smolagents学习笔记系列（七）Examples-Self-correcting Text-to-SQL

这篇文章锁定官网教程中 Examples 章节中的 Self-correcting Text-to-SQL 文章，主要介绍了如何使用 Agent 对数据库进行查找。

• 官网链接：https://huggingface.co/docs/smolagents/v1.9.2/en/examples/text_to_sql；

【注意事项】：

1. 这个教程涉及到一些SQL的部分，如果你对这里不熟悉也不必担心，你更应该关注的是如何定义 tool 以及给 agent 怎样提供合适的 prompt；
2. 官网在这部分做的有一点不好，他们将数据准备、tool、agent定义都拆开到不同的代码段里，导致你直接运行单独部分会没有任何输出，我这里将数据准备和tool agent定义拆分成两个文件，然后利用python特性分次执行这两个文件，这样能够更好地模拟真实生产环境，即先有数据然后才会让你写agent，你不必关心SQL部分怎么实现因为在这个系列文章中你要学习的是如何实现tool和agent；

Text-to-SQL

官网在一开头就给出了一个核心问题：为什么不直接使用LLM生成的SQL语句而需要通过Agent执行，他们为此给出了以下几个理由：

1. LLM生成的SQL语句很容易出错；
2. SQL语句可能执行成功但由于存在逻辑错误，导致答案不是我们想要的，这个过程并不会触发异常；
3. Agent可以帮我们检查输出结果，并决定是否需要修改向LLM的查询prompt；

准备工作

安装依赖：

$ pip install smolagents python-dotenv sqlalchemy --upgrade -q

如果你已经在环境变量中配置好 HF_TOKEN 则可以跳过下面这行命令：

$ export HF_TOKEN="你的huggingface token"

准备一个SQL环境并添加一些数据，这里我将官方提供的两段代码合成了一段，因为这两段都是数据库的前期准备工作，实现的是创建表、插入数据、显示数据的功能，将这个文件保存为 prepare_sql.py，尽管可以直接将后面的 tool 和 agent 写在这个文件后面，但真实情况下都是已经有一个完整的 SQL ，然后才去写 tool 和 agent，所以我这里将两部分内容拆开成两个文件。

• prepare_sql.py 文件：

# 代码段一：
from dotenv import load_dotenv
load_dotenv()

from sqlalchemy import (
    create_engine,
    MetaData,
    Table,
    Column,
    String,
    Integer,
    Float,
    insert,
    inspect,
    text,
)

engine = create_engine("sqlite:///:memory:")
metadata_obj = MetaData()

def insert_rows_into_table(rows, table, engine=engine):
    for row in rows:
        stmt = insert(table).values(**row)
        with engine.begin() as connection:
            connection.execute(stmt)

table_name = "receipts"
receipts = Table(
    table_name,
    metadata_obj,
    Column("receipt_id", Integer, primary_key=True),
    Column("customer_name", String(16), primary_key=True),
    Column("price", Float),
    Column("tip", Float),
)
metadata_obj.create_all(engine)

rows = [
    {"receipt_id": 1, "customer_name": "Alan Payne", "price": 12.06, "tip": 1.20},
    {"receipt_id": 2, "customer_name": "Alex Mason", "price": 23.86, "tip": 0.24},
    {"receipt_id": 3, "customer_name": "Woodrow Wilson", "price": 53.43, "tip": 5.43},
    {"receipt_id": 4, "customer_name": "Margaret James", "price": 21.11, "tip": 1.00},
]
insert_rows_into_table(rows, receipts)

# 代码段二：
inspector = inspect(engine)
columns_info = [(col["name"], col["type"]) for col in inspector.get_columns("receipts")]

table_description = "Columns:\n" + "\n".join([f"  - {name}: {col_type}" for name, col_type in columns_info])
print(table_description)

Build our agent

接下来就是定义一个 somlagents 的 tool ，这个实现的就是对 “receipts” 表进行SQL查询，然后以字符串的形式返回。我这里同样将官网上的两段代码合并成一段，因为这两段代码分别是定义 tool 和 Agent。Agent没有使用官网上的 meta-llama/Meta-Llama-3.1-8B-Instruct ，而是不指定由 HuggingFace 自动分配：

• agent.py 文件：

from smolagents import tool
from smolagents import CodeAgent, HfApiModel

# 定义tool
@tool
def sql_engine(query: str) -> str:
    """
    Allows you to perform SQL queries on the table. Returns a string representation of the result.
    The table is named 'receipts'. Its description is as follows:
        Columns:
        - receipt_id: INTEGER
        - customer_name: VARCHAR(16)
        - price: FLOAT
        - tip: FLOAT

