用于与 SQL 数据库交互的 Agent¶
在本教程中，我们将逐步介绍如何构建一个可以回答有关 SQL 数据库问题的 Agent。
在较高层面上，Agent 将：1. 从数据库中获取可用的表 2. 确定哪些表与问题相关 3. 获取相关表的 DDL 4. 根据问题和来自 DDL 的信息生成查询 5. 使用 LLM 检查查询中常见的错误 6. 执行查询并返回结果 7. 纠正数据库引擎发现的错误，直到查询成功 8. 根据结果制定响应
端到端的工作流程如下所示
设置¶
首先，让我们安装所需的软件包并设置 API 密钥
import getpass
import os
def _set_env(key: str):
if key not in os.environ:
os.environ[key] = getpass.getpass(f"{key}:")
_set_env("OPENAI_API_KEY")
设置 LangSmith 以进行 LangGraph 开发
注册 LangSmith 以快速发现问题并提高 LangGraph 项目的性能。LangSmith 允许您使用跟踪数据来调试、测试和监控使用 LangGraph 构建的 LLM 应用程序——在此处阅读有关如何开始使用的更多信息 here。
配置数据库¶
在本教程中，我们将创建一个 SQLite 数据库。SQLite 是一个轻量级数据库，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。在此处查找有关数据库的更多信息 here。
为了方便起见，我们已将数据库 (Chinook.db) 托管在公共 GCS 存储桶上。
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}")
langchain_community 包中提供的便捷 SQL 数据库包装器与数据库进行交互。该包装器提供了一个简单的接口来执行 SQL 查询并获取结果。我们还将使用 langchain_openai 包在教程后面与 OpenAI API for language models 进行交互。
from langchain_community.utilities import SQLDatabase
db = SQLDatabase.from_uri("sqlite:///Chinook.db")
print(db.dialect)
print(db.get_usable_table_names())
db.run("SELECT * FROM Artist LIMIT 10;")
API 参考：SQLDatabase
sqlite
['Album', 'Artist', 'Customer', 'Employee', 'Genre', 'Invoice', 'InvoiceLine', 'MediaType', 'Playlist', 'PlaylistTrack', 'Track']
"[(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')]"
实用函数¶
我们将定义一些实用函数来帮助我们实现 Agent。具体来说，我们将使用回退包装 ToolNode 以处理错误并将其呈现给 Agent。
from typing import Any
from langchain_core.messages import ToolMessage
from langchain_core.runnables import RunnableLambda, RunnableWithFallbacks
from langgraph.prebuilt import ToolNode
def create_tool_node_with_fallback(tools: list) -> RunnableWithFallbacks[Any, dict]:
"""
Create a ToolNode with a fallback to handle errors and surface them to the agent.
"""
return ToolNode(tools).with_fallbacks(
[RunnableLambda(handle_tool_error)], exception_key="error"
)
def handle_tool_error(state) -> dict:
error = state.get("error")
tool_calls = state["messages"][-1].tool_calls
return {
"messages": [
ToolMessage(
content=f"Error: {repr(error)}\n please fix your mistakes.",
tool_call_id=tc["id"],
)
for tc in tool_calls
]
}
API 参考：ToolMessage | RunnableLambda | RunnableWithFallbacks | ToolNode
为 Agent 定义工具¶
我们将定义一些 Agent 将用于与数据库交互的工具。
list_tables_tool：从数据库中获取可用的表get_schema_tool：获取表的 DDLdb_query_tool：执行查询并获取结果；如果查询失败，则返回错误消息
对于前两个工具，我们将从 SQLDatabaseToolkit 中获取它们，它也位于 langchain_community 包中。
from langchain_community.agent_toolkits import SQLDatabaseToolkit
from langchain_openai import ChatOpenAI
toolkit = SQLDatabaseToolkit(db=db, llm=ChatOpenAI(model="gpt-4o"))
tools = toolkit.get_tools()
list_tables_tool = next(tool for tool in tools if tool.name == "sql_db_list_tables")
get_schema_tool = next(tool for tool in tools if tool.name == "sql_db_schema")
print(list_tables_tool.invoke(""))
print(get_schema_tool.invoke("Artist"))
API 参考：SQLDatabaseToolkit | ChatOpenAI
Album, Artist, Customer, Employee, Genre, Invoice, InvoiceLine, MediaType, Playlist, PlaylistTrack, Track
CREATE TABLE "Artist" (
"ArtistId" INTEGER NOT NULL,
"Name" NVARCHAR(120),
PRIMARY KEY ("ArtistId")
)
/*
3 rows from Artist table:
ArtistId Name
1 AC/DC
2 Accept
3 Aerosmith
*/
db_query_tool，我们将针对数据库执行查询并返回结果。
from langchain_core.tools import tool
@tool
def db_query_tool(query: str) -> str:
"""
Execute a SQL query against the database and get back the result.
