评估复杂代理
关键概念
智能体评估 | 评估器 | LLM-as-judge 评估器
在本教程中，我们将构建一个帮助客户在数字音乐商店中导航的客服机器人。然后，我们将介绍三种最有效的聊天机器人评估方法：
我们将使用 LangGraph 构建我们的智能体，但此处展示的技术和 LangSmith 功能与框架无关。
设置
配置环境
让我们安装所需的依赖项：
pip install -U langgraph langchain[openai]
让我们为 OpenAI 和 LangSmith 设置环境变量：
⌵
import getpass
import os
def _set_env(var: str) -> None:
if not os.environ.get(var):
os.environ[var] = getpass.getpass(f"Set {var}: ")
os.environ["LANGSMITH_TRACING"] = "true"
_set_env("LANGSMITH_API_KEY")
_set_env("OPENAI_API_KEY")
下载数据库
我们将为本教程创建一个 SQLite 数据库。SQLite 是一个轻量级的数据库，易于设置和使用。 我们将加载 chinook 数据库，这是一个代表数字媒体商店的示例数据库。 有关该数据库的更多信息，请点击此处查看。
为方便起见，我们将数据库托管在公共 GCS 存储桶中：
⌵
import requests
url = "https://storage.googleapis.com/benchmarks-artifacts/chinook/Chinook.db"
response = requests.get(url)
if response.status_code == 200:
# Open a local file in binary write mode
with open("chinook.db", "wb") as file:
# Write the content of the response (the file) to the local file
file.write(response.content)
print("File downloaded and saved as Chinook.db")
else:
print(f"Failed to download the file. Status code: {response.status_code}")
这是数据库中的数据示例：
⌵
import sqlite3 ...
[(1, 'AC/DC'), (2, 'Accept'), (3, 'Aerosmith'), (4, 'Alanis Morissette'), (5, 'Alice In Chains'), (6, 'Antônio Carlos Jobim'), (7, 'Apocalyptica'), (8, 'Audioslave'), (9, 'BackBeat'), (10, 'Billy Cobham')]
这是数据库模式（图片来自 https://github.com/lerocha/chinook-database）：
定义客户支持代理
我们将创建一个 LangGraph 智能体，该智能体对数据库的访问权限受到限制。出于演示目的，我们的智能体将支持两种基本类型的请求：
- 查找：客户可以根据其他识别信息查找歌曲标题、艺术家名称和专辑。例如：“您有哪些吉米·亨德里克斯的歌曲？”
- 退款：客户可以就其过去的购买申请退款。例如：“我的名字是克劳德·香农，我想对上周的一笔购买申请退款，您能帮我吗？”
为了简化本演示，我们将通过删除相应的数据库记录来实现退款功能。我们将跳过用户身份验证和其他生产环境安全措施的实现。
代理的逻辑将构建为两个独立的子图（一个用于查询，另一个用于退款），并有一个父图将请求路由到相应的子图。
退款代理
让我们构建退款处理代理。该代理需要：
- 在数据库中查找客户的购买记录
- 删除相关的发票和发票行记录以处理退款
我们将创建两个 SQL 辅助函数：
- 一个通过删除记录来执行退款的功能
- 用于查找客户购买历史的功能
为了简化测试，我们将为这些函数添加一个“模拟”模式。当启用模拟模式时，这些函数将模拟数据库操作，而不会实际修改任何数据。
⌵
⌵
import sqlite3
def _refund(invoice_id: int | None, invoice_line_ids: list[int] | None, mock: bool = False) -> float: ...
def _lookup( ...
现在让我们定义我们的图。我们将使用一个简单的架构，包含三个主要路径：
- 从对话中提取客户和购买信息
- 将请求路由到以下三个路径之一：
- 退款路径：如果我们拥有足够的购买详情（发票 ID 或发票行 ID）来处理退款
- 查询路径：如果我们拥有足够的客户信息（姓名和电话号码）来搜索其购买历史
- 响应路径：如果我们需要更多信息，则回复用户并请求所需的具体详情
图的狀態将跟踪：
- 对话历史记录（用户与代理之间的消息）
- 从对话中提取的所有客户和购买信息
- 下一条发送给用户的消息（后续文本）
⌵
⌵
⌵
⌵
⌵
⌵
⌵
from typing import Literal
import json
from langchain.chat_models import init_chat_model
from langchain_core.runnables import RunnableConfig
from langgraph.graph import END, StateGraph
from langgraph.graph.message import AnyMessage, add_messages
from langgraph.types import Command, interrupt
from tabulate import tabulate
from typing_extensions import Annotated, TypedDict
# Graph state.
class State(TypedDict):
"""Agent state."""
