构建自定义 SQL 智能体 - AI 开发中文文档

在本教程中，我们将使用 LangGraph 构建一个自定义智能体，该智能体能够回答有关 SQL 数据库的问题。 LangChain 提供了内置的智能体实现，这些实现使用 LangGraph 基本结构实现。如果需要更深入的定制，智能体可以直接在 LangGraph 中实现。本指南演示了一个 SQL 智能体的示例实现。有关实际入门介绍，请参阅使用更高层次的 LangChain 抽象构建 SQL 智能体。

构建 SQL 数据库的 Q&A 系统需要执行模型生成的 SQL 查询。这样做存在固有的风险。确保您的数据库连接权限始终尽可能狭窄地限制在智能体的需求范围内。这将减轻（尽管不能完全消除）构建模型驱动系统的风险。

预构建的智能体可以让我们快速上手，但我们依赖于系统提示词来约束其行为——例如，我们指示智能体始终从 “列出表” 工具开始，并在执行查询之前始终运行查询检查器工具。我们可以通过自定义智能体在 LangGraph 中实现更高程度的控制。在这里，我们实现了一个简单的 ReAct 智能体设置，为特定的工具调用设置了专用节点。我们将使用与预构建智能体相同的 [状态]。

概念

我们将涵盖以下概念：

用于从 SQL 数据库读取的工具
LangGraph 的图 API，包括状态、节点、边和条件边。
人工参与流程

设置

安装

:::python

pip install langchain langgraph

::: :::js

npm i langchain @langchain/core @langchain/langgraph @langchain/openai sqlite3 zod

yarn add langchain @langchain/core @langchain/langgraph @langchain/openai sqlite3 zod

pnpm add langchain @langchain/core @langchain/langgraph @langchain/openai sqlite3 zod

:::

LangSmith

设置 LangSmith 以检查链或智能体内部的运行情况。然后设置以下环境变量：

export LANGSMITH_TRACING="true"
export LANGSMITH_API_KEY="..."

1. 选择 LLM

选择一个支持工具调用的模型： :::python

👉 请阅读 OpenAI 对话模型集成文档

pip install -U "langchain[openai]"

uv add "langchain[openai]"

import os
from langchain.chat_models import init_chat_model

os.environ["OPENAI_API_KEY"] = "sk-..."

model = init_chat_model("gpt-5.5")

import os
from langchain_openai import ChatOpenAI

os.environ["OPENAI_API_KEY"] = "sk-..."

model = ChatOpenAI(model="gpt-5.5")

👉 请阅读 Anthropic 对话模型集成文档

pip install -U "langchain[anthropic]"

uv add "langchain[anthropic]"

import os
from langchain.chat_models import init_chat_model

os.environ["ANTHROPIC_API_KEY"] = "sk-..."

model = init_chat_model("claude-sonnet-4-6")

import os
from langchain_anthropic import ChatAnthropic

os.environ["ANTHROPIC_API_KEY"] = "sk-..."

model = ChatAnthropic(model="claude-sonnet-4-6")

👉 请阅读 Azure 对话模型集成文档

pip install -U "langchain[openai]"

uv add "langchain[openai]"

import os
from langchain.chat_models import init_chat_model

os.environ["AZURE_OPENAI_API_KEY"] = "..."
os.environ["AZURE_OPENAI_ENDPOINT"] = "..."
os.environ["OPENAI_API_VERSION"] = "2025-03-01-preview"

model = init_chat_model(
    "azure_openai:gpt-5.5",
    azure_deployment=os.environ["AZURE_OPENAI_DEPLOYMENT_NAME"],
)

import os
from langchain_openai import AzureChatOpenAI

os.environ["AZURE_OPENAI_API_KEY"] = "..."
os.environ["AZURE_OPENAI_ENDPOINT"] = "..."
os.environ["OPENAI_API_VERSION"] = "2025-03-01-preview"

model = AzureChatOpenAI(
    model="gpt-5.5",
    azure_deployment=os.environ["AZURE_OPENAI_DEPLOYMENT_NAME"]
)

👉 请阅读 Google GenAI 对话模型集成文档

pip install -U "langchain[google-genai]"

uv add "langchain[google-genai]"

import os
from langchain.chat_models import init_chat_model

os.environ["GOOGLE_API_KEY"] = "..."

model = init_chat_model("google_genai:gemini-3.7-flash")

import os
from langchain_google_genai import ChatGoogleGenerativeAI

os.environ["GOOGLE_API_KEY"] = "..."

model = ChatGoogleGenerativeAI(model="gemini-3.7-flash")

👉 请阅读 AWS Bedrock 对话模型集成文档

pip install -U "langchain[aws]"

uv add "langchain[aws]"

from langchain.chat_models import init_chat_model

# Follow the steps here to configure your credentials:
# https://docs.aws.amazon.com/bedrock/latest/userguide/getting-started.html

model = init_chat_model(
    "us.anthropic.claude-sonnet-4-6",
    model_provider="bedrock_converse",
)

from langchain_aws import ChatBedrock

model = ChatBedrock(model="us.anthropic.claude-sonnet-4-6")

👉 请阅读 HuggingFace 对话模型集成文档

pip install -U "langchain[huggingface]"

uv add "langchain[huggingface]"

import os
from langchain.chat_models import init_chat_model

os.environ["HUGGINGFACEHUB_API_TOKEN"] = "hf_..."