If the query is not correct, an error message will be returned.
If an error is returned, rewrite the query, check the query, and try again.
"""
result = db.run_no_throw(query)
if not result:
return "Error: Query failed. Please rewrite your query and try again."
return result
print(db_query_tool.invoke("SELECT * FROM Artist LIMIT 10;"))
API 参考：tool
[(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')]
from langchain_core.prompts import ChatPromptTemplate
query_check_system = """You are a SQL expert with a strong attention to detail.
Double check the SQLite query for common mistakes, including:
- Using NOT IN with NULL values
- Using UNION when UNION ALL should have been used
- Using BETWEEN for exclusive ranges
- Data type mismatch in predicates
- Properly quoting identifiers
- Using the correct number of arguments for functions
- Casting to the correct data type
- Using the proper columns for joins
If there are any of the above mistakes, rewrite the query. If there are no mistakes, just reproduce the original query.
You will call the appropriate tool to execute the query after running this check."""
query_check_prompt = ChatPromptTemplate.from_messages(
[("system", query_check_system), ("placeholder", "{messages}")]
)
query_check = query_check_prompt | ChatOpenAI(model="gpt-4o", temperature=0).bind_tools(
[db_query_tool], tool_choice="required"
)
query_check.invoke({"messages": [("user", "SELECT * FROM Artist LIMIT 10;")]})
API 参考：ChatPromptTemplate
AIMessage(content='', additional_kwargs={'tool_calls': [{'id': 'call_la8JTjHox6P1VjTqc15GSgdk', 'function': {'arguments': '{"query":"SELECT * FROM Artist LIMIT 10;"}', 'name': 'db_query_tool'}, 'type': 'function'}], 'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 20, 'prompt_tokens': 221, 'total_tokens': 241}, 'model_name': 'gpt-4o-2024-05-13', 'system_fingerprint': 'fp_a2ff031fb5', 'finish_reason': 'stop', 'logprobs': None}, id='run-dd7873ef-d2f7-4769-a5c0-e6776ec2c515-0', tool_calls=[{'name': 'db_query_tool', 'args': {'query': 'SELECT * FROM Artist LIMIT 10;'}, 'id': 'call_la8JTjHox6P1VjTqc15GSgdk', 'type': 'tool_call'}], usage_metadata={'input_tokens': 221, 'output_tokens': 20, 'total_tokens': 241})
定义工作流程¶
然后，我们将定义 Agent 的工作流程。Agent 将首先强制调用 list_tables_tool 以从数据库中获取可用的表，然后按照教程开头提到的步骤进行操作。
将 Pydantic 与 LangChain 一起使用
此 notebook 使用 Pydantic v2 BaseModel，这需要 langchain-core >= 0.3。使用 langchain-core < 0.3 会因混合使用 Pydantic v1 和 v2 BaseModels 而导致错误。
from typing import Annotated, Literal
from langchain_core.messages import AIMessage
from langchain_openai import ChatOpenAI
from pydantic import BaseModel, Field
from typing_extensions import TypedDict
from langgraph.graph import END, StateGraph, START
from langgraph.graph.message import AnyMessage, add_messages
# Define the state for the agent
class State(TypedDict):
messages: Annotated[list[AnyMessage], add_messages]
# Define a new graph
workflow = StateGraph(State)
# Add a node for the first tool call
def first_tool_call(state: State) -> dict[str, list[AIMessage]]:
return {
"messages": [
AIMessage(
content="",
tool_calls=[
{
"name": "sql_db_list_tables",
"args": {},
"id": "tool_abcd123",
}
],
)
]
}
def model_check_query(state: State) -> dict[str, list[AIMessage]]:
"""
Use this tool to double-check if your query is correct before executing it.
