概述

在本教程中,您将学习如何使用 LangChain 智能体构建可以回答有关 SQL 数据库的问题的智能体。 在高层次上,智能体将:
1

从数据库获取可用的表和模式

2

决定哪些表与问题相关

3

获取相关表的模式

4

基于问题和模式中的信息生成查询

5

使用 LLM 仔细检查查询中的常见错误

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

基于结果制定响应

构建 SQL 数据库的问答系统需要执行模型生成的 SQL 查询。这样做存在固有风险。确保您的数据库连接权限始终尽可能狭窄地限定为智能体的需求。这将减轻(但不会消除)构建模型驱动系统的风险。

概念

我们将涵盖以下概念:

设置

安装

npm i langchain @langchain/core typeorm sqlite3 zod

LangSmith

设置 LangSmith 以检查链或智能体内部发生的情况。然后设置以下环境变量:
export LANGSMITH_TRACING="true"
export LANGSMITH_API_KEY="..."

1. 选择 LLM

选择支持工具调用的模型:
  • OpenAI
  • Anthropic
  • Azure
  • Google Gemini
  • Bedrock Converse
👉 Read the OpenAI chat model integration docs
npm install @langchain/openai
import { initChatModel } from "langchain";

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

const model = await initChatModel("gpt-4.1");
下面示例中显示的输出使用了 OpenAI。

2. 配置数据库

您将为本教程创建一个 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。 为方便起见,我们已将数据库(Chinook.db)托管在公共 GCS 存储桶上。
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() {
  if (await fs.exists(localPath)) {
    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);
  return localPath;
}

3. Add tools for database interactions

Use the SqlDatabase wrapper available in the langchain/sql_db to interact with the database. The wrapper provides a simple interface to execute SQL queries and fetch results:
import { SqlDatabase } from "@langchain/classic/sql_db";
import { DataSource } from "typeorm";

let db: SqlDatabase | undefined;
async function getDb() {
  if (!db) {
    const dbPath = await resolveDbFile();
    const datasource = new DataSource({ type: "sqlite", database: dbPath });
    db = await SqlDatabase.fromDataSourceParams({ appDataSource: datasource });
  }
  return db;
}

async function getSchema() {
  const db = await getDb();
  return await db.getTableInfo();
}

6. Implement human-in-the-loop review

It can be prudent to check the agent’s SQL queries before they are executed for any unintended actions or inefficiencies. LangChain agents feature support for built-in human-in-the-loop middleware to add oversight to agent tool calls. Let’s configure the agent to pause for human review on calling the sql_db_query tool:
from langchain.agents import create_agent
from langchain.agents.middleware import HumanInTheLoopMiddleware 
from langgraph.checkpoint.memory import InMemorySaver 


agent = create_agent(
    model,
    tools,
    system_prompt=system_prompt,
    middleware=[ 
        HumanInTheLoopMiddleware( 
            interrupt_on={"sql_db_query": True}, 
            description_prefix="Tool execution pending approval", 
        ), 
    ], 
    checkpointer=InMemorySaver(), 
)
We’ve added a checkpointer to our agent to allow execution to be paused and resumed. See the human-in-the-loop guide for detalis on this as well as available middleware configurations.
On running the agent, it will now pause for review before executing the sql_db_query tool:
question = "Which genre on average has the longest tracks?"
config = {"configurable": {"thread_id": "1"}} 

for step in agent.stream(
    {"messages": [{"role": "user", "content": question}]},
    config, 
    stream_mode="values",
):
    if "messages" in step:
        step["messages"][-1].pretty_print()
    elif "__interrupt__" in step: 
        print("INTERRUPTED:") 
        interrupt = step["__interrupt__"][0] 
        for request in interrupt.value["action_requests"]: 
            print(request["description"]) 
    else:
        pass
...

INTERRUPTED:
Tool execution pending approval

Tool: sql_db_query
Args: {'query': 'SELECT g.Name AS Genre, AVG(t.Milliseconds) AS AvgTrackLength FROM Track t JOIN Genre g ON t.GenreId = g.GenreId GROUP BY g.Name ORDER BY AvgTrackLength DESC LIMIT 1;'}
We can resume execution, in this case accepting the query, using Command:
from langgraph.types import Command 