model = init_chat_model(
    "microsoft/Phi-3-mini-4k-instruct",
    model_provider="huggingface",
    temperature=0.7,
    max_tokens=1024,
)

import os
from langchain_huggingface import ChatHuggingFace, HuggingFaceEndpoint

os.environ["HUGGINGFACEHUB_API_TOKEN"] = "hf_..."

llm = HuggingFaceEndpoint(
    repo_id="microsoft/Phi-3-mini-4k-instruct",
    temperature=0.7,
    max_length=1024,
)
model = ChatHuggingFace(llm=llm)

👉 请阅读 OpenRouter 对话模型集成文档

pip install -U "langchain-openrouter"

uv add "langchain-openrouter"

import os
from langchain.chat_models import init_chat_model

os.environ["OPENROUTER_API_KEY"] = "sk-..."

model = init_chat_model(
    "auto",
    model_provider="openrouter",
)

import os
from langchain_openrouter import ChatOpenRouter

os.environ["OPENROUTER_API_KEY"] = "sk-..."

model = ChatOpenRouter(model="auto")

::: :::js

OpenAI
Anthropic
Azure
Google Gemini
Bedrock Converse

👉 Read the OpenAI 对话模型集成文档

npm install @langchain/openai

pnpm install @langchain/openai

yarn add @langchain/openai

bun add @langchain/openai

import { initChatModel } from "langchain";

process.env.OPENAI_API_KEY = "your-api-key";

const model = await initChatModel("gpt-5.5");

import { ChatOpenAI } from "@langchain/openai";

const model = new ChatOpenAI({
  model: "gpt-5.5",
  apiKey: "your-api-key"
});

👉 Read the Anthropic 对话模型集成文档

npm install @langchain/anthropic

pnpm install @langchain/anthropic

yarn add @langchain/anthropic

pnpm add @langchain/anthropic

import { initChatModel } from "langchain";

process.env.ANTHROPIC_API_KEY = "your-api-key";

const model = await initChatModel("claude-sonnet-4-6");

import { ChatAnthropic } from "@langchain/anthropic";

const model = new ChatAnthropic({
  model: "claude-sonnet-4-6",
  apiKey: "your-api-key"
});

👉 Read the Azure 对话模型集成文档

npm install @langchain/azure

pnpm install @langchain/azure

yarn add @langchain/azure

bun add @langchain/azure

import { initChatModel } from "langchain";

process.env.AZURE_OPENAI_API_KEY = "your-api-key";
process.env.AZURE_OPENAI_ENDPOINT = "your-endpoint";
process.env.OPENAI_API_VERSION = "your-api-version";

const model = await initChatModel("azure_openai:gpt-5.5");

import { AzureChatOpenAI } from "@langchain/openai";

const model = new AzureChatOpenAI({
  model: "gpt-5.5",
  azureOpenAIApiKey: "your-api-key",
  azureOpenAIApiEndpoint: "your-endpoint",
  azureOpenAIApiVersion: "your-api-version"
});

👉 Read the ChatGoogle 对话模型集成文档

npm install @langchain/google

pnpm install @langchain/google

yarn add @langchain/google

bun add @langchain/google

import { initChatModel } from "langchain";

process.env.GOOGLE_API_KEY = "your-api-key";

const model = await initChatModel("google:gemini-3.7-flash");

import { ChatGoogle } from "@langchain/google/node";

const model = new ChatGoogle({
  model: "gemini-3.7-flash",
  apiKey: "your-api-key"
});

👉 Read the AWS Bedrock 对话模型集成文档

npm install @langchain/aws

pnpm install @langchain/aws

yarn add @langchain/aws

bun add @langchain/aws

import { initChatModel } from "langchain";

// Follow the steps here to configure your credentials:
// https://docs.aws.amazon.com/bedrock/latest/userguide/getting-started.html

const model = await initChatModel("bedrock:gpt-5.5");

import { ChatBedrockConverse } from "@langchain/aws";

// Follow the steps here to configure your credentials:
// https://docs.aws.amazon.com/bedrock/latest/userguide/getting-started.html

const model = new ChatBedrockConverse({
  model: "gpt-5.5",
  region: "us-east-2"
});

::: 示例中的输出使用了 OpenAI。

2. 配置数据库

在本教程中，您将创建一个 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，这是一个表示数字媒体商店的示例数据库。为了方便起见，我们将数据库（Chinook.db）托管在公共 GCS 存储桶上。 :::python

import pathlib
import requests

url = "https://storage.googleapis.com/benchmarks-artifacts/chinook/Chinook.db"
local_path = pathlib.Path("Chinook.db")

if local_path.exists():
    print(f"{local_path} already exists, skipping download.")
else:
    response = requests.get(url, timeout=60)
    if response.status_code == 200:
        local_path.write_bytes(response.content)
        print(f"File downloaded and saved as {local_path}")
    else:
        print(f"Failed to download the file. Status code: {response.status_code}")

我们将使用 Python 的内置 sqlite3 模块与数据库交互：

import sqlite3

con = sqlite3.connect("Chinook.db")
cursor = con.cursor()

cursor.execute("SELECT name FROM sqlite_master WHERE type='table';")
tables = [row[0] for row in cursor.fetchall() if not row[0].startswith("sqlite_")]

print("Dialect: sqlite")
print(f"Available tables: {tables}")

cursor.execute("SELECT * FROM Artist LIMIT 5;")
print(f"Sample output: {cursor.fetchall()}")
con.close()