messages: Annotated[list[AnyMessage], add_messages]
followup: str | None
invoice_id: int | None
invoice_line_ids: list[int] | None
customer_first_name: str | None
customer_last_name: str | None
customer_phone: str | None
track_name: str | None
album_title: str | None
artist_name: str | None
purchase_date_iso_8601: str | None
# Instructions for extracting the user/purchase info from the conversation.
gather_info_instructions = """You are managing an online music store that sells song tracks. \
Customers can buy multiple tracks at a time and these purchases are recorded in a database as \
an Invoice per purchase and an associated set of Invoice Lines for each purchased track.
Your task is to help customers who would like a refund for one or more of the tracks they've \
purchased. In order for you to be able refund them, the customer must specify the Invoice ID \
to get a refund on all the tracks they bought in a single transaction, or one or more Invoice \
Line IDs if they would like refunds on individual tracks.
Often a user will not know the specific Invoice ID(s) or Invoice Line ID(s) for which they \
would like a refund. In this case you can help them look up their invoices by asking them to \
specify:
- Required: Their first name, last name, and phone number.
- Optionally: The track name, artist name, album name, or purchase date.
If the customer has not specified the required information (either Invoice/Invoice Line IDs \
or first name, last name, phone) then please ask them to specify it."""
# Extraction schema, mirrors the graph state.
class PurchaseInformation(TypedDict):
"""All of the known information about the invoice / invoice lines the customer would like refunded. Do not make up values, leave fields as null if you don't know their value."""
invoice_id: int | None
invoice_line_ids: list[int] | None
customer_first_name: str | None
customer_last_name: str | None
customer_phone: str | None
track_name: str | None
album_title: str | None
artist_name: str | None
purchase_date_iso_8601: str | None
followup: Annotated[
str | None,
...,
"If the user hasn't enough identifying information, please tell them what the required information is and ask them to specify it.",
]
# Model for performing extraction.
info_llm = init_chat_model("gpt-4o-mini").with_structured_output(
PurchaseInformation, method="json_schema", include_raw=True
)
# Graph node for extracting user info and routing to lookup/refund/END.
async def gather_info(state: State) -> Command[Literal["lookup", "refund", END]]:
info = await info_llm.ainvoke(
[
{"role": "system", "content": gather_info_instructions},
*state["messages"],
]
)
parsed = info["parsed"]
if any(parsed[k] for k in ("invoice_id", "invoice_line_ids")):
goto = "refund"
elif all(
parsed[k]
for k in ("customer_first_name", "customer_last_name", "customer_phone")
):
goto = "lookup"
else:
goto = END
update = {"messages": [info["raw"]], **parsed}
return Command(update=update, goto=goto)
# Graph node for executing the refund.
# Note that here we inspect the runtime config for an "env" variable.
# If "env" is set to "test", then we don't actually delete any rows from our database.
# This will become important when we're running our evaluations.
def refund(state: State, config: RunnableConfig) -> dict:
# Whether to mock the deletion. True if the configurable var 'env' is set to 'test'.
mock = config.get("configurable", {}).get("env", "prod") == "test"
refunded = _refund(
invoice_id=state["invoice_id"], invoice_line_ids=state["invoice_line_ids"], mock=mock
)
response = f"You have been refunded a total of: ${refunded:.2f}. Is there anything else I can help with?"
return {
"messages": [{"role": "assistant", "content": response}],
"followup": response,
}
# Graph node for looking up the users purchases
def lookup(state: State) -> dict:
args = (
state[k]
for k in (
"customer_first_name",
"customer_last_name",
"customer_phone",
"track_name",
"album_title",
"artist_name",
"purchase_date_iso_8601",
)
)
results = _lookup(*args)
if not results:
response = "We did not find any purchases associated with the information you've provided. Are you sure you've entered all of your information correctly?"
