Autogen for data analysis including numbers and text?

[analyticsinsights]

I would like to use Autogen to analyse data stored in a CSV-file. When I set this up as a GroupChat, I am able to have, for example, a (1) DataLoader, (2) Coder and (3) CodeExecutor agent which are together able to write and execute a Python code which create a chart that show basic statistics of the data in the CSV-file. However, in the CSV-file are also text data in each row and I would like Autogen to summarise the text in each row briefly, or classify the text into topics using the LLM and then store the results back into a pandas dataframe and then work with the new content of the dataframe to create graphs.

Unfortunately, I am not able to create an agent that is able to use the LLM for this task and apply it to filtered rows of the dataframe. Somehow, the coder agent mentioned above tries instead to install NLTK or Scikit-Learn and count words of the text instead of simply using the LLM to summarise the text. How can I achieve this?

This example of using RAG seems to allow a specific file with data, but I would need the data to be preprocessed in interim steps by the other agents (as described above) and only the filtered Pandas dataframe should then be used by the RetrieveUserProxyAgent.

[rysweet]

@analyticsinsights prompting for coding tasks can be tricky. unfortunately without an example of your code to look at I think its hard to advise you on next steps.

[ekzhu]

@analyticsinsights in that example the description field would be helpful.

I think you can start with a single agent and built tools.

Tools -- cheaper, fast, robust, explainable
LLM -- expensive, slow, unreliable, unexplainable

Use tools and orchestrated workflows to reduce as much stochasticity as possible.

Only go multi agent when you have exhausted the capability of single agent.

[analyticsinsights]

Unfortunately, this is not too helpful.

I know what tools are, and I have used a single agent at the beginning which did not get me further. It is not apparent from the description of tool use which elements of my described workflow I need to embed in custom tools.

I found this more elaborate example code and will take a closer look at what functions are defined and how. I suppose that I need to implement (1) loading of the CSV-file, (2) filtering of the CSV-file and (3) calling LLM for summarisation as functions. But I am concerned that the final tool becomes very static if I am pre-defining all these elements. Then the tool would lack flexibility, which is precisely one of the key advantages that agentic workflows promised to provide.

What is also needed with Autogen is a list of standard tools like CrewAI and LangChain have it - and to document them and their application well.

[ekzhu]

Looks like you have a good sense of where you want to go. My impression was that you had a fixed workflow so my suggestion was to apply determinism and no need to worry about generalization. Apologies.

If you want your agent to generalize to a wide range of tasks, code executor or functions that writes code are helpful. But they come with the additional cost and risk.

You can use langchain tools through a wrapper: https://microsoft.github.io/autogen/dev/reference/python/autogen_ext.tools.langchain.html

[ekzhu]

Our work on general capability is MagenticOne. It will be available in the 0.4 version.

Are you prototyping or looking for something to work in production scenarios?

[analyticsinsights]

This is for prototyping, I am trying to implement various use-cases. Analysing data stored in CSV-files is clearly one of them.

[analyticsinsights]

I just wanted to share with the community that answering questions on the content of a CSV-file is actually very straightforward --out of the box-- with LangChain tools.

import pandas as pd
from langchain_experimental.agents.agent_toolkits import create_pandas_dataframe_agent
from langchain_openai import OpenAI

df = pd.read_csv("https://raw.githubusercontent.com/pandas-dev/pandas/main/doc/data/titanic.csv")
agent = create_pandas_dataframe_agent(OpenAI(temperature=0), df, verbose=True, allow_dangerous_code=True)
agent.invoke("What is the percentage of passengers that were in first class?")

The AG2 release to integrate tools from third-party providers shows how important tools are for agentic workflows. It would be great to have this also in Autogen.

[ekzhu]

How about this?

import asyncio
import pandas as pd
from langchain_experimental.tools.python.tool import PythonAstREPLTool
from autogen_ext.tools.langchain import LangChainToolAdapter
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.ui import Console

async def main() -> None:
    df = pd.read_csv("https://raw.githubusercontent.com/pandas-dev/pandas/main/doc/data/titanic.csv")
    tool = LangChainToolAdapter(PythonAstREPLTool(locals={"df": df}))
    model_client = OpenAIChatCompletionClient(model="gpt-4o")
    agent = AssistantAgent("assistant", tools=[tool], model_client=model_client, system_message="Use the `df` variable to access the dataset.")
    await Console(agent.run_stream(task="What's the average age of the passengers?"))

asyncio.run(main())

We already support Langchain tool integration through LangchainToolWrapper: https://microsoft.github.io/autogen/dev/reference/python/autogen_ext.tools.langchain.html

[analyticsinsights]

Many thanks @ekzhu! This works.

Is it possible to return the dataframe when my task is "Filter the dataframe for first class passengers"?

[ekzhu]

Unless you want to convert the returned dataframe in text message, I would use a separate variable to store the result or saving the dataframe on disk.

import asyncio
import pandas as pd
from langchain_experimental.tools.python.tool import PythonAstREPLTool
from autogen_ext.tools.langchain import LangChainToolAdapter
from autogen_ext.models.openai import OpenAIChatCompletionClient
from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.ui import Console

async def main() -> None:
    df = pd.read_csv("https://raw.githubusercontent.com/pandas-dev/pandas/main/doc/data/titanic.csv")
    repl = PythonAstREPLTool(locals={"df": df})
    tool = LangChainToolAdapter(repl)
    model_client = OpenAIChatCompletionClient(model="gpt-4o")
    agent = AssistantAgent("assistant", tools=[tool], model_client=model_client, system_message="Use the `df` variable to access the dataset.")
    await Console(agent.run_stream(task="Filter the dataframe for first class passengers and save it to a variable named `first_class`."))
    print(repl.locals.get("first_class"))

asyncio.run(main())

[analyticsinsights]

Many thanks @ekzhu! With this code one can really start analysing pandas dataframes and get the results using Autogen.