Fatal process exit when calling df_pai.chat() with SQL agent (no error, no output)

[KANISHKSHUKLA]

System Info

OS version: Windows 11
Python version: 3.10.11
pandasai version: 3.x (latest beta, installed with --pre)
pandasai-litellm version: 0.0.1
DuckDB version: 1.4.0.dev1184
LLM backend: Gemini via LiteLLM

🐛 Describe the bug

When calling df_pai.chat("...") with a simple SQL question, the Python process exits immediately with no error, traceback, or output. This happens even with a tiny CSV and a simple question. Direct DuckDB queries work fine. This appears to be a fatal bug in the pandas-ai 3.x SQL agent or code execution sandbox.

Steps to reproduce:

1. Use this minimal CSV (Volumes.csv):
   Port,State,Commodity,Entity,Type,Period,Value
   PortA,Fresh,Apple,EntityX,Group,2024-25,12.5
   PortA,Fresh,Banana,EntityY,Tied,2024-25,8.2
   PortB,Dry,Orange,EntityZ,Tied,2024-25,15.0
   PortB,Fresh,Apple,EntityX,Group,2023-24,7.1
   PortC,Dry,Grape,EntityW,Tied,2024-25,5.0
   PortA,Fresh,Apple,EntityY,Tied,2023-24,3.3
   PortC,Fresh,Banana,EntityZ,Group,2023-24,6.7
   PortB,Dry,Grape,EntityW,Tied,2023-24,2.2

2. Use this minimal script:
   import pandas as pd
   import pandasai as pai
   from pandasai_litellm import LiteLLM
   from pandasai.dataframe import DataFrame as PandasAIDataFrame
   import os
   from dotenv import load_dotenv

   load_dotenv()
   df = pd.read_csv("Volumes.csv")
   llm = LiteLLM(model="gemini/gemini-2.5-flash", api_key=os.environ.get("GEMINI_API_KEY"))
   pai.config.set({"llm": llm})
   df_pai = PandasAIDataFrame(df)
   print("Before LLM call")
   print(df_pai.chat("What is the total value for Mundra in 2024-25?"))
   print("After LLM call")

3. Run the script:
   python test.py

Expected behavior:

• The script should print "Before LLM call", then the answer, then "After LLM call".

Actual behavior:

• The script prints "Before LLM call" and then the process exits immediately.
• No error, no traceback, no output.
• No crash when running the equivalent SQL directly with DuckDB.

What I tried:

• Used a tiny CSV (8 rows).
• Ran direct DuckDB queries (they work fine).
• Tried in both Streamlit and plain Python script (same result).
• Upgraded all packages to latest versions.
• Tried with both gemini/gemini-2.5-flash and gemini/gemini-2.5-pro.

Additional info:

• This looks like a fatal bug in the pandas-ai 3.x SQL agent or its code execution sandbox.
• Please advise on a workaround or fix.

[KGomez-git]

I also had this issue, but it was fixed once I downgraded DuckDB to version 1.3.0. It might be that more recent versions are incompatible. I wasn't able to run direct DuckDB queries though, so it might not be the same problem, but it's worth trying.

[KANISHKSHUKLA]

I also had this issue, but it was fixed once I downgraded DuckDB to version 1.3.0. It might be that more recent versions are incompatible. I wasn't able to run direct DuckDB queries though, so it might not be the same problem, but it's worth trying.

Can you please tell which python version were you using? Can you please tell?

[KGomez-git]

I was using python 3.11.9

Here is my requirements.txt:

• pandas
• pandasai>=3.0.0b2
• pandasai-openai
• pandasai-litellm
• numpy<2
• pyyaml
• duckdb==1.3.0

[KANISHKSHUKLA]

I was using python 3.11.9

Here is my requirements.txt:

• pandas
• pandasai>=3.0.0b2
• pandasai-openai
• pandasai-litellm
• numpy<2
• pyyaml
• duckdb==1.3.0

Thank You for the suggestion, It really helped me, everything is fixed now, but i think there should be error handling in the pandasai for such errors, as it leaves us just guessing around.

[imanrofiqsy]

Sorry, I encountered a similar issue when using pandasai on a Django web server. Currently, I am running the server using the “VS Code Debugger” and everything is working fine as expected. However, when I run the code using a cmd script, only the message “Write Query” appears, and then the process stops without any errors or tracebacks.

my requirements
pandasai 3.0.0b19
python 3.11
pandasai-litellm version: 0.0.1
LLM backend: gpt-4.1-mini
Postgre

here is the code

llm = LiteLLM(model="gpt-4.1-mini", api_key=MyAPI)

pai.config.set({
    "llm": llm
})
@csrf_protect
def generate_pandasai_response(request):
    if request.method == 'POST':
        try:
            # Ambil data dari request
            data = json.loads(request.body)
            query = data.get('query', '')
            type = data.get('report_type', 'energy_usage')
            df = DataLogGeneral.objects.all().values()
            if type == "billing":
                df = BillingData.objects.all().values()
            try:
                print("Write Query")
                response = pai.DataFrame(df).chat(query)
                print("get response")
            except Exception as e:
                print("Error in PandasAI:", e)

            return JsonResponse({'response': str(response)})

        except Exception as e:
            print("Error:", e)
            return JsonResponse({'error': str(e)}, status=500)

    return JsonResponse({'error': 'Invalid request method'}, status=405)

here is the launch.json

{
    // Use IntelliSense to learn about possible attributes.
    // Hover to view descriptions of existing attributes.
    // For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
    "version": "0.2.0",
    "configurations": [

        {
            "name": "Python Debugger: Django",
            "type": "debugpy",
            "request": "launch",
            "args": [
                "runserver"
            ],
            "django": true,
            "autoStartBrowser": true,
            "program": "${workspaceFolder}\\enosys\\manage.py",
            "justMyCode": true
        }
    ]
}

how I run in cmd script

1. Activate the env
2. run the manage.py run server

Thanks.