Dialect: sqlite
Available tables: ['Album', 'Artist', 'Customer', 'Employee', 'Genre', 'Invoice', 'InvoiceLine', 'MediaType', 'Playlist', 'PlaylistTrack', 'Track']
Sample output: [(1, 'AC/DC'), (2, 'Accept'), (3, 'Aerosmith'), (4, 'Alanis Morissette'), (5, 'Alice In Chains')]

::: :::js

import fs from "node:fs/promises";
import path from "node:path";

const url =
  "https://storage.googleapis.com/benchmarks-artifacts/chinook/Chinook.db";
const localPath = path.resolve("Chinook.db");

async function resolveDbPath() {
  const exists = await fs
    .access(localPath)
    .then(() => true)
    .catch(() => false);
  if (exists) {
    console.log(`${localPath} already exists, skipping download.`);
    return localPath;
  }
  const resp = await fetch(url);
  if (!resp.ok)
    throw new Error(`Failed to download DB. Status code: ${resp.status}`);
  const buf = Buffer.from(await resp.arrayBuffer());
  await fs.writeFile(localPath, buf);
  console.log(`File downloaded and saved as ${localPath}`);
  return localPath;
}

我们将使用 sqlite3 库与数据库交互：

import sqlite3 from "sqlite3";

const dialect = "sqlite";

async function runQuery(query: string, params: unknown[] = []): Promise<any[]> {
  const dbPath = await resolveDbPath();
  const db = new sqlite3.Database(dbPath);
  return new Promise((resolve, reject) => {
    db.all(query, params, (err, rows) => {
      db.close();
      if (err) reject(err);
      else resolve(rows);
    });
  });
}

const tableRows = await runQuery(
  "SELECT name FROM sqlite_master WHERE type='table' AND name NOT LIKE 'sqlite_%';",
);
const tableNames = tableRows.map((row) => String(row.name));
console.log(`Dialect: ${dialect}`);
console.log(`Available tables: ${tableNames.join(", ")}`);
const sampleResults = await runQuery("SELECT * FROM Artist LIMIT 5;");
console.log(`Sample output: ${JSON.stringify(sampleResults)}`);

Dialect: sqlite
Available tables: Album, Artist, Customer, Employee, Genre, Invoice, InvoiceLine, MediaType, Playlist, PlaylistTrack, Track
Sample output: [{"ArtistId":1,"Name":"AC/DC"},{"ArtistId":2,"Name":"Accept"},{"ArtistId":3,"Name":"Aerosmith"},{"ArtistId":4,"Name":"Alanis Morissette"},{"ArtistId":5,"Name":"Alice In Chains"}]

:::

3. 添加用于数据库交互的工具

以下数据库工具仅用于演示目的，是极简的包装器。它们不打算用于生产环境或保证安全性。在使用模型生成的 SQL 之前，请使用狭窄范围的数据库权限并添加应用程序特定的验证。

:::python 我们可以使用 @tool 装饰器从 langchain.tools 实现数据库工具作为薄包装器：

import sqlite3
from langchain.tools import tool

# Below are minimal tools for demonstration purposes.


@tool
def sql_db_list_tables() -> str:
    """Input is an empty string, output is a comma-separated list of tables in the database."""
    con = sqlite3.connect("Chinook.db")
    try:
        cursor = con.cursor()
        cursor.execute("SELECT name FROM sqlite_master WHERE type='table';")
        tables = [
            row[0]
            for row in cursor.fetchall()
            if not row[0].startswith("sqlite_")
        ]
        return ", ".join(tables)
    finally:
        con.close()


@tool
def sql_db_schema(table_names: str) -> str:
    """Input to this tool is a comma-separated list of tables, output is the schema and sample rows for those tables.
    Be sure that the tables actually exist by calling sql_db_list_tables first!
    Example Input: table1, table2, table3"""
    con = sqlite3.connect("Chinook.db")
    try:
        cursor = con.cursor()
        cursor.execute("SELECT name FROM sqlite_master WHERE type='table';")
        valid_tables = {
            row[0] for row in cursor.fetchall() if not row[0].startswith("sqlite_")
        }
        results = []
        for table in table_names.split(","):
            table = table.strip()
            if table not in valid_tables:
                results.append(
                    f"Error: table_names {{{table!r}}} not found in database"
                )
                continue
            cursor.execute(
                "SELECT sql FROM sqlite_master WHERE type='table' AND name=?;",
                (table,),
            )
            schema_row = cursor.fetchone()
            if schema_row:
                results.append(schema_row[0])
                try:
                    quoted_table = '"' + table.replace('"', '""') + '"'
                    cursor.execute(f"SELECT * FROM {quoted_table} LIMIT 3;")
                    rows = cursor.fetchall()
                    if rows:
                        col_names = [description[0] for description in cursor.description]
                        results.append(
                            f"/*\n3 rows from {table} table:\n"
                            + "\t".join(col_names)
                            + "\n"
                            + "\n".join(
                                "\t".join(str(x) for x in row) for row in rows
                            )
                            + "\n*/"
                        )
                except Exception as e:
                    results.append(f"Error fetching sample rows: {e}")
        return "\n\n".join(results)
    finally:
        con.close()


@tool
def sql_db_query(query: str) -> str:
    """Input to this tool is a detailed and correct SQL query, output is a result from the database.
    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.
    If you encounter an issue with Unknown column 'xxxx' in 'field list', use sql_db_schema to query the correct table fields."""
    con = sqlite3.connect("Chinook.db")
    try:
        cursor = con.cursor()
        cursor.execute(query)
        res = cursor.fetchall()
        return str(res)
    except Exception as e:
        return f"Error: {e}"
    finally:
        con.close()


tools = [sql_db_list_tables, sql_db_schema, sql_db_query]