followup = response
else:
response = f"Which of the following purchases would you like to be refunded for?\n\n```json{json.dumps(results, indent=2)}\n```"
followup = f"Which of the following purchases would you like to be refunded for?\n\n{tabulate(results, headers='keys')}"
return {
"messages": [{"role": "assistant", "content": response}],
"followup": followup,
"invoice_line_ids": [res["invoice_line_id"] for res in results],
}
# Building our graph
graph_builder = StateGraph(State)
graph_builder.add_node(gather_info)
graph_builder.add_node(refund)
graph_builder.add_node(lookup)
graph_builder.set_entry_point("gather_info")
graph_builder.add_edge("lookup", END)
graph_builder.add_edge("refund", END)
refund_graph = graph_builder.compile()
我们可以可视化我们的退款图：
⌵
# Assumes you're in an interactive Python environment
from IPython.display import Image, display ...
查找代理
对于查找（即问答）代理，我们将使用一个简单的 ReACT 架构，并为代理提供工具，以便根据各种过滤器查找曲目名称、艺术家名称和专辑名称。例如，您可以按特定艺术家查找专辑，查找发布过特定名称歌曲的艺术家等。
⌵
⌵
⌵
⌵
⌵
⌵
from langchain.embeddings import init_embeddings
from langchain_core.tools import tool
from langchain_core.vectorstores import InMemoryVectorStore
from langgraph.prebuilt import create_react_agent
# Our SQL queries will only work if we filter on the exact string values that are in the DB.
# To ensure this, we'll create vectorstore indexes for all of the artists, tracks and albums
# ahead of time and use those to disambiguate the user input. E.g. if a user searches for
# songs by "prince" and our DB records the artist as "Prince", ideally when we query our
# artist vectorstore for "prince" we'll get back the value "Prince", which we can then
# use in our SQL queries.
def index_fields() -> tuple[InMemoryVectorStore, InMemoryVectorStore, InMemoryVectorStore]: ...
track_store, artist_store, album_store = index_fields()
# Agent tools
@tool
def lookup_track( ...
@tool
def lookup_album( ...
@tool
def lookup_artist( ...
# Agent model
qa_llm = init_chat_model("claude-3-5-sonnet-latest")
# The prebuilt ReACT agent only expects State to have a 'messages' key, so the
# state we defined for the refund agent can also be passed to our lookup agent.
qa_graph = create_react_agent(qa_llm, [lookup_track, lookup_artist, lookup_album])
display(Image(qa_graph.get_graph(xray=True).draw_mermaid_png()))
父级代理
现在让我们定义一个父代理，它结合了我们两个特定任务的代理。 父代理的唯一工作是通过分类用户的当前意图路由到其中一个子代理，并将输出编译为后续消息。
⌵
⌵
⌵
⌵
# Schema for routing user intent.
# We'll use structured outputs to enforce that the model returns only
# the desired output.
class UserIntent(TypedDict):
"""The user's current intent in the conversation"""
intent: Literal["refund", "question_answering"]
# Routing model with structured output
router_llm = init_chat_model("gpt-4o-mini").with_structured_output(
UserIntent, method="json_schema", strict=True
)
# Instructions for routing.
route_instructions = """You are managing an online music store that sells song tracks. \
You can help customers in two types of ways: (1) answering general questions about \
tracks sold at your store, (2) helping them get a refund on a purhcase they made at your store.
Based on the following conversation, determine if the user is currently seeking general \
information about song tracks or if they are trying to refund a specific purchase.
Return 'refund' if they are trying to get a refund and 'question_answering' if they are \
asking a general music question. Do NOT return anything else. Do NOT try to respond to \
the user.
