string | Record<string, unknown> | null)[];
};
outputs: GraphTrajectory;
referenceOutputs?: GraphTrajectory;
[key: string]: unknown;
}) => ...Where inputs is a list of inputs (or a dict with a key named "inputs") to the graph whose items each represent the start of a new invocation in a thread, results representing the final output from each turn in the thread, and steps representing the internal steps taken for each turn.
This evaluator is similar to the trajectory_llm_as_judge evaluator, but it works with graph trajectories instead of message trajectories. Below, we set up a LangGraph agent, extract a trajectory from it using the built-in utils, and pass it to the evaluator. First, let's setup our graph, call it, and then extract the trajectory:
Python
from agentevals.graph_trajectory.utils import (
extract_langgraph_trajectory_from_thread,
)
from agentevals.graph_trajectory.llm import create_graph_trajectory_llm_as_judge
from langgraph.prebuilt import create_react_agent
from langgraph.checkpoint.memory import MemorySaver
from langgraph.types import Command, interrupt
from langchain_core.tools import tool
@tool
def search(query: str):
"""Call to surf the web."""
user_answer = interrupt("Tell me the answer to the question.")
return user_answer
tools = [search]
checkpointer = MemorySaver()
graph = create_react_agent(
model="gpt-4o-mini",
checkpointer=checkpointer,
tools=[search],
)
graph.invoke(
{"messages": [{"role": "user", "content": "what's the weather in sf?"}]},
config={"configurable": {"thread_id": "1"}},
)
# Resume the agent with a new command, simulating a human-in-the-loop workflow
graph.invoke(
Command(resume="It is rainy and 70 degrees!"),
config={"configurable": {"thread_id": "1"}},
)
# Extract the trajectory from the first two thread runs
extracted_trajectory = extract_langgraph_trajectory_from_thread(
graph, {"configurable": {"thread_id": "1"}}
)
print(extracted_trajectory){
'inputs': [{
'__start__': {
'messages': [
{'role': 'user', 'content': "what's the weather in sf?"}
]}
},
'__resuming__': {
'messages': [
{'role': 'user', 'content': 'It is rainy and 70 degrees!'}
]}
],
'outputs': {
'results': [
{},
{
'messages': [
{'role': 'ai', 'content': 'The current weather in San Francisco is rainy, with a temperature of 70 degrees.'}
]
}
],
'steps': [
['__start__', 'agent', 'tools', '__interrupt__'],
['agent']
]
}
}
TypeScript
import { tool } from "@langchain/core/tools";
import { ChatOpenAI } from "@langchain/openai";
import { createReactAgent } from "@langchain/langgraph/prebuilt";
import { MemorySaver, interrupt } from "@langchain/langgraph";
import { z } from "zod";
import { extractLangGraphTrajectoryFromThread } from "agentevals";
const search = tool((_): string => {
const userAnswer = interrupt("Tell me the answer to the question.")
return userAnswer;
}, {
name: "search",
description: "Call to surf the web.",
schema: z.object({
query: z.string()
})
})
const tools = [search];
// Create a checkpointer
const checkpointer = new MemorySaver();
// Create the React agent
const graph = createReactAgent({
llm: new ChatOpenAI({ model: "gpt-4o-mini" }),
tools,
checkpointer,
});
// Invoke the graph with initial message
await graph.invoke(
{ messages: [{ role: "user", content: "what's the weather in sf?" }] },
{ configurable: { thread_id: "1" } }
);
// Resume the agent with a new command (simulating human-in-the-loop)
await graph.invoke(
{ messages: [{ role: "user", content: "It is rainy and 70 degrees!" }] },
{ configurable: { thread_id: "1" } }
);
const extractedTrajectory = await extractLangGraphTrajectoryFromThread(
graph,
{ configurable: { thread_id: "1" } },
);
console.log(extractedTrajectory);{
'inputs': [{
'__start__': {
'messages': [
{'role': 'user', 'content': "what's the weather in sf?"}
]}
},
'__resuming__': {
'messages': [
{'role': 'user', 'content': 'It is rainy and 70 degrees!'}
]}
],
'outputs': {
'results': [
{},
{
'messages': [
{'role': 'ai', 'content': 'The current weather in San Francisco is rainy, with a temperature of 70 degrees.'}
]
}
],
'steps': [
['__start__', 'agent', 'tools', '__interrupt__'],
['agent']
]
}
}
Now, we can pass the extracted trajectory to the evaluator:
Python
graph_trajectory_evaluator = create_graph_trajectory_llm_as_judge(
model="openai:o3-mini",
)
res = graph_trajectory_evaluator(
inputs=extracted_trajectory["inputs"],
outputs=extracted_trajectory["outputs"],
)
print(res){
'key': 'graph_trajectory_accuracy',
'score': True,
'comment': 'The overall process follows a logical progression: the conversation begins with the user’s request, the agent then processes the request through its own internal steps (including calling tools), interrupts to obtain further input, and finally resumes to provide a natural language answer. Each step is consistent with the intended design in the rubric, and the overall path is relatively efficient and semantically aligns with a typical query resolution trajectory. Thus, the score should be: true.'