"""
return {"messages": [query_check.invoke({"messages": [state["messages"][-1]]})]}
workflow.add_node("first_tool_call", first_tool_call)
# Add nodes for the first two tools
workflow.add_node(
"list_tables_tool", create_tool_node_with_fallback([list_tables_tool])
)
workflow.add_node("get_schema_tool", create_tool_node_with_fallback([get_schema_tool]))
# Add a node for a model to choose the relevant tables based on the question and available tables
model_get_schema = ChatOpenAI(model="gpt-4o", temperature=0).bind_tools(
[get_schema_tool]
)
workflow.add_node(
"model_get_schema",
lambda state: {
"messages": [model_get_schema.invoke(state["messages"])],
},
)
# Describe a tool to represent the end state
class SubmitFinalAnswer(BaseModel):
"""Submit the final answer to the user based on the query results."""
final_answer: str = Field(..., description="The final answer to the user")
# Add a node for a model to generate a query based on the question and schema
query_gen_system = """You are a SQL expert with a strong attention to detail.
Given an input question, output a syntactically correct SQLite query to run, then look at the results of the query and return the answer.
DO NOT call any tool besides SubmitFinalAnswer to submit the final answer.
When generating the query:
Output the SQL query that answers the input question without a tool call.
Unless the user specifies a specific number of examples they wish to obtain, always limit your query to at most 5 results.
You can order the results by a relevant column to return the most interesting examples in the database.
Never query for all the columns from a specific table, only ask for the relevant columns given the question.
If you get an error while executing a query, rewrite the query and try again.
If you get an empty result set, you should try to rewrite the query to get a non-empty result set.
NEVER make stuff up if you don't have enough information to answer the query... just say you don't have enough information.
If you have enough information to answer the input question, simply invoke the appropriate tool to submit the final answer to the user.
DO NOT make any DML statements (INSERT, UPDATE, DELETE, DROP etc.) to the database."""
query_gen_prompt = ChatPromptTemplate.from_messages(
[("system", query_gen_system), ("placeholder", "{messages}")]
)
query_gen = query_gen_prompt | ChatOpenAI(model="gpt-4o", temperature=0).bind_tools(
[SubmitFinalAnswer]
)
def query_gen_node(state: State):
message = query_gen.invoke(state)
