Build a SQL agent

Overview

In this tutorial, you will learn how to build an agent that can answer questions about a SQL database using LangChain agents. At a high level, the agent will:

Fetch the available tables and schemas from the database

Decide which tables are relevant to the question

Fetch the schemas for the relevant tables

Generate a query based on the question and information from the schemas

Double-check the query for common mistakes using an LLM

Execute the query and return the results

Correct mistakes surfaced by the database engine until the query is successful

Formulate a response based on the results

Building Q&A systems of SQL databases requires executing model-generated SQL queries. There are inherent risks in doing this. Make sure that your database connection permissions are always scoped as narrowly as possible for your agent’s needs. This will mitigate, though not eliminate, the risks of building a model-driven system.

Concepts

We will cover the following concepts:

Tools for reading from SQL databases
LangChain agents
Human-in-the-loop processes

Setup

Installation

npm i langchain @langchain/core typeorm sqlite3 zod

LangSmith

Set up LangSmith to inspect what is happening inside your chain or agent. Then set the following environment variables:

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

1. Select an LLM

Select a model that supports tool-calling:

OpenAI
Anthropic
Google Gemini
Fireworks AI
MistralAI
VertexAI

npm i @langchain/openai

OPENAI_API_KEY=your-api-key

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

const llm = new ChatOpenAI({
  model: "gpt-4o-mini",
});

The output shown in the examples below used OpenAI.

2. Configure the database

You will be creating a SQLite database for this tutorial. SQLite is a lightweight database that is easy to set up and use. We will be loading the chinook database, which is a sample database that represents a digital media store. For convenience, we have hosted the database (Chinook.db) on a public GCS bucket.

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/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(
    llm,
    tools,
    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: 
            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=[{"type": "accept"}]), 
    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:
            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 { 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: "openai:gpt-5",
  tools: [executeSql],
  prompt: 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.

Run your agent in Studio

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/sql_db";
import { DataSource } from "typeorm";
import { SystemMessage, createAgent, tool } from "langchain"
import { 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: "openai:gpt-5",
  tools: [executeSql],
  prompt: getSystemPrompt,
});

Next steps

For deeper customization, check out this tutorial for implementing a SQL agent directly using LangGraph primitives.

Edit the source of this page on GitHub

Tutorials

Conceptual overviews

LangChain Academy

Additional resources

Overview

Concepts

Setup

Installation

LangSmith

1. Select an LLM

2. Configure the database

3. Add tools for database interactions

6. Implement human-in-the-loop review

4. Execute SQL queries

5. Use `createAgent`

6. Run the agent

(Optional) Use Studio

Next steps

Tutorials

Conceptual overviews

LangChain Academy

Additional resources

​Overview

​Concepts

​Setup

​Installation

​LangSmith

​1. Select an LLM

​2. Configure the database

​3. Add tools for database interactions

​6. Implement human-in-the-loop review

​4. Execute SQL queries

​5. Use createAgent

​6. Run the agent

​(Optional) Use Studio

​Next steps

Overview

Concepts

Setup

Installation

LangSmith

1. Select an LLM

2. Configure the database

3. Add tools for database interactions

6. Implement human-in-the-loop review

4. Execute SQL queries

5. Use `createAgent`

6. Run the agent

(Optional) Use Studio

Next steps