app.py

import os
import random
import torch
import pandas as pd
from flask import Flask, render_template, request, jsonify
from sentence_transformers import SentenceTransformer, util
import gradio as gr
import uvicorn
import nest_asyncio
from fastapi import FastAPI
from fastapi.middleware.wsgi import WSGIMiddleware

# ==============================
# CONFIG
# ==============================
BASE_DIR = os.path.abspath(os.path.dirname(__file__))

CSV_DATA = "dataset_2026.csv"
EMB_FILE = "embeddings_questions.pt"
TOP_K_RECOMMANDATIONS = 5

# ==============================
# FLASK APP
# ==============================
app = Flask(
    __name__,
    template_folder=os.path.join(BASE_DIR, "templates"),
    static_folder=os.path.join(BASE_DIR, "static")
)

# ==============================
# MODEL
# ==============================
print("🔄 Chargement du modèle...")

try:
    model = SentenceTransformer(
        "OrdalieTech/Solon-embeddings-mini-beta-1.1",
        device="cpu",
        trust_remote_code=True
    )
    print("✓ Modèle principal chargé")
except Exception as e:
    print("⚠️ Modèle principal échoué:", e)
    model = SentenceTransformer(
        "paraphrase-multilingual-MiniLM-L12-v2",
        device="cpu"
    )
    print("✓ Modèle fallback chargé")

# ==============================
# GLOBAL CACHE (IMPORTANT FIX)
# ==============================
df = None
embeddings = None


# ==============================
# DATA LOADING (ROBUST FIX)
# ==============================
def load_data():
    global df

    try:
        # auto-detect separator (FIX IMPORTANT)
        df = pd.read_csv(CSV_DATA, sep=None, engine="python")

        # normalize column names (VERY IMPORTANT FIX)
        df.columns = df.columns.str.strip()

        print(f"✓ Données chargées: {len(df)} lignes")
        print("📌 Colonnes:", df.columns.tolist())

        return df

    except FileNotFoundError:
        print("❌ Dataset introuvable → création...")

        df = pd.DataFrame({
            "Question": ["Bonjour", "Comment ça va?", "Qu'est-ce que c'est?"],
            "Response": [
                "Bonjour! Comment puis-je vous aider?",
                "Je vais bien, merci!",
                "C'est une application Q/A"
            ],
            "Intent": ["salutation", "conversation", "information"]
        })

        df.to_csv(CSV_DATA, index=False)
        return df


# ==============================
# EMBEDDINGS (CACHE FIX)
# ==============================
def load_embeddings():
    global embeddings, df

    if embeddings is not None:
        return embeddings

    if os.path.exists(EMB_FILE):
        print("📂 Chargement embeddings...")
        embeddings = torch.load(EMB_FILE, map_location="cpu")
        return embeddings

    print("🔨 Création embeddings...")

    questions = df["Question"].astype(str).tolist()

    embeddings = model.encode(
        questions,
        convert_to_tensor=True,
        normalize_embeddings=True,
        show_progress_bar=True
    )

    torch.save(embeddings, EMB_FILE)
    print("✓ Embeddings sauvegardés")

    return embeddings


# ==============================
# UTILS
# ==============================
def enrich_message(text):
    prefixes = [
        "Bonne question 🙂",
        "Voici la réponse :",
        "Intéressant !",
        "D'après mes données :",
        "Réponse :",
        "🤖"
    ]
    return f"{random.choice(prefixes)} {text}"


def get_column(df, name):
    """

    SAFE column getter (fixes Intent/intent/spacing issues)

    """
    for col in df.columns:
        if col.lower() == name.lower():
            return col
    raise KeyError(f"Column '{name}' not found. Available: {df.columns.tolist()}")


# ==============================
# CORE LOGIC (FIXED)
# ==============================
def process_question(question):
    global df, embeddings

    if not question or not question.strip():
        return {
            "response": "Veuillez poser une question valide.",
            "confidence": 0,
            "matched": "—",
            "intent": "Invalid",
            "recs": []
        }

    try:
        df = load_data()
        embeddings = load_embeddings()

        q_col = get_column(df, "Question")
        r_col = get_column(df, "Response")
        i_col = get_column(df, "Intent")

        emb_q = model.encode(
            question,
            convert_to_tensor=True,
            normalize_embeddings=True
        )

        scores = util.pytorch_cos_sim(emb_q, embeddings)[0]

        best_idx = torch.argmax(scores).item()
        confidence = int(scores[best_idx].item() * 100)

        # LOW CONFIDENCE
        if confidence < 40:
            return {
                "response": "Désolé, je n'ai pas trouvé de réponse.",
                "confidence": confidence,
                "matched": "—",
                "intent": "Not found",
                "recs": []
            }

        # MEDIUM CONFIDENCE (suggestions)
        if confidence < 80:
            k = min(TOP_K_RECOMMANDATIONS + 1, len(scores))
            top_indices = torch.topk(scores, k).indices.tolist()

            recs = [
                df[q_col].iloc[i]
                for i in top_indices
                if i != best_idx
            ][:TOP_K_RECOMMANDATIONS]

            return {
                "response": "Je ne suis pas sûr. Voulez-vous dire :",
                "confidence": confidence,
                "matched": df[q_col].iloc[best_idx],
                "intent": df[i_col].iloc[best_idx],
                "recs": recs
            }

        # HIGH CONFIDENCE (final answer)
        answer = df[r_col].iloc[best_idx]
        intent = df[i_col].iloc[best_idx]

        return {
            "response": enrich_message(answer),
            "confidence": confidence,
            "matched": df[q_col].iloc[best_idx],
            "intent": intent,
            "recs": []
        }

    except Exception as e:
        print("❌ Erreur:", e)
        return {
            "response": "Erreur technique.",
            "confidence": 0,
            "matched": "—",
            "intent": "Error",
            "recs": []
        }


# ==============================
# FLASK ROUTES
# ==============================
@app.route("/")
def index():
    return render_template("index.html")


@app.route("/ask", methods=["POST"])
def ask():
    try:
        data = request.get_json()
        question = data.get("question", "")
        return jsonify(process_question(question))
    except Exception as e:
        print(e)
        return jsonify({"response": "Erreur serveur"})


# ==============================
# GRADIO
# ==============================
def gradio_chat(message, history):
    return process_question(message)["response"]

iface = gr.ChatInterface(
    fn=gradio_chat,
    title="AskLaQ Assistant",
    description="Posez vos questions"
)


# ==============================
# FASTAPI WRAPPER
# ==============================
fastapi_app = FastAPI(title="AskLaQ API")

fastapi_app.mount("/", WSGIMiddleware(app))
fastapi_app = gr.mount_gradio_app(fastapi_app, iface, path="/chat")


# ==============================
# MAIN
# ==============================
if __name__ == "__main__":
    nest_asyncio.apply()

    print("=" * 60)
    print("🚀 ASKLAQ SYSTEM (ROBUST VERSION)")
    print("=" * 60)

    uvicorn.run(
        fastapi_app,
        host="0.0.0.0",
        port=7860,
        log_level="info"
    )