# Sometimes, the LLM will hallucinate and call the wrong tool. We need to catch this and return an error message.
tool_messages = []
if message.tool_calls:
for tc in message.tool_calls:
if tc["name"] != "SubmitFinalAnswer":
tool_messages.append(
ToolMessage(
content=f"Error: The wrong tool was called: {tc['name']}. Please fix your mistakes. Remember to only call SubmitFinalAnswer to submit the final answer. Generated queries should be outputted WITHOUT a tool call.",
tool_call_id=tc["id"],
)
)
else:
tool_messages = []
return {"messages": [message] + tool_messages}
workflow.add_node("query_gen", query_gen_node)
# Add a node for the model to check the query before executing it
workflow.add_node("correct_query", model_check_query)
# Add node for executing the query
workflow.add_node("execute_query", create_tool_node_with_fallback([db_query_tool]))
# Define a conditional edge to decide whether to continue or end the workflow
def should_continue(state: State) -> Literal[END, "correct_query", "query_gen"]:
messages = state["messages"]
last_message = messages[-1]
# If there is a tool call, then we finish
if getattr(last_message, "tool_calls", None):
return END
if last_message.content.startswith("Error:"):
return "query_gen"
else:
return "correct_query"
# Specify the edges between the nodes
workflow.add_edge(START, "first_tool_call")
workflow.add_edge("first_tool_call", "list_tables_tool")
workflow.add_edge("list_tables_tool", "model_get_schema")
workflow.add_edge("model_get_schema", "get_schema_tool")
workflow.add_edge("get_schema_tool", "query_gen")
workflow.add_conditional_edges(
"query_gen",
should_continue,
)
workflow.add_edge("correct_query", "execute_query")
workflow.add_edge("execute_query", "query_gen")
# Compile the workflow into a runnable
app = workflow.compile()
API 参考：AIMessage | ChatOpenAI | END | StateGraph | START | add_messages
可视化图¶
from IPython.display import Image, display
from langchain_core.runnables.graph import MermaidDrawMethod
display(
Image(
app.get_graph().draw_mermaid_png(
draw_method=MermaidDrawMethod.API,
)
)
)
API 参考：MermaidDrawMethod
运行 Agent¶
messages = app.invoke(
{"messages": [("user", "Which sales agent made the most in sales in 2009?")]}
)
json_str = messages["messages"][-1].tool_calls[0]["args"]["final_answer"]
json_str
for event in app.stream(
{"messages": [("user", "Which sales agent made the most in sales in 2009?")]}
):
print(event)
{'first_tool_call': {'messages': [AIMessage(content='', tool_calls=[{'name': 'sql_db_list_tables', 'args': {}, 'id': 'tool_abcd123', 'type': 'tool_call'}])]}}
{'list_tables_tool': {'messages': [ToolMessage(content='Album, Artist, Customer, Employee, Genre, Invoice, InvoiceLine, MediaType, Playlist, PlaylistTrack, Track', name='sql_db_list_tables', tool_call_id='tool_abcd123')]}}
