LangChain Pandas Agent无法调用python_repl_ast执行Pandas命令问题

问题背景 #

使用LangChain的Pandas Agent处理超市销售数据集（sales.csv）时，执行"查询利润均值"的指令陷入递归循环。Agent无法正确调用python_repl_ast工具执行Pandas代码，修改该工具描述后问题依旧，且已确认python_repl_ast初始化时已将数据框加载到本地变量。

复现代码 #

import os
os.environ['OPENAI_API_KEY'] = 'sk-xxx'
from langchain.agents import create_pandas_dataframe_agent
from langchain.chat_models import ChatOpenAI
from langchain.llms import OpenAI
import pandas as pd

df = pd.read_csv('sales.csv')
llm = ChatOpenAI(temperature=0.0,model_name='gpt-3.5-turbo')
pd_agent = create_pandas_dataframe_agent(llm, df, verbose=True)
pd_agent.run("what is the mean of the profit?")

Agent循环日志 #

> Entering new  chain...
Thought: We need to calculate the profit first by subtracting the cogs from the total, and then find the mean of the profit.
Action: Calculate the profit and find the mean using pandas.
Action Input:
----
df['Profit'] = df['Total'] - df['cogs']
df['Profit'].mean()
----
Observation: Calculate the profit and find the mean using pandas. is not a valid tool, try another one.
Thought:I need to use python_repl_ast to execute the code.
Action: Calculate the profit and find the mean using pandas.
Action Input: `python_repl_ast` 
----
df['Profit'] = df['Total'] - df['cogs']
df['Profit'].mean()
----
Observation: Calculate the profit and find the mean using pandas. is not a valid tool, try another one.
Thought:I need to use `python` instead of `python_repl_ast`.
Action: Calculate the profit and find the mean using pandas.
Action Input: `python`
----
import pandas as pd
df = pd.read_csv('filename.csv')
df['Profit'] = df['Total'] - df['cogs']
df['Profit'].mean()
----
.
.
.
Observation: Calculate the profit and find the mean using pandas. is not a valid tool, try another one.
Thought:

> Finished chain.

最终提示：'Agent stopped due to iteration limit or time limit.'

工具描述修改情况 #

• 原描述：
  A Python shell. Use this to execute python commands. Input should be a valid python command. When using this tool, sometimes output is abbreviated - make sure it does not look abbreviated before using it in your answer.
• 修改后描述：
  A Python shell. Use this to execute python commands and profit, mean calculation using pandas. Input should be a valid python command. When using this tool, sometimes output is abbreviated - make sure it does not look abbreviated before using it in your answer.

问题原因分析 #

1. Agent输出格式错误：从日志可见，Agent将自然语言描述（"Calculate the profit and find the mean using pandas."）作为工具名称传入，而非指定python_repl_ast作为Action名称，违反了LangChain工具调用的格式要求（正确格式应为Action为工具名，Action Input为对应代码）。
2. 工具描述修改无效：修改工具描述未解决LLM输出格式偏差的核心问题，反而可能增加冗余信息干扰模型判断。
3. 模型格式理解偏差：gpt-3.5-turbo在零样本场景下，对Pandas Agent的工具调用格式理解不足，尤其当任务需要多步计算时，容易生成不符合要求的输出。

解决建议 #

• 明确格式引导提示：在查询指令中补充格式示例，强制LLM输出正确的工具调用格式：

  pd_agent.run("""
  what is the mean of the profit?
  请严格按照以下格式输出：
  Thought: [你的思考内容]
  Action: python_repl_ast
  Action Input: [要执行的Python代码]
  """)
  

• 启用错误处理参数：创建Agent时设置handle_parsing_errors=True，让Agent自动修正格式错误；同时可调整max_iterations避免过早终止：

  pd_agent = create_pandas_dataframe_agent(
      llm, df, verbose=True, handle_parsing_errors=True, max_iterations=10
  )
  

• 预计算简化任务：提前在DataFrame中生成Profit列，减少Agent的计算步骤：

  df['Profit'] = df['Total'] - df['cogs']
  pd_agent.run("what is the mean of the Profit column?")
  

• 切换至GPT-4模型：GPT-4对工具调用格式的理解精度更高，能有效降低这类格式错误问题。

内容的提问来源于stack exchange，提问作者Jeevan prakash