# Use a distinct loop variable so it does not shadow the `tool` decorator,
# which is reused later to wrap the query tool for human review.
for t in tools:
    print(f"{t.name}: {t.description}\n")

sql_db_list_tables: Input is an empty string, output is a comma-separated list of tables in the database.

sql_db_schema: Input to this tool is a comma-separated list of tables, output is the schema and sample rows for those tables.
    Be sure that the tables actually exist by calling sql_db_list_tables first!
    Example Input: table1, table2, table3

sql_db_query: Input to this tool is a detailed and correct SQL query, output is a result from the database.
    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.
    If you encounter an issue with Unknown column 'xxxx' in 'field list', use sql_db_schema to query the correct table fields.

::: :::js 我们将创建自定义工具来与数据库交互：

import { tool } from "langchain";
import * as z from "zod";

async function getTableNames() {
  const rows = await runQuery(
    "SELECT name FROM sqlite_master WHERE type='table' AND name NOT LIKE 'sqlite_%';",
  );
  return rows.map((row) => String(row.name));
}

function quoteSqliteIdentifier(identifier: string) {
  return `"${identifier.replaceAll('"', '""')}"`;
}

const listTablesTool = tool(
  async () => {
    const tableNames = await getTableNames();
    return tableNames.join(", ");
  },
  {
    name: "sql_db_list_tables",
    description:
      "Input is an empty string, output is a comma-separated list of tables in the database.",
    schema: z.object({}),
  },
);

const getSchemaTool = tool(
  async ({ table_names }) => {
    const validTables = new Set(await getTableNames());
    const results: string[] = [];
    for (const table of table_names.split(",").map((t) => t.trim())) {
      if (!validTables.has(table)) {
        results.push(`Error: table_names {'${table}'} not found in database`);
        continue;
      }
      const schemaRows = await runQuery(
        "SELECT sql FROM sqlite_master WHERE type='table' AND name=?;",
        [table],
      );
      const schema = schemaRows[0]?.sql;
      if (schema) {
        results.push(String(schema));
        try {
          const rows = await runQuery(
            `SELECT * FROM ${quoteSqliteIdentifier(table)} LIMIT 3;`,
          );
          if (rows.length > 0) {
            const colNames = Object.keys(rows[0]);
            results.push(
              `/*\n3 rows from ${table} table:\n${colNames.join("\t")}\n` +
                rows
                  .map((row) =>
                    colNames.map((col) => String(row[col])).join("\t"),
                  )
                  .join("\n") +
                "\n*/",
            );
          }
        } catch (e) {
          results.push(`Error fetching sample rows: ${e}`);
        }
      }
    }
    return results.join("\n\n");
  },
  {
    name: "sql_db_schema",
    description:
      "Input to this tool is a comma-separated list of tables, output is the schema and sample rows for those tables. Be sure that the tables actually exist by calling sql_db_list_tables first! Example Input: table1, table2, table3",
    schema: z.object({
      table_names: z.string().describe("Comma-separated list of table names"),
    }),
  },
);

const queryTool = tool(
  async ({ query }) => {
    try {
      const result = await runQuery(query);
      return JSON.stringify(result);
    } catch (error) {
      return `Error: ${error instanceof Error ? error.message : String(error)}`;
    }
  },
  {
    name: "sql_db_query",
    description:
      "Input to this tool is a detailed and correct SQL query, output is a result from the database. 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.",
    schema: z.object({
      query: z.string().describe("SQL query to execute"),
    }),
  },
);

const tools = [listTablesTool, getSchemaTool, queryTool];

for (const toolItem of tools) {
  console.log(`${toolItem.name}: ${toolItem.description}\n`);
}

sql_db_list_tables: Input is an empty string, output is a comma-separated list of tables in the database.

sql_db_schema: Input to this tool is a comma-separated list of tables, output is the schema and sample rows for those tables. Be sure that the tables actually exist by calling sql_db_list_tables first! Example Input: table1, table2, table3

sql_db_query: Input to this tool is a detailed and correct SQL query, output is a result from the database. 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.

:::

4. 定义应用步骤

我们为以下步骤构建专用节点：

列出数据库表
调用 “获取架构” 工具
生成查询
检查查询

将这些步骤放入专用节点让我们能够（1）在需要时强制执行工具调用，以及（2）自定义与每个步骤关联的提示词。 :::python

from typing import Literal

from langchain.messages import AIMessage
from langchain_core.runnables import RunnableConfig
from langgraph.graph import END, START, MessagesState, StateGraph
from langgraph.prebuilt import ToolNode

get_schema_tool = next(tool for tool in tools if tool.name == "sql_db_schema")
get_schema_node = ToolNode([get_schema_tool], name="get_schema")

run_query_tool = next(tool for tool in tools if tool.name == "sql_db_query")
run_query_node = ToolNode([run_query_tool], name="run_query")


# Example: create a predetermined tool call
def list_tables(state: MessagesState):
    tool_call = {
        "name": "sql_db_list_tables",
        "args": {},
        "id": "abc123",
        "type": "tool_call",
    }
    tool_call_message = AIMessage(content="", tool_calls=[tool_call])

    list_tables_tool = next(tool for tool in tools if tool.name == "sql_db_list_tables")
    tool_message = list_tables_tool.invoke(tool_call)
    response = AIMessage(f"Available tables: {tool_message.content}")

    return {"messages": [tool_call_message, tool_message, response]}


# Example: force a model to create a tool call
def call_get_schema(state: MessagesState):
    # Note that LangChain enforces that all models accept `tool_choice="any"`
    # as well as `tool_choice=<string name of tool>`.
    llm_with_tools = model.bind_tools([get_schema_tool], tool_choice="any")
    response = llm_with_tools.invoke(state["messages"])

    return {"messages": [response]}


generate_query_system_prompt = """
You are an agent designed to interact with a SQL database.
Given an input question, create a syntactically correct {dialect} query to run,
then look at the results of the query and return the answer. Unless the user
specifies a specific number of examples they wish to obtain, always limit your
query to at most {top_k} 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.