{'model_get_schema': {'messages': [AIMessage(content='', additional_kwargs={'tool_calls': [{'id': 'call_z1tyC7cEAawi5oIQn731Uknp', 'function': {'arguments': '{"table_names":"Employee, Invoice"}', 'name': 'sql_db_schema'}, 'type': 'function'}], 'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 18, 'prompt_tokens': 177, 'total_tokens': 195}, 'model_name': 'gpt-4o-2024-05-13', 'system_fingerprint': 'fp_a2ff031fb5', 'finish_reason': 'tool_calls', 'logprobs': None}, id='run-c91a5aad-fc05-4881-87f9-0662d703c3c8-0', tool_calls=[{'name': 'sql_db_schema', 'args': {'table_names': 'Employee, Invoice'}, 'id': 'call_z1tyC7cEAawi5oIQn731Uknp', 'type': 'tool_call'}], usage_metadata={'input_tokens': 177, 'output_tokens': 18, 'total_tokens': 195})]}}
{'get_schema_tool': {'messages': [ToolMessage(content='\nCREATE TABLE "Employee" (\n\t"EmployeeId" INTEGER NOT NULL, \n\t"LastName" NVARCHAR(20) NOT NULL, \n\t"FirstName" NVARCHAR(20) NOT NULL, \n\t"Title" NVARCHAR(30), \n\t"ReportsTo" INTEGER, \n\t"BirthDate" DATETIME, \n\t"HireDate" DATETIME, \n\t"Address" NVARCHAR(70), \n\t"City" NVARCHAR(40), \n\t"State" NVARCHAR(40), \n\t"Country" NVARCHAR(40), \n\t"PostalCode" NVARCHAR(10), \n\t"Phone" NVARCHAR(24), \n\t"Fax" NVARCHAR(24), \n\t"Email" NVARCHAR(60), \n\tPRIMARY KEY ("EmployeeId"), \n\tFOREIGN KEY("ReportsTo") REFERENCES "Employee" ("EmployeeId")\n)\n\n/*\n3 rows from Employee table:\nEmployeeId\tLastName\tFirstName\tTitle\tReportsTo\tBirthDate\tHireDate\tAddress\tCity\tState\tCountry\tPostalCode\tPhone\tFax\tEmail\n1\tAdams\tAndrew\tGeneral Manager\tNone\t1962-02-18 00:00:00\t2002-08-14 00:00:00\t11120 Jasper Ave NW\tEdmonton\tAB\tCanada\tT5K 2N1\t+1 (780) 428-9482\t+1 (780) 428-3457\tandrew@chinookcorp.com\n2\tEdwards\tNancy\tSales Manager\t1\t1958-12-08 00:00:00\t2002-05-01 00:00:00\t825 8 Ave SW\tCalgary\tAB\tCanada\tT2P 2T3\t+1 (403) 262-3443\t+1 (403) 262-3322\tnancy@chinookcorp.com\n3\tPeacock\tJane\tSales Support Agent\t2\t1973-08-29 00:00:00\t2002-04-01 00:00:00\t1111 6 Ave SW\tCalgary\tAB\tCanada\tT2P 5M5\t+1 (403) 262-3443\t+1 (403) 262-6712\tjane@chinookcorp.com\n*/\n\n\nCREATE TABLE "Invoice" (\n\t"InvoiceId" INTEGER NOT NULL, \n\t"CustomerId" INTEGER NOT NULL, \n\t"InvoiceDate" DATETIME NOT NULL, \n\t"BillingAddress" NVARCHAR(70), \n\t"BillingCity" NVARCHAR(40), \n\t"BillingState" NVARCHAR(40), \n\t"BillingCountry" NVARCHAR(40), \n\t"BillingPostalCode" NVARCHAR(10), \n\t"Total" NUMERIC(10, 2) NOT NULL, \n\tPRIMARY KEY ("InvoiceId"), \n\tFOREIGN KEY("CustomerId") REFERENCES "Customer" ("CustomerId")\n)\n\n/*\n3 rows from Invoice table:\nInvoiceId\tCustomerId\tInvoiceDate\tBillingAddress\tBillingCity\tBillingState\tBillingCountry\tBillingPostalCode\tTotal\n1\t2\t2009-01-01 00:00:00\tTheodor-Heuss-Straße 34\tStuttgart\tNone\tGermany\t70174\t1.98\n2\t4\t2009-01-02 00:00:00\tUllevålsveien 14\tOslo\tNone\tNorway\t0171\t3.96\n3\t8\t2009-01-03 00:00:00\tGrétrystraat 63\tBrussels\tNone\tBelgium\t1000\t5.94\n*/', name='sql_db_schema', tool_call_id='call_z1tyC7cEAawi5oIQn731Uknp')]}}