}
TypeScript
import { createGraphTrajectoryLLMAsJudge } from "agentevals";
const graphTrajectoryEvaluator = createGraphTrajectoryLLMAsJudge({
model: "openai:o3-mini",
})
const res = await graphTrajectoryEvaluator(
inputs=extractedTrajectory.inputs,
outputs=extractedTrajectory.outputs,
)
console.log(res);{
'key': 'graph_trajectory_accuracy',
'score': True,
'comment': 'The overall process follows a logical progression: the conversation begins with the user’s request, the agent then processes the request through its own internal steps (including calling tools), interrupts to obtain further input, and finally resumes to provide a natural language answer. Each step is consistent with the intended design in the rubric, and the overall path is relatively efficient and semantically aligns with a typical query resolution trajectory. Thus, the score should be: true.'
}
Note that though this evaluator takes the typical inputs, outputs, and reference_outputs parameters, it internally combines inputs and outputs to form a thread. Therefore, if you want to customize the prompt, your prompt should also contain a thread input variable:
Python
CUSTOM_PROMPT = """You are an expert data labeler.
Your task is to grade the accuracy of an AI agent's internal steps in resolving a user queries.
<Rubric>
An accurate trajectory:
- Makes logical sense between steps
- Shows clear progression
- Is perfectly efficient, with no more than one tool call
- Is semantically equivalent to the provided reference trajectory, if present
</Rubric>
<Instructions>
Grade the following thread, evaluating whether the agent's overall steps are logical and relatively efficient.
For the trajectory, "__start__" denotes an initial entrypoint to the agent, and "__interrupt__" corresponds to the agent
interrupting to await additional data from another source ("human-in-the-loop"):
</Instructions>
<thread>
{thread}
</thread>
{reference_outputs}
"""
evaluator = create_graph_trajectory_llm_as_judge(
prompt=CUSTOM_PROMPT,
model="openai:o3-mini",
)
res = await evaluator(
inputs=extracted_trajectory["inputs"],
outputs=extracted_trajectory["outputs"],
)TypeScript
const CUSTOM_PROMPT = `You are an expert data labeler.
Your task is to grade the accuracy of an AI agent's internal steps in resolving a user queries.
<Rubric>
An accurate trajectory:
- Makes logical sense between steps
- Shows clear progression
- Is perfectly efficient, with no more than one tool call
- Is semantically equivalent to the provided reference trajectory, if present
</Rubric>
<Instructions>
Grade the following thread, evaluating whether the agent's overall steps are logical and relatively efficient.
For the trajectory, "__start__" denotes an initial entrypoint to the agent, and "__interrupt__" corresponds to the agent
interrupting to await additional data from another source ("human-in-the-loop"):
</Instructions>
<thread>
{thread}
</thread>
{reference_outputs}
`
const graphTrajectoryEvaluator = createGraphTrajectoryLLMAsJudge({
prompt: CUSTOM_PROMPT,
model: "openai:o3-mini",
})
res = await graphTrajectoryEvaluator(
inputs=extractedTrajectory.inputs,
outputs=extractedTrajectory.outputs,
)In order to format them properly into the prompt, reference_outputs should be passed in as a GraphTrajectory object like outputs.
Also note that like other LLM-as-judge evaluators, you can pass extra kwargs into the evaluator to format them into the prompt.
The graph_trajectory_strict_match evaluator is a simple evaluator that checks if the steps in the provided graph trajectory match the reference trajectory exactly.