DO NOT make any DML statements (INSERT, UPDATE, DELETE, DROP etc.) to the database.
""".format(
    dialect="sqlite",
    top_k=5,
)


def generate_query(state: MessagesState):
    system_message = {
        "role": "system",
        "content": generate_query_system_prompt,
    }
    # We do not force a tool call here, to allow the model to
    # respond naturally when it obtains the solution.
    llm_with_tools = model.bind_tools([run_query_tool])
    response = llm_with_tools.invoke([system_message] + state["messages"])

    return {"messages": [response]}


check_query_system_prompt = """
You are a SQL expert with a strong attention to detail.
Double check the {dialect} 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.
""".format(dialect="sqlite")


def check_query(state: MessagesState):
    system_message = {
        "role": "system",
        "content": check_query_system_prompt,
    }

    # Generate an artificial user message to check
    tool_call = state["messages"][-1].tool_calls[0]
    user_message = {"role": "user", "content": tool_call["args"]["query"]}
    llm_with_tools = model.bind_tools([run_query_tool], tool_choice="any")
    response = llm_with_tools.invoke([system_message, user_message])
    response.id = state["messages"][-1].id

    return {"messages": [response]}

::: :::js

import {
  AIMessage,
  HumanMessage,
  SystemMessage,
  ToolMessage,
} from "@langchain/core/messages";
import { ToolNode } from "@langchain/langgraph/prebuilt";
import {
  END,
  GraphNode,
  MessagesValue,
  START,
  StateGraph,
  StateSchema,
} from "@langchain/langgraph";

// Create tool nodes for schema and query execution
const getSchemaNode = new ToolNode([getSchemaTool]);
const runQueryNode = new ToolNode([queryTool]);

// Define state schema
const MessagesState = new StateSchema({
  messages: MessagesValue,
});

// Example: create a predetermined tool call
const listTables: GraphNode<typeof MessagesState> = async (state) => {
  const toolCall = {
    name: "sql_db_list_tables",
    args: {},
    id: "abc123",
    type: "tool_call" as const,
  };
  const toolCallMessage = new AIMessage({
    content: "",
    tool_calls: [toolCall],
  });

  const toolMessage = await listTablesTool.invoke({});
  const response = new AIMessage(`Available tables: ${toolMessage}`);

  return {
    messages: [
      toolCallMessage,
      new ToolMessage({ content: toolMessage, tool_call_id: "abc123" }),
      response,
    ],
  };
};

// Example: force a model to create a tool call
const callGetSchema: GraphNode<typeof MessagesState> = async (state) => {
  const llmWithTools = model!.bindTools([getSchemaTool], {
    tool_choice: "any",
  });
  const response = await llmWithTools.invoke(state.messages);

  return { messages: [response] };
};

const topK = 5;

const generateQuerySystemPrompt = `
You are an agent designed to interact with a SQL database.
Given an input question, create a syntactically correct ${dialect}
query to run, then look at the results of the query and return the answer. Unless
the user specifies a specific number of examples they wish to obtain, always limit
your query to at most ${topK} 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.

DO NOT make any DML statements (INSERT, UPDATE, DELETE, DROP etc.) to the database.
`;

const generateQuery: GraphNode<typeof MessagesState> = async (state) => {
  const systemMessage = new SystemMessage(generateQuerySystemPrompt);
  // We do not force a tool call here, to allow the model to
  // respond naturally when it obtains the solution.
  const llmWithTools = model!.bindTools([queryTool]);
  const response = await llmWithTools.invoke([
    systemMessage,
    ...state.messages,
  ]);

  return { messages: [response] };
};

const checkQuerySystemPrompt = `
You are a SQL expert with a strong attention to detail.
Double check the ${dialect} 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.
`;

const checkQuery: GraphNode<typeof MessagesState> = async (state) => {
  const systemMessage = new SystemMessage(checkQuerySystemPrompt);

  // Generate an artificial user message to check
  const lastMessage = state.messages[state.messages.length - 1];
  if (!AIMessage.isInstance(lastMessage) || !lastMessage.tool_calls?.length) {
    throw new Error("No tool calls found in the last message");
  }
  const toolCall = lastMessage.tool_calls[0];
  const userMessage = new HumanMessage(toolCall.args.query);
  const llmWithTools = model!.bindTools([queryTool], {
    tool_choice: "any",
  });
  const response = await llmWithTools.invoke([systemMessage, userMessage]);
  // Preserve the original message ID
  response.id = lastMessage.id;

  return { messages: [response] };
};

:::