"""
# Node for routing.
async def intent_classifier(
state: State,
) -> Command[Literal["refund_agent", "question_answering_agent"]]:
response = router_llm.invoke(
[{"role": "system", "content": route_instructions}, *state["messages"]]
)
return Command(goto=response["intent"] + "_agent")
# Node for making sure the 'followup' key is set before our agent run completes.
def compile_followup(state: State) -> dict:
"""Set the followup to be the last message if it hasn't explicitly been set."""
if not state.get("followup"):
return {"followup": state["messages"][-1].content}
return {}
# Agent definition
graph_builder = StateGraph(State)
graph_builder.add_node(intent_classifier)
# Since all of our subagents have compatible state,
# we can add them as nodes directly.
graph_builder.add_node("refund_agent", refund_graph)
graph_builder.add_node("question_answering_agent", qa_graph)
graph_builder.add_node(compile_followup)
graph_builder.set_entry_point("intent_classifier")
graph_builder.add_edge("refund_agent", "compile_followup")
graph_builder.add_edge("question_answering_agent", "compile_followup")
graph_builder.add_edge("compile_followup", END)
graph = graph_builder.compile()
我们可以可视化我们编译的父图，包括其所有子图：
display(Image(graph.get_graph().draw_mermaid_png()))
试试
让我们试用一下我们的自定义支持代理！
⌵
state = await graph.ainvoke(
{"messages": [{"role": "user", "content": "what james brown songs do you have"}]}
)
print(state["followup"])
⌵
I found 20 James Brown songs in the database, all from the album "Sex Machine". Here they are: ...
⌵
state = await graph.ainvoke({"messages": [
{
"role": "user",
"content": "my name is Aaron Mitchell and my number is +1 (204) 452-6452. I bought some songs by Led Zeppelin that i'd like refunded",
}
]})
print(state["followup"])
⌵
Which of the following purchases would you like to be refunded for? ...
评估
现在我们已经有了一个可测试的智能体版本，让我们运行一些评估。 智能体评估可以至少关注以下三个方面：
- 最终响应: 输入是一个提示词和可选的工具列表。输出是最终的智能体响应。
- 轨迹: 与之前一样，输入是一个提示词和可选的工具列表。输出是工具调用列表
- 单步操作: 与之前一样，输入是一个提示词和一个可选的工具列表。输出是工具调用。
让我们运行每种类型的评估：
最终响应评估器
首先，让我们创建一个 数据集，用于评估代理的端到端性能。 为简化起见，我们将使用相同的数据集进行最终响应和轨迹评估，因此我们将为每个示例问题添加真实响应和轨迹。 我们将在下一节中介绍轨迹。
⌵
⌵
from langsmith import Client
client = Client()
# Create a dataset
examples = [
{
"inputs": {
"question": "How many songs do you have by James Brown",
},
"outputs": {
"response": "We have 20 songs by James Brown",
"trajectory": ["question_answering_agent", "lookup_track"]
}
},
{
"inputs": {
"question": "My name is Aaron Mitchell and I'd like a refund.",
},
"outputs": {
"response": "I need some more information to help you with the refund. Please specify your phone number, the invoice ID, or the line item IDs for the purchase you'd like refunded.",
"trajectory": ["refund_agent"],
}
},
{
"inputs": {
"question": "My name is Aaron Mitchell and I'd like a refund on my Led Zeppelin purchases. My number is +1 (204) 452-6452",
},
"outputs": {
"response": 'Which of the following purchases would you like to be refunded for?\n\n invoice_line_id track_name artist_name purchase_date quantity_purchased price_per_unit\n----------------- -------------------------------- ------------- ------------------- -------------------- ----------------\n 267 How Many More Times Led Zeppelin 2009-08-06 00:00:00 1 0.99\n 268 What Is And What Should Never Be Led Zeppelin 2009-08-06 00:00:00 1 0.99',
"trajectory": ["refund_agent", "lookup"],
},
},
{
"inputs": {
"question": "Who recorded Wish You Were Here again? What other albums of there's do you have?",
},
"outputs": {
"response": "Wish You Were Here is an album by Pink Floyd",
"trajectory": ["question_answering_agent", "lookup_album"],
},
},
{
"inputs": {
"question": "I want a full refund for invoice 237",
},
"outputs": {
"response": "You have been refunded $0.99.",
"trajectory": ["refund_agent", "refund"],
}
},
]
dataset_name = "Chinook Customer Service Bot: E2E"
if not client.has_dataset(dataset_name=dataset_name):
dataset = client.create_dataset(dataset_name=dataset_name)
client.create_examples(
dataset_id=dataset.id,
examples=examples
)
我们将创建一个自定义的 LLM-as-judge 评估器，该评估器使用另一个模型将我们的代理在每个示例上的输出与参考响应进行比较，并判断它们是否等价：
⌵
⌵
⌵
# LLM-as-judge instructions
grader_instructions = """You are a teacher grading a quiz.