Python
from agentevals.graph_trajectory.utils import (
extract_langgraph_trajectory_from_thread,
)
from agentevals.graph_trajectory.strict import graph_trajectory_strict_match
from langgraph.prebuilt import create_react_agent
from langgraph.checkpoint.memory import MemorySaver
from langgraph.types import Command, interrupt
from langchain_core.tools import tool
@tool
def search(query: str):
"""Call to surf the web."""
user_answer = interrupt("Tell me the answer to the question.")
return user_answer
tools = [search]
checkpointer = MemorySaver()
graph = create_react_agent(
model="gpt-4o-mini",
checkpointer=checkpointer,
tools=[search],
)
graph.invoke(
{"messages": [{"role": "user", "content": "what's the weather in sf?"}]},
config={"configurable": {"thread_id": "1"}},
)
# Resume the agent with a new command, simulating a human-in-the-loop workflow
graph.invoke(
Command(resume="It is rainy and 70 degrees!"),
config={"configurable": {"thread_id": "1"}},
)
# Extract the trajectory from the first two thread runs
extracted_trajectory = extract_langgraph_trajectory_from_thread(
graph, {"configurable": {"thread_id": "1"}}
)
reference_trajectory = {
# not used for strict match
"results": [],
"steps": [["__start__", "agent", "tools", "__interrupt__"], ["agent"]],
}
res = graph_trajectory_strict_match(
outputs=extracted_trajectory["outputs"],
reference_outputs=reference_trajectory,
)
print(res){
'key': 'graph_trajectory_strict_match',
'score': True,
}
TypeScript
import { tool } from "@langchain/core/tools";
import { ChatOpenAI } from "@langchain/openai";
import { createReactAgent } from "@langchain/langgraph/prebuilt";
import { MemorySaver, interrupt } from "@langchain/langgraph";
import { z } from "zod";
import { extractLangGraphTrajectoryFromThread, graphTrajectoryStrictMatch } from "agentevals";
const search = tool((_): string => {
const userAnswer = interrupt("Tell me the answer to the question.")
return userAnswer;
}, {
name: "search",
description: "Call to surf the web.",
schema: z.object({
query: z.string()
})
})
const tools = [search];
// Create a checkpointer
const checkpointer = new MemorySaver();
// Create the React agent
const graph = createReactAgent({
llm: new ChatOpenAI({ model: "gpt-4o-mini" }),
tools,
checkpointer,
});
// Invoke the graph with initial message
await graph.invoke(
{ messages: [{ role: "user", content: "what's the weather in sf?" }] },
{ configurable: { thread_id: "1" } }
);
// Resume the agent with a new command (simulating human-in-the-loop)
await graph.invoke(
{ messages: [{ role: "user", content: "It is rainy and 70 degrees!" }] },
{ configurable: { thread_id: "1" } }
);
const extractedTrajectory = await extractLangGraphTrajectoryFromThread(
graph,
{ configurable: { thread_id: "1" } },
);
const referenceTrajectory = {
results: [],
steps: [["__start__", "agent", "tools", "__interrupt__"], ["agent"]],
}
const result = await graphTrajectoryStrictMatch({
outputs: trajectory.outputs,
referenceOutputs: referenceOutputs!,
});
console.log(result);{
'key': 'graph_trajectory_strict_match',
'score': True,
}
All agentevals evaluators support Python asyncio. As a convention, evaluators that use a factory function will have async put immediately after create_ in the function name (for example, create_async_trajectory_llm_as_judge), and evaluators used directly will end in async (e.g. trajectory_strict_match_async).
Here's an example of how to use the create_async_llm_as_judge evaluator asynchronously:
from agentevals.trajectory.llm import create_async_trajectory_llm_as_judge
evaluator = create_async_llm_as_judge(
prompt="What is the weather in {inputs}?",
)
result = await evaluator(inputs="San Francisco")If you are using the OpenAI client directly, remember to pass in AsyncOpenAI as the judge parameter:
from openai import AsyncOpenAI
evaluator = create_async_llm_as_judge(
prompt="What is the weather in {inputs}?",
judge=AsyncOpenAI(),
model="o3-mini",
)
result = await evaluator(inputs="San Francisco")For tracking experiments over time, you can log evaluator results to LangSmith, a platform for building production-grade LLM applications that includes tracing, evaluation, and experimentation tools.
LangSmith currently offers two ways to run evals: a pytest (Python) or Vitest/Jest integration and the evaluate function. We'll give a quick example of how to run evals using both.