5. 实现智能体

现在我们可以使用 Graph API 将这些步骤组装成一个工作流。我们在查询生成步骤定义了一个条件边，如果生成了查询则路由到查询检查器，如果没有工具调用则结束，这样 LLM 就能对查询做出响应。 :::python

def should_continue(state: MessagesState) -> Literal[END, "check_query"]:
    messages = state["messages"]
    last_message = messages[-1]
    if not last_message.tool_calls:
        return END
    else:
        return "check_query"


builder = StateGraph(MessagesState)
builder.add_node(list_tables)
builder.add_node(call_get_schema)
builder.add_node(get_schema_node, "get_schema")
builder.add_node(generate_query)
builder.add_node(check_query)
builder.add_node(run_query_node, "run_query")

builder.add_edge(START, "list_tables")
builder.add_edge("list_tables", "call_get_schema")
builder.add_edge("call_get_schema", "get_schema")
builder.add_edge("get_schema", "generate_query")
builder.add_conditional_edges(
    "generate_query",
    should_continue,
)
builder.add_edge("check_query", "run_query")
builder.add_edge("run_query", "generate_query")

agent = builder.compile()

我们在下方可视化应用：

import pathlib

pathlib.Path("graph.png").write_bytes(agent.get_graph().draw_mermaid_png())

::: :::js

import { ConditionalEdgeRouter } from "@langchain/langgraph";

const shouldContinue: ConditionalEdgeRouter<{ InputSchema: typeof MessagesState; Nodes: "check_query" }> = (state) => {
  const messages = state.messages;
  const lastMessage = messages[messages.length - 1];
  if (!AIMessage.isInstance(lastMessage) || !lastMessage.tool_calls?.length) {
    return END;
  } else {
    return "check_query";
  }
};

const builder = new StateGraph(MessagesState)
  .addNode("list_tables", listTables)
  .addNode("call_get_schema", callGetSchema)
  .addNode("get_schema", getSchemaNode)
  .addNode("generate_query", generateQuery)
  .addNode("check_query", checkQuery)
  .addNode("run_query", runQueryNode)
  .addEdge(START, "list_tables")
  .addEdge("list_tables", "call_get_schema")
  .addEdge("call_get_schema", "get_schema")
  .addEdge("get_schema", "generate_query")
  .addConditionalEdges("generate_query", shouldContinue)
  .addEdge("check_query", "run_query")
  .addEdge("run_query", "generate_query");

const agent = builder.compile();

我们在下方可视化应用：

import * as fs from "node:fs/promises";

const drawableGraph = await agent.getGraphAsync();
const image = await drawableGraph.drawMermaidPng();
const imageBuffer = new Uint8Array(await image.arrayBuffer());

await fs.writeFile("graph.png", imageBuffer);

:::

现在我们可以调用图： :::python

question = "Which genre on average has the longest tracks?"

stream = agent.stream_events(
    {"messages": [{"role": "user", "content": question}]},
    version="v3",
)
for message in stream.messages:
    for token in message.text:
        print(token, end="", flush=True)

final_state = stream.output

::: :::js

const question = "Which genre on average has the longest tracks?";

const stream = await agent.streamEvents(
  { messages: [{ role: "user", content: question }] },
  { version: "v3" },
);

for await (const message of stream.messages) {
  for await (const token of message.text) {
    process.stdout.write(token);
  }
}

const finalState = await stream.output;

:::

================================ Human Message =================================

Which genre on average has the longest tracks?
================================== Ai Message ==================================

Available tables: Album, Artist, Customer, Employee, Genre, Invoice, InvoiceLine, MediaType, Playlist, PlaylistTrack, Track
================================== Ai Message ==================================
Tool Calls:
  sql_db_schema (call_yzje0tj7JK3TEzDx4QnRR3lL)
 Call ID: call_yzje0tj7JK3TEzDx4QnRR3lL
  Args:
    table_names: Genre, Track
================================= Tool Message =================================
Name: sql_db_schema


CREATE TABLE "Genre" (
	"GenreId" INTEGER NOT NULL,
	"Name" NVARCHAR(120),
	PRIMARY KEY ("GenreId")
)

/*
3 rows from Genre table:
GenreId	Name
1	Rock
2	Jazz
3	Metal
*/


CREATE TABLE "Track" (
	"TrackId" INTEGER NOT NULL,
	"Name" NVARCHAR(200) NOT NULL,
	"AlbumId" INTEGER,
	"MediaTypeId" INTEGER NOT NULL,
	"GenreId" INTEGER,
	"Composer" NVARCHAR(220),
	"Milliseconds" INTEGER NOT NULL,
	"Bytes" INTEGER,
	"UnitPrice" NUMERIC(10, 2) NOT NULL,
	PRIMARY KEY ("TrackId"),
	FOREIGN KEY("MediaTypeId") REFERENCES "MediaType" ("MediaTypeId"),
	FOREIGN KEY("GenreId") REFERENCES "Genre" ("GenreId"),
	FOREIGN KEY("AlbumId") REFERENCES "Album" ("AlbumId")
)