You will be given a QUESTION, the GROUND TRUTH (correct) RESPONSE, and the STUDENT RESPONSE.
Here is the grade criteria to follow:
(1) Grade the student responses based ONLY on their factual accuracy relative to the ground truth answer.
(2) Ensure that the student response does not contain any conflicting statements.
(3) It is OK if the student response contains more information than the ground truth response, as long as it is factually accurate relative to the ground truth response.
Correctness:
True means that the student's response meets all of the criteria.
False means that the student's response does not meet all of the criteria.
Explain your reasoning in a step-by-step manner to ensure your reasoning and conclusion are correct."""
# LLM-as-judge output schema
class Grade(TypedDict):
"""Compare the expected and actual answers and grade the actual answer."""
reasoning: Annotated[str, ..., "Explain your reasoning for whether the actual response is correct or not."]
is_correct: Annotated[bool, ..., "True if the student response is mostly or exactly correct, otherwise False."]
# Judge LLM
grader_llm = init_chat_model("gpt-4o-mini", temperature=0).with_structured_output(Grade, method="json_schema", strict=True)
# Evaluator function
async def final_answer_correct(inputs: dict, outputs: dict, reference_outputs: dict) -> bool:
"""Evaluate if the final response is equivalent to reference response."""
# Note that we assume the outputs has a 'response' dictionary. We'll need to make sure
# that the target function we define includes this key.
user = f"""QUESTION: {inputs['question']}
GROUND TRUTH RESPONSE: {reference_outputs['response']}
STUDENT RESPONSE: {outputs['response']}"""
grade = await grader_llm.ainvoke([{"role": "system", "content": grader_instructions}, {"role": "user", "content": user}])
return grade["is_correct"]
现在我们可以运行我们的评估。我们的评估器假设我们的目标函数返回一个'response'键，因此让我们定义一个这样做的目标函数。
还要记住，在我们的退款图中，我们使退款节点可配置，这样如果我们指定config={"env": "test"}，就会模拟退款而无需实际更新数据库。 我们将在调用图时，在目标run_graph方法中使用这个可配置的变量：
⌵
# Target function
async def run_graph(inputs: dict) -> dict:
"""Run graph and track the trajectory it takes along with the final response."""
result = await graph.ainvoke({"messages": [
{ "role": "user", "content": inputs['question']},
]}, config={"env": "test"})
return {"response": result["followup"]}
# Evaluation job and results
experiment_results = await client.aevaluate(
run_graph,
data=dataset_name,
evaluators=[final_answer_correct],
experiment_prefix="sql-agent-gpt4o-e2e",
num_repetitions=1,
max_concurrency=4,
)
experiment_results.to_pandas()
您可以在这里查看这些结果的样子：LangSmith 链接。
轨迹评估器
随着代理变得越来越复杂，它们出现潜在故障点的机会也更多。与其使用简单的通过/失败评估，不如采用能够给予部分评分的评估方式：即使代理未能得出正确答案，只要它采取了某些正确的步骤，也能获得相应的分数。
这就是轨迹评估发挥作用的地方。一个轨迹评估：
- 比较代理实际执行的步骤序列与预期步骤序列
- 根据正确完成的预期步骤数量计算分数
对于本示例，我们的端到端数据集包含一个有序的步骤列表，我们期望代理执行这些步骤。让我们创建一个评估器，用于将代理的实际轨迹与这些预期步骤进行比较，并计算完成步骤的百分比：
⌵
def trajectory_subsequence(outputs: dict, reference_outputs: dict) -> float:
"""Check how many of the desired steps the agent took."""