{'query_gen': {'messages': [AIMessage(content='', additional_kwargs={'tool_calls': [{'id': 'call_ErWLktUfxKsHGNGr74m72yYD', 'function': {'arguments': '{"table_names":"Customer"}', 'name': 'sql_db_schema'}, 'type': 'function'}], 'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 16, 'prompt_tokens': 1179, 'total_tokens': 1195}, 'model_name': 'gpt-4o-2024-05-13', 'system_fingerprint': 'fp_a2ff031fb5', 'finish_reason': 'tool_calls', 'logprobs': None}, id='run-19e02169-5e1e-40d9-90a2-384336ca5069-0', tool_calls=[{'name': 'sql_db_schema', 'args': {'table_names': 'Customer'}, 'id': 'call_ErWLktUfxKsHGNGr74m72yYD', 'type': 'tool_call'}], usage_metadata={'input_tokens': 1179, 'output_tokens': 16, 'total_tokens': 1195}), ToolMessage(content='Error: The wrong tool was called: sql_db_schema. Please fix your mistakes. Remember to only call SubmitFinalAnswer to submit the final answer. Generated queries should be outputted WITHOUT a tool call.', id='de5d25f5-b891-4e47-8282-d04dc9b93e9e', tool_call_id='call_ErWLktUfxKsHGNGr74m72yYD')]}}
{'query_gen': {'messages': [AIMessage(content='', additional_kwargs={'tool_calls': [{'id': 'call_TFaA52SbhgEqm3ElEAd4HCsn', 'function': {'arguments': '{"table_names":["Customer"]}', 'name': 'sql_db_schema'}, 'type': 'function'}], 'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 17, 'prompt_tokens': 1245, 'total_tokens': 1262}, 'model_name': 'gpt-4o-2024-05-13', 'system_fingerprint': 'fp_a2ff031fb5', 'finish_reason': 'tool_calls', 'logprobs': None}, id='run-2c5f800f-43dc-4224-847b-49b5079efd2a-0', tool_calls=[{'name': 'sql_db_schema', 'args': {'table_names': ['Customer']}, 'id': 'call_TFaA52SbhgEqm3ElEAd4HCsn', 'type': 'tool_call'}], usage_metadata={'input_tokens': 1245, 'output_tokens': 17, 'total_tokens': 1262}), ToolMessage(content='Error: The wrong tool was called: sql_db_schema. Please fix your mistakes. Remember to only call SubmitFinalAnswer to submit the final answer. Generated queries should be outputted WITHOUT a tool call.', id='6c962a35-fc24-4f27-86f0-6ec05256d478', tool_call_id='call_TFaA52SbhgEqm3ElEAd4HCsn')]}}
{'query_gen': {'messages': [AIMessage(content="To determine which sales agent made the most in sales in 2009, we need to join the `Invoice`, `Customer`, and `Employee` tables. Here is the query to find the top sales agent:\n\n\`\`\`sql\nSELECT e.FirstName, e.LastName, SUM(i.Total) as TotalSales\nFROM Invoice i\nJOIN Customer c ON i.CustomerId = c.CustomerId\nJOIN Employee e ON c.SupportRepId = e.EmployeeId\nWHERE strftime('%Y', i.InvoiceDate) = '2009'\nGROUP BY e.EmployeeId\nORDER BY TotalSales DESC\nLIMIT 1;\n\`\`\`", additional_kwargs={'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 125, 'prompt_tokens': 1312, 'total_tokens': 1437}, 'model_name': 'gpt-4o-2024-05-13', 'system_fingerprint': 'fp_3aa7262c27', 'finish_reason': 'stop', 'logprobs': None}, id='run-6cacd10d-d3aa-49ae-b9d7-8cc209fc4ccc-0', usage_metadata={'input_tokens': 1312, 'output_tokens': 125, 'total_tokens': 1437})]}}