for step in agent.stream(
    Command(resume={"decisions": [{"type": "approve"}]}), 
    config,
    stream_mode="values",
):
    if "messages" in step:
        step["messages"][-1].pretty_print()
    elif "__interrupt__" in step:
        print("INTERRUPTED:")
        interrupt = step["__interrupt__"][0]
        for request in interrupt.value["action_requests"]:
            print(request["description"])
    else:
        pass
================================== Ai Message ==================================
Tool Calls:
  sql_db_query (call_7oz86Epg7lYRqi9rQHbZPS1U)
 Call ID: call_7oz86Epg7lYRqi9rQHbZPS1U
  Args:
    query: SELECT Genre.Name, AVG(Track.Milliseconds) AS AvgDuration FROM Track JOIN Genre ON Track.GenreId = Genre.GenreId GROUP BY Genre.Name ORDER BY AvgDuration 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 duration of about 2,911,783 milliseconds, followed by "Science Fiction" and "Drama."
Refer to the human-in-the-loop guide for details.

4. Execute SQL queries

Before running the command, do a check to check the LLM generated command in _safe_sql:

const DENY_RE = /\b(INSERT|UPDATE|DELETE|ALTER|DROP|CREATE|REPLACE|TRUNCATE)\b/i;
const HAS_LIMIT_TAIL_RE = /\blimit\b\s+\d+(\s*,\s*\d+)?\s*;?\s*$/i;

function sanitizeSqlQuery(q) {
  let query = String(q ?? "").trim();

  // block multiple statements (allow one optional trailing ;)
  const semis = [...query].filter((c) => c === ";").length;
  if (semis > 1 || (query.endsWith(";") && query.slice(0, -1).includes(";"))) {
    throw new Error("multiple statements are not allowed.")
  }
  query = query.replace(/;+\s*$/g, "").trim();

  // read-only gate
  if (!query.toLowerCase().startsWith("select")) {
    throw new Error("Only SELECT statements are allowed")
  }
  if (DENY_RE.test(query)) {
    throw new Error("DML/DDL detected. Only read-only queries are permitted.")
  }

  // append LIMIT only if not already present
  if (!HAS_LIMIT_TAIL_RE.test(query)) {
    query += " LIMIT 5";
  }
  return query;
}

Then, use run from SQLDatabase to execute commands with an execute_sql tool:
import { tool } from "langchain"
import * as z from "zod";

const executeSql = tool(
  async ({ query }) => {
    const q = sanitizeSqlQuery(query);
    try {
      const result = await db.run(q);
      return typeof result === "string" ? result : JSON.stringify(result, null, 2);
    } catch (e) {
      throw new Error(e?.message ?? String(e))
    }
  },
  {
    name: "execute_sql",
    description: "Execute a READ-ONLY SQLite SELECT query and return results.",
    schema: z.object({
      query: z.string().describe("SQLite SELECT query to execute (read-only)."),
    }),
  }
);

5. Use createAgent

Use createAgent to build a ReAct agent with minimal code. The agent will interpret the request and generate a SQL command. The tools will check the command for safety and then try to execute the command. If the command has an error, the error message is returned to the model. The model can then examine the original request and the new error message and generate a new command. This can continue until the LLM generates the command successfully or reaches an end count. This pattern of providing a model with feedback - error messages in this case - is very powerful. Initialize the agent with a descriptive system prompt to customize its behavior:
import { SystemMessage } from "langchain";

const getSystemPrompt = async () => new SystemMessage(`You are a careful SQLite analyst.

Authoritative schema (do not invent columns/tables):
${await getSchema()}

Rules:
- Think step-by-step.
- When you need data, call the tool \`execute_sql\` with ONE SELECT query.
- Read-only only; no INSERT/UPDATE/DELETE/ALTER/DROP/CREATE/REPLACE/TRUNCATE.
- Limit to 5 rows unless user explicitly asks otherwise.
- If the tool returns 'Error:', revise the SQL and try again.
- Limit the number of attempts to 5.
- If you are not successful after 5 attempts, return a note to the user.
- Prefer explicit column lists; avoid SELECT *.
`);
Now, create an agent with the model, tools, and prompt:
import { createAgent } from "langchain";

const agent = createAgent({
  model: "gpt-5",
  tools: [executeSql],
  systemPrompt: getSystemPrompt,
});

6. Run the agent

Run the agent on a sample query and observe its behavior:
const question = "Which genre, on average, has the longest tracks?";
const stream = await agent.stream(
  { messages: [{ role: "user", content: question }] },
  { streamMode: "values" }
);
for await (const step of stream) {
  const message = step.messages.at(-1);
  console.log(`${message.role}: ${JSON.stringify(message.content, null, 2)}`);
}
human: Which genre, on average, has the longest tracks?
ai:
tool: [{"Genre":"Sci Fi & Fantasy","AvgMilliseconds":2911783.0384615385}]
ai: Sci Fi & Fantasy — average track length ≈ 48.5 minutes (about 2,911,783 ms).
The agent correctly wrote a query, checked the query, and ran it to inform its final response.
You can inspect all aspects of the above run, including steps taken, tools invoked, what prompts were seen by the LLM, and more in the LangSmith trace.

