import os
import pandas as pd
from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
from langchain_openai import ChatOpenAI
from langchain_experimental.agents import create_pandas_dataframe_agent

app = FastAPI(title="BiteSight Analytics API")

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"], # In production, replace with your Vercel URL
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

# Load data on startup
DATA_FILE = "data_client.csv"
try:
    df = pd.read_csv(DATA_FILE)
    print(f"Successfully loaded {DATA_FILE} with {len(df)} rows.")
except Exception as e:
    print(f"Warning: Could not load {DATA_FILE}: {e}")
    df = pd.DataFrame() # Fallback empty dataframe

class ChatRequest(BaseModel):
    question: str

@app.post("/chat")
async def chat(request: ChatRequest):
    try:
        if df.empty:
            return {"answer": "Error: Data source is unavailable on the server."}

        akashml_api_key = os.environ.get("AKASHML_API_KEY")
        if not akashml_api_key:
            return {"answer": "Error: AKASHML_API_KEY environment variable tidak diset di Hugging Face Space Anda."}

        # Initialize AkashML LLM via ChatOpenAI (OpenAI-compatible)
        llm = ChatOpenAI(
            openai_api_key=akashml_api_key,
            openai_api_base="https://api.akashml.com/v1",
            model_name="MiniMaxAI/MiniMax-M2.5",
            temperature=0.2,
            max_tokens=8192
        )
        
        system_message = (
            "You are a world-class F&B Executive Business Consultant and Data Scientist "
            "with 20+ years of experience advising top restaurant chains and retail F&B brands. "
            "You have deep expertise in sales analytics, consumer behavior, operational efficiency, and growth strategy.\n\n"

            "RESPONSE STYLE:\n"
            "- If the user sends a general greeting, respond warmly and naturally in Indonesian.\n"
            "- If the user asks a business/data question, give a THOROUGH, INSIGHTFUL, and EXECUTIVE-LEVEL analysis. "
            "Do NOT give shallow or overly brief answers. Go deep. Uncover layers of insight.\n"
            "- ALWAYS include: key data facts WITH specific numbers, analytical commentary explaining the 'why', "
            "patterns/anomalies/trends, business implications, AND concrete actionable recommendations.\n"
            "- Structure your response FREELY using rich Markdown: use headers (##, ###), bold text, "
            "bullet lists, and comparison tables when helpful. DO NOT use any rigid fixed template.\n"
            "- ALWAYS respond in fluent, professional INDONESIAN.\n\n"

            "PANDAS DATA ENGINE RULES (CRITICAL):\n"
            "- A pandas DataFrame named `df` with ALL rows of client transaction data is ALREADY LOADED.\n"
            "- ALWAYS run actual pandas operations to get real numbers. NEVER estimate or fabricate data.\n"
            "- DO NOT recreate the dataframe using StringIO or hardcoded text. Use `df` directly.\n"
            "- Run MULTIPLE pandas queries if needed to build a complete picture "
            "(e.g., value_counts, groupby, mean, corr, describe, resample, etc.).\n"
            "- Cross-reference multiple dimensions when relevant (e.g., payment method vs. branch vs. time of day).\n\n"

            "DATA PRESENTATION RULES:\n"
            "- ALWAYS state specific numbers, percentages, and timeframes. Vague statements are not acceptable.\n"
            "- Currency is US Dollars ($). ALWAYS format as $XX.XX. NEVER use 'Rp' or 'Rupiah'.\n"
            "- Do NOT explain the dataset schema or mention row counts unless explicitly asked.\n"
            "- When comparing metrics, provide both absolute numbers AND percentages for clarity.\n"
        )

        agent = create_pandas_dataframe_agent(
            llm,
            df,
            agent_type="tool-calling",
            allow_dangerous_code=True,
            handle_parsing_errors=True,
            verbose=True,
            prefix=system_message
        )

        response = agent.invoke({"input": request.question})
        return {"answer": response.get("output", "No output generated.")}


    except Exception as e:
        error_msg = str(e)
        if "Could not parse" in error_msg:
            extracted = error_msg.split("Could not parse")[-1].replace("LLM output:", "").strip(" `'\"")
            return {"answer": extracted}
            
        return {"answer": f"Terjadi kendala saat memproses instruksi. Jika Anda sekadar menyapa, halo juga! Jika ini pertanyaan data, format Model mungkin tidak sesuai.\n\n**Log Server:** `{error_msg}`"}

if __name__ == "__main__":
    import uvicorn
    uvicorn.run(app, host="0.0.0.0", port=7860) # Hugging Face Spaces typical port