    Args:
        query: The query to perform. This should be correct SQL.
    """
    output = ""
    with engine.connect() as con:
        rows = con.execute(text(query))
        for row in rows:
            output += "\n" + str(row)
    return output

# 定义agent
agent = CodeAgent(
    tools=[sql_engine],
    model=HfApiModel(),
)

# 让agent执行你与LLM交互的内容
agent.run("Can you give me the name of the client who got the most expensive receipt?")

为了将两个文件依次执行以模拟真实情况，我们需要额外创建一个文件 merge.py：

• merge.py 文件：

# 先执行 prepare_sql.py 将数据准备好
with open("prepare_sql.py", encoding="utf-8") as f:
    exec(f.read())

# 再执行 agent.py 实现agent调用tool功能
with open("agent.py", encoding="utf-8") as f:
    exec(f.read())

此时你的文件结构应该如下：

(base) ~/Desktop/LLM $ tree
.
├── agent
├── merge.py
└── prepare_sql.py

运行 merge.py 文件：

$ python merge.py

Level 2: Table joins

官网除了提供上面那个SQL查询的示例以外，还提供了另一个 SQL 任务：处理多个表之间的连接。这里同样将其拆成三个文件：

• prepare_sql.py 文件：

from dotenv import load_dotenv
load_dotenv()

from sqlalchemy import (
    create_engine,
    MetaData,
    Table,
    Column,
    String,
    Integer,
    Float,
    insert,
    inspect,
    text,
)

engine = create_engine("sqlite:///:memory:")
metadata_obj = MetaData()

def insert_rows_into_table(rows, table, engine=engine):
    for row in rows:
        stmt = insert(table).values(**row)
        with engine.begin() as connection:
            connection.execute(stmt)

# 第一张 Table 
table_name = "receipts"
receipts = Table(
    table_name,
    metadata_obj,
    Column("receipt_id", Integer, primary_key=True),
    Column("customer_name", String(16), primary_key=True),
    Column("price", Float),
    Column("tip", Float),
)
metadata_obj.create_all(engine)

rows = [
    {"receipt_id": 1, "customer_name": "Alan Payne", "price": 12.06, "tip": 1.20},
    {"receipt_id": 2, "customer_name": "Alex Mason", "price": 23.86, "tip": 0.24},
    {"receipt_id": 3, "customer_name": "Woodrow Wilson", "price": 53.43, "tip": 5.43},
    {"receipt_id": 4, "customer_name": "Margaret James", "price": 21.11, "tip": 1.00},
]
insert_rows_into_table(rows, receipts)

# 第二张 Table
table_name = "waiters"
waiters = Table(
    table_name,
    metadata_obj,
    Column("receipt_id", Integer, primary_key=True),
    Column("waiter_name", String(16), primary_key=True),
)
metadata_obj.create_all(engine)

rows = [
    {"receipt_id": 1, "waiter_name": "Corey Johnson"},
    {"receipt_id": 2, "waiter_name": "Michael Watts"},
    {"receipt_id": 3, "waiter_name": "Michael Watts"},
    {"receipt_id": 4, "waiter_name": "Margaret James"},
]
insert_rows_into_table(rows, waiters)

updated_description = """Allows you to perform SQL queries on the table. Beware that this tool's output is a string representation of the execution output.
It can use the following tables:"""

inspector = inspect(engine)
for table in ["receipts", "waiters"]:
    columns_info = [(col["name"], col["type"]) for col in inspector.get_columns(table)]
    table_description = f"Table '{table}':\n"
    table_description += "Columns:\n" + "\n".join([f"  - {name}: {col_type}" for name, col_type in columns_info])
    updated_description += "\n\n" + table_description

print(updated_description)

• agent.py 文件：

from smolagents import tool
from smolagents import CodeAgent, HfApiModel


@tool
def sql_engine(query: str) -> str:
    """
    Allows you to perform SQL queries on the table. Returns a string representation of the result.
    The table is named 'receipts'. Its description is as follows:
        Columns:
        - receipt_id: INTEGER
        - customer_name: VARCHAR(16)
        - price: FLOAT
        - tip: FLOAT

    Args:
        query: The query to perform. This should be correct SQL.
    """
    output = ""
    with engine.connect() as con:
        rows = con.execute(text(query))
        for row in rows:
            output += "\n" + str(row)
    return output


sql_engine.description = updated_description

agent = CodeAgent(
    tools=[sql_engine],
    model=HfApiModel(),
)

agent.run("Which waiter got more total money from tips?")

• merge.py 文件：

# 先执行 prepare_sql.py 将数据准备好
with open("prepare_sql.py", encoding="utf-8") as f:
    exec(f.read())

# 再执行 agent.py 实现agent调用tool功能
with open("agent.py", encoding="utf-8") as f:
    exec(f.read())

然后执行 merge.py 文件：

$ python merge.py