{'correct_query': {'messages': [AIMessage(content='', additional_kwargs={'tool_calls': [{'id': 'call_FwCE2c7WORU7lKHdSWqMv0ON', 'function': {'arguments': '{"query":"SELECT e.FirstName, e.LastName, SUM(i.Total) as TotalSales\\nFROM Invoice i\\nJOIN Customer c ON i.CustomerId = c.CustomerId\\nJOIN Employee e ON c.SupportRepId = e.EmployeeId\\nWHERE strftime(\'%Y\', i.InvoiceDate) = \'2009\'\\nGROUP BY e.EmployeeId\\nORDER BY TotalSales DESC\\nLIMIT 1;"}', 'name': 'db_query_tool'}, 'type': 'function'}], 'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 90, 'prompt_tokens': 337, 'total_tokens': 427}, 'model_name': 'gpt-4o-2024-05-13', 'system_fingerprint': 'fp_a2ff031fb5', 'finish_reason': 'stop', 'logprobs': None}, id='run-71067e75-80f6-4356-8239-518e466b3526-0', tool_calls=[{'name': 'db_query_tool', 'args': {'query': "SELECT e.FirstName, e.LastName, SUM(i.Total) as TotalSales\nFROM Invoice i\nJOIN Customer c ON i.CustomerId = c.CustomerId\nJOIN Employee e ON c.SupportRepId = e.EmployeeId\nWHERE strftime('%Y', i.InvoiceDate) = '2009'\nGROUP BY e.EmployeeId\nORDER BY TotalSales DESC\nLIMIT 1;"}, 'id': 'call_FwCE2c7WORU7lKHdSWqMv0ON', 'type': 'tool_call'}], usage_metadata={'input_tokens': 337, 'output_tokens': 90, 'total_tokens': 427})]}}
{'execute_query': {'messages': [ToolMessage(content="[('Steve', 'Johnson', 164.34)]", name='db_query_tool', tool_call_id='call_FwCE2c7WORU7lKHdSWqMv0ON')]}}
{'query_gen': {'messages': [AIMessage(content='', additional_kwargs={'tool_calls': [{'id': 'call_fHJ4lvdiFM9HY6gupE6vLZV4', 'function': {'arguments': '{"final_answer":"The sales agent who made the most in sales in 2009 is Steve Johnson with total sales of 164.34."}', 'name': 'SubmitFinalAnswer'}, 'type': 'function'}], 'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 41, 'prompt_tokens': 1553, 'total_tokens': 1594}, 'model_name': 'gpt-4o-2024-05-13', 'system_fingerprint': 'fp_cb7cc8e106', 'finish_reason': 'tool_calls', 'logprobs': None}, id='run-2ec7bf3a-2a16-47bd-aa9c-b7d6dc531c1b-0', tool_calls=[{'name': 'SubmitFinalAnswer', 'args': {'final_answer': 'The sales agent who made the most in sales in 2009 is Steve Johnson with total sales of 164.34.'}, 'id': 'call_fHJ4lvdiFM9HY6gupE6vLZV4', 'type': 'tool_call'}], usage_metadata={'input_tokens': 1553, 'output_tokens': 41, 'total_tokens': 1594})]}}
评估¶
现在，我们可以评估此 Agent！我们之前在 LangSmith 评估 Cookbook 中定义了 简单的 SQL Agent，并评估了对我们数据库的 5 个问题的响应。我们可以将此 Agent 与我们之前的 Agent 在同一数据集上进行比较。Agent 评估 可以关注 3 个方面
响应：输入是提示和工具列表。输出是 Agent 响应。单个工具：与之前一样，输入是提示和工具列表。输出是工具调用。轨迹：与之前一样，输入是提示和工具列表。输出是工具调用列表
响应¶
我们将评估 Agent 相对于参考答案的端到端响应。让我们在 同一数据集上 运行 响应评估。
import json
def predict_sql_agent_answer(example: dict):
"""Use this for answer evaluation"""
msg = {"messages": ("user", example["input"])}
messages = app.invoke(msg)
json_str = messages["messages"][-1].tool_calls[0]["args"]
response = json_str["final_answer"]
return {"response": response}
from langchain import hub
from langchain_openai import ChatOpenAI
# Grade prompt
grade_prompt_answer_accuracy = prompt = hub.pull("langchain-ai/rag-answer-vs-reference")
def answer_evaluator(run, example) -> dict:
"""
A simple evaluator for RAG answer accuracy
"""
# Get question, ground truth answer, chain
input_question = example.inputs["input"]
reference = example.outputs["output"]
prediction = run.outputs["response"]
# LLM grader
llm = ChatOpenAI(model="gpt-4-turbo", temperature=0)
# Structured prompt
answer_grader = grade_prompt_answer_accuracy | llm
# Run evaluator
score = answer_grader.invoke(
{
"question": input_question,