/*
3 rows from Track table:
TrackId	Name	AlbumId	MediaTypeId	GenreId	Composer	Milliseconds	Bytes	UnitPrice
1	For Those About To Rock (We Salute You)	1	1	1	Angus Young, Malcolm Young, Brian Johnson	343719	11170334	0.99
2	Balls to the Wall	2	2	1	U. Dirkschneider, W. Hoffmann, H. Frank, P. Baltes, S. Kaufmann, G. Hoffmann	342562	5510424	0.99
3	Fast As a Shark	3	2	1	F. Baltes, S. Kaufman, U. Dirkscneider & W. Hoffman	230619	3990994	0.99
*/
================================== Ai Message ==================================
Tool Calls:
  sql_db_query (call_cb9ApLfZLSq7CWg6jd0im90b)
 Call ID: call_cb9ApLfZLSq7CWg6jd0im90b
  Args:
    query: SELECT Genre.Name, AVG(Track.Milliseconds) AS AvgMilliseconds FROM Track JOIN Genre ON Track.GenreId = Genre.GenreId GROUP BY Genre.GenreId ORDER BY AvgMilliseconds DESC LIMIT 5;
================================== Ai Message ==================================
Tool Calls:
  sql_db_query (call_DMVALfnQ4kJsuF3Yl6jxbeAU)
 Call ID: call_DMVALfnQ4kJsuF3Yl6jxbeAU
  Args:
    query: SELECT Genre.Name, AVG(Track.Milliseconds) AS AvgMilliseconds FROM Track JOIN Genre ON Track.GenreId = Genre.GenreId GROUP BY Genre.GenreId ORDER BY AvgMilliseconds DESC LIMIT 5;
================================= Tool Message =================================
Name: sql_db_query

[('Sci Fi & Fantasy', 2911783.0384615385), ('Science Fiction', 2625549.076923077), ('Drama', 2575283.78125), ('TV Shows', 2145041.0215053763), ('Comedy', 1585263.705882353)]
================================== Ai Message ==================================

The genre with the longest tracks on average is "Sci Fi & Fantasy," with an average track length of approximately 2,911,783 milliseconds. Other genres with relatively long tracks include "Science Fiction," "Drama," "TV Shows," and "Comedy."

:::python

参见 LangSmith 跟踪了解上述运行情况。

::: :::js

参见 LangSmith 跟踪了解上述运行情况。

:::

6. 实现人工审核

在执行任何可能产生意外行为或低效性的 SQL 查询之前，检查智能体的 SQL 查询是明智的。这里我们利用 LangGraph 的人工审核功能在执行 SQL 查询前暂停运行并等待人工审核。使用 LangGraph 的持久化层，我们可以无限期地暂停运行（至少只要持久化层存活）。让我们将 sql_db_query 工具封装在一个接收人类输入的节点中。我们可以使用中断函数实现这一点。下面，我们允许输入以批准工具调用、编辑其参数或提供用户反馈。 :::python

from langchain.tools import tool
from langgraph.types import interrupt
from langchain_core.runnables import RunnableConfig


@tool(
    run_query_tool.name,
    description=run_query_tool.description,
    args_schema=run_query_tool.args_schema,
)
def run_query_tool_with_interrupt(config: RunnableConfig, **tool_input):
    request = {
        "action": run_query_tool.name,
        "args": tool_input,
        "description": "Please review the tool call",
    }
    response = interrupt([request])  
    # approve the tool call
    if response["type"] == "accept":
        tool_response = run_query_tool.invoke(tool_input, config)
    # update tool call args
    elif response["type"] == "edit":
        tool_input = response["args"]["args"]
        tool_response = run_query_tool.invoke(tool_input, config)
    # respond to the LLM with user feedback
    elif response["type"] == "response":
        user_feedback = response["args"]
        tool_response = user_feedback
    else:
        raise ValueError(f"Unsupported interrupt response type: {response['type']}")

    return tool_response


# Redefine the tool node to use the interrupt version
run_query_node = ToolNode([run_query_tool_with_interrupt], name="run_query")  

::: :::js

import { RunnableConfig } from "@langchain/core/runnables";
import { interrupt } from "@langchain/langgraph";

const queryToolWithInterrupt = tool(
  async (input, config: RunnableConfig) => {
    const request = {
      action: queryTool.name,
      args: input,
      description: "Please review the tool call",
    };
    const response = interrupt([request]); 
    // approve the tool call
    if (response.type === "accept") {
      const toolResponse = await queryTool.invoke(input, config);
      return toolResponse;
    }
    // update tool call args
    else if (response.type === "edit") {
      const editedInput = response.args.args;
      const toolResponse = await queryTool.invoke(editedInput, config);
      return toolResponse;
    }
    // respond to the LLM with user feedback
    else if (response.type === "response") {
      const userFeedback = response.args;
      return userFeedback;
    } else {
      throw new Error(`Unsupported interrupt response type: ${response.type}`);
    }
  },
  {
    name: queryTool.name,
    description: queryTool.description,
    schema: queryTool.schema,
  },
);

:::

上述实现遵循了工具中断示例在更广泛的人工审核指南中的内容。参考该指南了解详情和替代方案。

现在让我们重新组装我们的图。我们将程序化检查替换为人工审核。注意我们现在包含了一个检查点；这是暂停和恢复运行所必需的。 :::python

from langgraph.checkpoint.memory import InMemorySaver

def should_continue(state: MessagesState) -> Literal[END, "run_query"]:
    messages = state["messages"]
    last_message = messages[-1]
    if not last_message.tool_calls:
        return END
    else:
        return "run_query"

builder = StateGraph(MessagesState)
builder.add_node(list_tables)
builder.add_node(call_get_schema)
builder.add_node(get_schema_node, "get_schema")
builder.add_node(generate_query)
builder.add_node(run_query_node, "run_query")

builder.add_edge(START, "list_tables")
builder.add_edge("list_tables", "call_get_schema")
builder.add_edge("call_get_schema", "get_schema")
builder.add_edge("get_schema", "generate_query")
builder.add_conditional_edges(
    "generate_query",
    should_continue,
)
builder.add_edge("run_query", "generate_query")

checkpointer = InMemorySaver()  
agent = builder.compile(checkpointer=checkpointer)  

