app.py

import pandas as pd
import gradio as gr
from transformers import pipeline
from langchain_community.llms import HuggingFacePipeline
from langchain_experimental.agents import create_pandas_dataframe_agent
from langchain.agents.agent_types import AgentType

# Load data
df = pd.read_csv("synthetic_profit.csv")

# Lightweight Hugging Face pipeline (Flan-T5-base)
hf_pipeline = pipeline(
    task="text2text-generation",
    model="google/flan-t5-base",
    device=-1  # CPU
)

# LangChain LLM
llm = HuggingFacePipeline(pipeline=hf_pipeline)

# Create LangChain agent with explicit parsing-error handling
agent = create_pandas_dataframe_agent(
    llm, 
    df, 
    verbose=True, 
    agent_type=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
    handle_parsing_errors=True,
    allow_dangerous_code=True
)

# Answer query function with clearer prompts
def answer(query: str) -> str:
    try:
        prompt = (
            f"Answer this clearly and numerically without scientific notation: {query}. "
            "If multiple numbers, provide their total sum clearly."
        )
        response = agent.run(prompt)
        return f"📊 {response}"
    except Exception as e:
        return f"❗ Error: {str(e)}"

# Gradio interface
demo = gr.Interface(
    fn=answer,
    inputs=gr.Textbox(
        lines=2, 
        placeholder="E.g., 'Total revenue for Product B in EMEA during Q2 2024'"
    ),
    outputs="text",
    title="🟢 SAP Profitability Data Chat (Flan-T5 + Pandas)",
    description="Ask clearly numeric questions about synthetic SAP profitability data. Results are precise and human-readable."
)

demo.launch()