First, follow these instructions to set up LangSmith's pytest runner, or these to set up Vitest or Jest, setting appropriate environment variables:
export LANGSMITH_API_KEY="your_langsmith_api_key"
export LANGSMITH_TRACING="true"Python
Then, set up a file named test_trajectory.py with the following contents:
import pytest
import json
from langsmith import testing as t
from agentevals.trajectory.llm import create_trajectory_llm_as_judge
trajectory_evaluator = create_trajectory_llm_as_judge(
model="openai:o3-mini",
)
@pytest.mark.langsmith
def test_trajectory_accuracy():
outputs = [
{"role": "user", "content": "What is the weather in SF?"},
{
"role": "assistant",
"tool_calls": [
{
"function": {
"name": "get_weather",
"arguments": json.dumps({"city": "SF"}),
}
}
],
},
{"role": "tool", "content": "It's 80 degrees and sunny in SF."},
{"role": "assistant", "content": "The weather in SF is 80 degrees and sunny."},
]
reference_outputs = [
{"role": "user", "content": "What is the weather in SF?"},
{
"role": "assistant",
"tool_calls": [
{
"function": {
"name": "get_weather",
"arguments": json.dumps({"city": "San Francisco"}),
}
}
],
},
{"role": "tool", "content": "It's 80 degrees and sunny in San Francisco."},
{"role": "assistant", "content": "The weather in SF is 80˚ and sunny."},
]
t.log_inputs({})
t.log_outputs({"messages": outputs})
t.log_reference_outputs({"messages": reference_outputs})
trajectory_evaluator(
outputs=outputs,
reference_outputs=reference_outputs
)Note that when creating the evaluator, we've added a feedback_key parameter. This will be used to name the feedback in LangSmith.
Now, run the eval with pytest:
pytest test_trajectory.py --langsmith-outputTypeScript
Then, set up a file named test_trajectory.eval.ts with the following contents:
import * as ls from "langsmith/vitest";
// import * as ls from "langsmith/jest";
import { createTrajectoryLLMAsJudge } from "agentevals";
const trajectoryEvaluator = createTrajectoryLLMAsJudge({
model: "openai:o3-mini",
});
ls.describe("trajectory accuracy", () => {
ls.test("accurate trajectory", {
inputs: {
messages: [
{
role: "user",
content: "What is the weather in SF?"
}
]
},
referenceOutputs: {
messages: [
{"role": "user", "content": "What is the weather in SF?"},
{
"role": "assistant",
"tool_calls": [
{
"function": {
"name": "get_weather",
"arguments": JSON.stringify({"city": "San Francisco"}),
}
}
],
},
{"role": "tool", "content": "It's 80 degrees and sunny in San Francisco."},
{"role": "assistant", "content": "The weather in SF is 80˚ and sunny."},
],
},
}, async ({ inputs, referenceOutputs }) => {
const outputs = [
{"role": "user", "content": "What is the weather in SF?"},
{
"role": "assistant",
"tool_calls": [
{
"function": {
"name": "get_weather",
"arguments": JSON.stringify({"city": "SF"}),
}
}
],
},
{"role": "tool", "content": "It's 80 degrees and sunny in SF."},
{"role": "assistant", "content": "The weather in SF is 80 degrees and sunny."},
];
ls.logOutputs({ messages: outputs });
await trajectoryEvaluator({
inputs,
outputs,
referenceOutputs,
});
});
});Now, run the eval with your runner of choice:
vitest run test_trajectory.eval.tsFeedback from the prebuilt evaluator will be automatically logged in LangSmith as a table of results like this in your terminal:
And you should also see the results in the experiment view in LangSmith:
Alternatively, you can create a dataset in LangSmith and use your created evaluators with LangSmith's evaluate function:
Python
from langsmith import Client
from agentevals.trajectory.llm import create_trajectory_llm_as_judge
client = Client()
trajectory_evaluator = create_trajectory_llm_as_judge(
model="openai:o3-mini",
)
experiment_results = client.evaluate(
# This is a dummy target function, replace with your actual LLM-based system
lambda inputs: "What color is the sky?",
data="Sample dataset",
evaluators=[
trajectory_evaluator
]
)TypeScript
import { evaluate } from "langsmith/evaluation";
import { createTrajectoryLLMAsJudge, TRAJECTORY_ACCURACY_PROMPT } from "agentevals";
const trajectoryEvaluator = createTrajectoryLLMAsJudge({
model: "openai:o3-mini",
prompt: TRAJECTORY_ACCURACY_PROMPT
});
await evaluate(
(inputs) => [
{role: "user", content: "What is the weather in SF?"},
{
role: "assistant",
tool_calls: [
{
function: {
name: "get_weather",
arguments: json.dumps({"city": "SF"}),
}
}
],
},
{role: "tool", content: "It's 80 degrees and sunny in SF."},
{role: "assistant", content: "The weather in SF is 80 degrees and sunny."},
],
{
data: datasetName,
evaluators: [trajectoryEvaluator],
}
);We hope that agentevals helps make evaluating your LLM agents easier!
If you have any questions, comments, or suggestions, please open an issue or reach out to us on X @LangChainAI.