::: :::js

import { Command, MemorySaver } from "@langchain/langgraph";

const shouldContinueWithHuman: ConditionalEdgeRouter<{ InputSchema: typeof MessagesState; Nodes: "run_query" }> = (state) => {
  const messages = state.messages;
  const lastMessage = messages[messages.length - 1];
  if (!AIMessage.isInstance(lastMessage) || !lastMessage.tool_calls?.length) {
    return END;
  } else {
    return "run_query";
  }
};

const builderWithHuman = new StateGraph(MessagesState)
  .addNode("list_tables", listTables)
  .addNode("call_get_schema", callGetSchema)
  .addNode("get_schema", getSchemaNode)
  .addNode("generate_query", generateQuery)
  .addNode("run_query", runQueryNodeWithInterrupt)
  .addEdge(START, "list_tables")
  .addEdge("list_tables", "call_get_schema")
  .addEdge("call_get_schema", "get_schema")
  .addEdge("get_schema", "generate_query")
  .addConditionalEdges("generate_query", shouldContinueWithHuman)
  .addEdge("run_query", "generate_query");

const checkpointer = new MemorySaver(); 
const agentWithHuman = builderWithHuman.compile({ checkpointer }); 

::: 我们可以像之前一样调用图。这次执行被中断： :::python

question = "Which genre on average has the longest tracks?"

stream = agent.stream_events(
    {"messages": [{"role": "user", "content": question}]},
    config,
    version="v3",
)
for message in stream.messages:
    for token in message.text:
        print(token, end="", flush=True)
if stream.interrupted:
    action = stream.interrupts[0]
    print("INTERRUPTED:")
    for request in action.value:
        print(json.dumps(request, indent=2))

::: :::js

const hitlQuestion = "Which genre on average has the longest tracks?";

const hitlStream = await agentWithHuman.streamEvents(
  { messages: [{ role: "user", content: hitlQuestion }] },
  { ...config, version: "v3" },
);

for await (const message of hitlStream.messages) {
  for await (const token of message.text) {
    process.stdout.write(token);
  }
}

// Check for interrupts
if (hitlStream.interrupted) {
  console.log("\nINTERRUPTED:");
  console.log(JSON.stringify(hitlStream.interrupts[0], null, 2));
}

:::

...

INTERRUPTED:
{
  "action": "sql_db_query",
  "args": {
    "query": "SELECT Genre.Name, AVG(Track.Milliseconds) AS AvgLength FROM Track JOIN Genre ON Track.GenreId = Genre.GenreId GROUP BY Genre.Name ORDER BY AvgLength DESC LIMIT 5;"
  },
  "description": "Please review the tool call"
}

我们可以使用 Command 接受或编辑工具调用： :::python

from langgraph.types import Command

stream = agent.stream_events(
    Command(resume={"type": "accept"}),
    # Command(resume={"type": "edit", "args": {"query": "..."}}),
    config,
    version="v3",
)
for message in stream.messages:
    for token in message.text:
        print(token, end="", flush=True)
if stream.interrupted:
    action = stream.interrupts[0]
    print("INTERRUPTED:")
    for request in action.value:
        print(json.dumps(request, indent=2))

::: :::js

const resumeStream = await agentWithHuman.streamEvents(
  new Command({ resume: { type: "accept" } }),
  // new Command({ resume: { type: "edit", args: { query: "..." } } }),
  { ...config, version: "v3" },
);

for await (const message of resumeStream.messages) {
  for await (const token of message.text) {
    process.stdout.write(token);
  }
}

:::

================================== Ai Message ==================================
Tool Calls:
  sql_db_query (call_t4yXkD6shwdTPuelXEmY3sAY)
 Call ID: call_t4yXkD6shwdTPuelXEmY3sAY
  Args:
    query: SELECT Genre.Name, AVG(Track.Milliseconds) AS AvgLength FROM Track JOIN Genre ON Track.GenreId = Genre.GenreId GROUP BY Genre.Name ORDER BY AvgLength DESC LIMIT 5;
================================= Tool Message =================================
Name: sql_db_query

[('Sci Fi & Fantasy', 2911783.0384615385), ('Science Fiction', 2625549.076923077), ('Drama', 2575283.78125), ('TV Shows', 2145041.0215053763), ('Comedy', 1585263.705882353)]
================================== Ai Message ==================================

The genre with the longest average track length is "Sci Fi & Fantasy" with an average length of about 2,911,783 milliseconds. Other genres with long average track lengths include "Science Fiction," "Drama," "TV Shows," and "Comedy."

参考人工审核指南了解详情。

下一步

查阅评估图指南，了解如何使用 LangSmith 评估 LangGraph 应用，包括像这样的 SQL 智能体。

官方原文：src/oss/langgraph/sql-agent.mdx · 非官方翻译，译文由 AI 生成，以英文原文为准。

​概念

​设置

​安装

​LangSmith

​1. 选择 LLM

​2. 配置数据库

​3. 添加用于数据库交互的工具

​4. 定义应用步骤

​5. 实现智能体

​6. 实现人工审核

​下一步

概念

设置

安装

LangSmith

1. 选择 LLM

2. 配置数据库

3. 添加用于数据库交互的工具

4. 定义应用步骤

5. 实现智能体

6. 实现人工审核

下一步