"correct_answer": reference,
"student_answer": prediction,
}
)
score = score["Score"]
return {"key": "answer_v_reference_score", "score": score}
API 参考：ChatOpenAI
from langsmith.evaluation import evaluate
dataset_name = "SQL Agent Response"
try:
experiment_results = evaluate(
predict_sql_agent_answer,
data=dataset_name,
evaluators=[answer_evaluator],
num_repetitions=3,
experiment_prefix="sql-agent-multi-step-response-v-reference",
metadata={"version": "Chinook, gpt-4o multi-step-agent"},
)
except:
print("Please setup LangSmith")
摘要指标（请参阅数据集 here）
- 此处的
multi-stepAgent 性能优于之前定义的 基本案例 SQL Agent
轨迹¶
让我们在此同一数据集上运行 轨迹评估。
# These are the tools that we expect the agent to use
expected_trajectory = [
"sql_db_list_tables", # first: list_tables_tool node
"sql_db_schema", # second: get_schema_tool node
"db_query_tool", # third: execute_query node
"SubmitFinalAnswer",
] # fourth: query_gen
def predict_sql_agent_messages(example: dict):
"""Use this for answer evaluation"""
msg = {"messages": ("user", example["input"])}
messages = app.invoke(msg)
return {"response": messages}
from langsmith.schemas import Example, Run
def find_tool_calls(messages):
"""
Find all tool calls in the messages returned
"""
tool_calls = [
tc["name"] for m in messages["messages"] for tc in getattr(m, "tool_calls", [])
]
return tool_calls
def contains_all_tool_calls_in_order_exact_match(
root_run: Run, example: Example
) -> dict:
"""
Check if all expected tools are called in exact order and without any additional tool calls.
"""
expected_trajectory = [
"sql_db_list_tables",
"sql_db_schema",
"db_query_tool",
"SubmitFinalAnswer",
]
messages = root_run.outputs["response"]
tool_calls = find_tool_calls(messages)
# Print the tool calls for debugging
print("Here are my tool calls:")
print(tool_calls)
# Check if the tool calls match the expected trajectory exactly
if tool_calls == expected_trajectory:
score = 1
else:
score = 0
return {"score": int(score), "key": "multi_tool_call_in_exact_order"}
def contains_all_tool_calls_in_order(root_run: Run, example: Example) -> dict:
"""
Check if all expected tools are called in order,
but it allows for other tools to be called in between the expected ones.
"""
messages = root_run.outputs["response"]
tool_calls = find_tool_calls(messages)
# Print the tool calls for debugging
print("Here are my tool calls:")
print(tool_calls)
it = iter(tool_calls)
if all(elem in it for elem in expected_trajectory):
score = 1
else:
score = 0
return {"score": int(score), "key": "multi_tool_call_in_order"}
try:
experiment_results = evaluate(
predict_sql_agent_messages,
data=dataset_name,
evaluators=[
contains_all_tool_calls_in_order,
contains_all_tool_calls_in_order_exact_match,
],
num_repetitions=3,
experiment_prefix="sql-agent-multi-step-tool-calling-trajecory-in-order",
metadata={"version": "Chinook, gpt-4o multi-step-agent"},
)
except:
print("Please setup LangSmith")
聚合分数表明我们从未以完全正确的顺序正确调用工具
查看日志记录，我们可以看到一些有趣的事情 -
我们似乎将一个幻觉工具调用 sql_db_query 注入到我们的大部分运行轨迹中。
这就是为什么 multi_tool_call_in_exact_order 失败，但 multi_tool_call_in_order 仍然通过的原因。
我们将在未来的 Cookbook 中探索使用 LangGraph 解决此问题的方法！