(Optional) Use Studio

Studio provides a “client side” loop as well as memory so you can run this as a chat interface and query the database. You can ask questions like “Tell me the scheme of the database” or “Show me the invoices for the 5 top customers”. You will see the SQL command that is generated and the resulting output. The details of how to get that started are below.
In addition to the previously mentioned packages, you will need to:
npm i -g langgraph-cli@latest
In directory you will run in, you will need a langgraph.json file with the following contents:
{
  "dependencies": ["."],
  "graphs": {
      "agent": "./sqlAgent.ts:agent",
      "graph": "./sqlAgentLanggraph.ts:graph"
  },
  "env": ".env"
}
import fs from "node:fs/promises";
import path from "node:path";
import { SqlDatabase } from "@langchain/classic/sql_db";
import { DataSource } from "typeorm";
import { SystemMessage, createAgent, tool } from "langchain"
import * as z from "zod";

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

async function resolveDbPath() {
  if (await fs.exists(localPath)) {
    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);
  return localPath;
}

let db: SqlDatabase | undefined;
async function getDb() {
  if (!db) {
    const dbPath = await resolveDbPath();
    const datasource = new DataSource({ type: "sqlite", database: dbPath });
    db = await SqlDatabase.fromDataSourceParams({ appDataSource: datasource });
  }
  return db;
}

async function getSchema() {
  const db = await getDb();
  return await db.getTableInfo();
}

const DENY_RE = /\b(INSERT|UPDATE|DELETE|ALTER|DROP|CREATE|REPLACE|TRUNCATE)\b/i;
const HAS_LIMIT_TAIL_RE = /\blimit\b\s+\d+(\s*,\s*\d+)?\s*;?\s*$/i;

function sanitizeSqlQuery(q) {
  let query = String(q ?? "").trim();

  // block multiple statements (allow one optional trailing ;)
  const semis = [...query].filter((c) => c === ";").length;
  if (semis > 1 || (query.endsWith(";") && query.slice(0, -1).includes(";"))) {
    throw new Error("multiple statements are not allowed.")
  }
  query = query.replace(/;+\s*$/g, "").trim();

  // read-only gate
  if (!query.toLowerCase().startsWith("select")) {
    throw new Error("Only SELECT statements are allowed")
  }
  if (DENY_RE.test(query)) {
    throw new Error("DML/DDL detected. Only read-only queries are permitted.")
  }

  // append LIMIT only if not already present
  if (!HAS_LIMIT_TAIL_RE.test(query)) {
    query += " LIMIT 5";
  }
  return query;
}

const executeSql = tool(
  async ({ query }) => {
    const q = sanitizeSqlQuery(query);
    try {
      const result = await db.run(q);
      return typeof result === "string" ? result : JSON.stringify(result, null, 2);
    } catch (e) {
      throw new Error(e?.message ?? String(e))
    }
  },
  {
    name: "execute_sql",
    description: "Execute a READ-ONLY SQLite SELECT query and return results.",
    schema: z.object({
      query: z.string().describe("SQLite SELECT query to execute (read-only)."),
    }),
  }
);

const getSystemPrompt = async () => new SystemMessage(`You are a careful SQLite analyst.

Authoritative schema (do not invent columns/tables):
${await getSchema()}

Rules:
- Think step-by-step.
- When you need data, call the tool \`execute_sql\` with ONE SELECT query.
- Read-only only; no INSERT/UPDATE/DELETE/ALTER/DROP/CREATE/REPLACE/TRUNCATE.
- Limit to 5 rows unless user explicitly asks otherwise.
- If the tool returns 'Error:', revise the SQL and try again.
- Limit the number of attempts to 5.
- If you are not successful after 5 attempts, return a note to the user.
- Prefer explicit column lists; avoid SELECT *.
`);

export const agent = createAgent({
  model: "gpt-5",
  tools: [executeSql],
  systemPrompt: getSystemPrompt,
});

后续步骤

要进行更深入的定制,请查看本教程,了解如何直接使用 LangGraph 原语实现 SQL 智能体。
Connect these docs programmatically to Claude, VSCode, and more via MCP for real-time answers.