if len(reference_outputs['trajectory']) > len(outputs['trajectory']):
return False
i = j = 0
while i < len(reference_outputs['trajectory']) and j < len(outputs['trajectory']):
if reference_outputs['trajectory'][i] == outputs['trajectory'][j]:
i += 1
j += 1
return i / len(reference_outputs['trajectory'])
现在我们可以运行评估。我们的评估器假设目标函数返回一个'trajectory'键，因此让我们定义一个这样做的目标函数。我们需要使用LangGraph的流式传输功能来记录轨迹。
请注意，我们复用了与最终响应评估相同的数据集，因此我们可以将两个评估器一起运行，并定义一个同时返回“response”和" }
⌵
async def run_graph(inputs: dict) -> dict:
"""Run graph and track the trajectory it takes along with the final response."""
trajectory = []
# Set subgraph=True to stream events from subgraphs of the main graph: https://langchain-ai.github.io/langgraph/how-tos/streaming-subgraphs/
# Set stream_mode="debug" to stream all possible events: https://langchain-ai.github.io/langgraph/concepts/streaming
async for namespace, chunk in graph.astream({"messages": [
{
"role": "user",
"content": inputs['question'],
}
]}, subgraphs=True, stream_mode="debug"):
# Event type for entering a node
if chunk['type'] == 'task':
# Record the node name
trajectory.append(chunk['payload']['name'])
# Given how we defined our dataset, we also need to track when specific tools are
# called by our question answering ReACT agent. These tool calls can be found
# when the ToolsNode (named "tools") is invoked by looking at the AIMessage.tool_calls
# of the latest input message.
if chunk['payload']['name'] == 'tools' and chunk['type'] == 'task':
for tc in chunk['payload']['input']['messages'][-1].tool_calls:
trajectory.append(tc['name'])
return {"trajectory": trajectory}
experiment_results = await client.aevaluate(
run_graph,
data=dataset_name,
evaluators=[trajectory_subsequence],
experiment_prefix="sql-agent-gpt4o-trajectory",
num_repetitions=1,
max_concurrency=4,
)
experiment_results.to_pandas()
您可以在这里查看这些结果的样子：LangSmith 链接。
单步评估器
虽然端到端测试能为您提供关于智能体性能的最多信号，但为了调试和迭代您的智能体， pinpoint 特定难以处理的步骤并直接评估它们可能很有帮助。
在我们的案例中，代理的一个关键部分是将用户的意图正确路由到“退款”路径或“问答”路径。让我们创建一个数据集并运行一些评估，以直接对这一组件进行压力测试。
⌵
⌵
# Create dataset
examples = [
{
"inputs": {"messages": [{"role": "user", "content": "i bought some tracks recently and i dont like them"}]},
"outputs": {"route": "refund_agent"},
},
{
"inputs": {"messages": [{"role": "user", "content": "I was thinking of purchasing some Rolling Stones tunes, any recommendations?"}]},
"outputs": {"route": "question_answering_agent"},
},
{
"inputs": {"messages": [{"role": "user", "content": "i want a refund on purchase 237"}, {"role": "assistant", "content": "I've refunded you a total of $1.98. How else can I help you today?"}, {"role": "user", "content": "did prince release any albums in 2000?"}]},
"outputs": {"route": "question_answering_agent"},
},
{
"inputs": {"messages": [{"role": "user", "content": "i purchased a cover of Yesterday recently but can't remember who it was by, which versions of it do you have?"}]},
"outputs": {"route": "question_answering_agent"},
},
]
dataset_name = "Chinook Customer Service Bot: Intent Classifier"
if not client.has_dataset(dataset_name=dataset_name):
dataset = client.create_dataset(dataset_name=dataset_name)
client.create_examples(
dataset_id=dataset.id,
examples=examples
)
# Evaluator
def correct(outputs: dict, reference_outputs: dict) -> bool:
"""Check if the agent chose the correct route."""
return outputs["route"] == reference_outputs["route"]
# Target function for running the relevant step
async def run_intent_classifier(inputs: dict) -> dict:
# Note that we can access and run the intent_classifier node of our graph directly.
command = await graph.nodes['intent_classifier'].ainvoke(inputs)
return {"route": command.goto}
# Run evaluation
experiment_results = await client.aevaluate(
run_intent_classifier,
data=dataset_name,
evaluators=[correct],
experiment_prefix="sql-agent-gpt4o-intent-classifier",
max_concurrency=4,
)
您可以在这里查看这些结果的样子：LangSmith 链接。
参考代码
这是上述所有代码的整合脚本：
⌵
import json ...