Upload app.py with huggingface_hub

app.py

@@ -0,0 +1,243 @@
+import os
+import random
+import torch
+import pandas as pd
+from functools import lru_cache
+
+from flask import Flask, render_template, request, jsonify
+from sentence_transformers import SentenceTransformer
+
+from fastapi import FastAPI
+from fastapi.middleware.wsgi import WSGIMiddleware
+import uvicorn
+import nest_asyncio
+
+# ===============================
+# OPTIONAL FAISS
+# ===============================
+try:
+    import faiss
+    FAISS_AVAILABLE = True
+    print("✅ FAISS available")
+except ImportError:
+    FAISS_AVAILABLE = False
+    print("⚠️ FAISS not available → torch fallback")
+
+# ===============================
+# CONFIG
+# ===============================
+BASE_DIR = os.path.abspath(os.path.dirname(__file__))
+CSV_DATA = "dataset_2026.csv"
+EMB_FILE = "embeddings_questions.pt"
+TOP_K_RECOMMANDATIONS = 5
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+
+# ===============================
+# APP
+# ===============================
+app = Flask(
+    __name__,
+    template_folder=os.path.join(BASE_DIR, "templates"),
+    static_folder=os.path.join(BASE_DIR, "static")
+)
+
+# ===============================
+# LOAD MODEL (ONCE)
+# ===============================
+print("🔹 Loading model...")
+model = SentenceTransformer(
+    "OrdalieTech/Solon-embeddings-mini-beta-1.1",
+    device=DEVICE,
+    trust_remote_code=True
+)
+
+# ===============================
+# LOAD & CLEAN DATASET
+# ===============================
+print("🔹 Loading dataset...")
+df = pd.read_csv(CSV_DATA, low_memory=False)
+
+df = df.dropna(subset=["question"]).reset_index(drop=True)
+
+if len(df) == 0:
+    raise RuntimeError("❌ Dataset has no valid questions")
+
+questions = df["question"].astype(str).tolist()
+NB_QUESTIONS = len(questions)
+
+print(f"✅ Valid questions: {NB_QUESTIONS}")
+
+# ===============================
+# LOAD / CREATE EMBEDDINGS
+# ===============================
+if os.path.exists(EMB_FILE):
+    emb_base = torch.load(EMB_FILE, map_location=DEVICE)
+else:
+    emb_base = model.encode(
+        questions,
+        convert_to_tensor=True,
+        normalize_embeddings=True,
+        batch_size=64
+    )
+    torch.save(emb_base, EMB_FILE)
+
+if emb_base.shape[0] != NB_QUESTIONS:
+    raise RuntimeError("❌ Embedding count mismatch")
+
+# ===============================
+# INDEX SETUP
+# ===============================
+K_SEARCH = max(1, min(TOP_K_RECOMMANDATIONS + 1, NB_QUESTIONS))
+
+if FAISS_AVAILABLE:
+    emb_np = emb_base.cpu().numpy()
+    dim = emb_np.shape[1]
+    index = faiss.IndexFlatIP(dim)
+    index.add(emb_np)
+else:
+    emb_base_cpu = emb_base.cpu()
+
+# ===============================
+# CACHE QUESTION EMBEDDING
+# ===============================
+@lru_cache(maxsize=1000)
+def encode_question_cached(q: str):
+    return model.encode(
+        q,
+        convert_to_tensor=True,
+        normalize_embeddings=True
+    )
+
+# ===============================
+# UTILS
+# ===============================
+def enrich_message(base):
+    return random.choice([
+        f"Bonne question 🙂 {base}",
+        f"Voici ce que je peux vous dire : {base}",
+        f"Intéressant ! {base}",
+        base
+    ])
+
+# ===============================
+# CORE LOGIC (BULLETPROOF)
+# ===============================
+def process_question(question: str):
+
+    if not question or not question.strip():
+        return {
+            "response": "Veuillez poser une question.",
+            "confidence": 0,
+            "matched": "—",
+            "intent": "Vide",
+            "recs": []
+        }
+
+    emb_q = encode_question_cached(question)
+
+    # ---------- FAISS ----------
+    if FAISS_AVAILABLE:
+        D, I = index.search(
+            emb_q.cpu().numpy().reshape(1, -1),
+            K_SEARCH
+        )
+
+        if D.size == 0:
+            return {
+                "response": "Aucune réponse trouvée",
+                "confidence": 0,
+                "matched": "—",
+                "intent": "Inconnu",
+                "recs": []
+            }
+
+        idxs = I[0].tolist()
+        scores = D[0].tolist()
+
+    # ---------- TORCH FALLBACK ----------
+    else:
+        # ensure shape [1, dim]
+        if emb_q.dim() == 1:
+            emb_q = emb_q.unsqueeze(0)
+
+        scores_all = torch.matmul(
+            emb_q,
+            emb_base_cpu.T
+        ).squeeze(0)
+
+        nb_scores = scores_all.numel()
+        k = min(K_SEARCH, nb_scores)
+
+        if k == 0:
+            return {
+                "response": "Aucune réponse trouvée",
+                "confidence": 0,
+                "matched": "—",
+                "intent": "Inconnu",
+                "recs": []
+            }
+
+        values, indices = torch.topk(scores_all, k)
+        idxs = indices.tolist()
+        scores = values.tolist()
+
+    # ---------- DECISION ----------
+    best_idx = idxs[0]
+    score = int(scores[0] * 100)
+
+    if score < 40:
+        return {
+            "response": "Aucune réponse trouvée",
+            "confidence": score,
+            "matched": "—",
+            "intent": "Inconnu",
+            "recs": []
+        }
+
+    if score < 80:
+        recs = [
+            df["question"].iloc[i]
+            for i in idxs[1:]
+            if i < NB_QUESTIONS
+        ][:TOP_K_RECOMMANDATIONS]
+
+        return {
+            "response": "Je ne suis pas totalement sûr.",
+            "confidence": score,
+            "matched": df["question"].iloc[best_idx],
+            "intent": "Incertain",
+            "recs": recs
+        }
+
+    return {
+        "response": enrich_message(df["rationale"].iloc[best_idx]),
+        "confidence": score,
+        "matched": df["question"].iloc[best_idx],
+        "intent": df["intent"].iloc[best_idx],
+        "recs": []
+    }
+
+# ===============================
+# ROUTES
+# ===============================
+@app.route("/")
+def index():
+    return render_template("index.html")
+
+@app.route("/ask", methods=["POST"])
+def ask():
+    return jsonify(process_question(request.json.get("question", "")))
+
+# ===============================
+# FASTAPI WRAPPER
+# ===============================
+fastapi_app = FastAPI()
+fastapi_app.mount("/", WSGIMiddleware(app))
+
+# ===============================
+# MAIN
+# ===============================
+if __name__ == "__main__":
+    nest_asyncio.apply()
+    print("🚀 Server running on http://localhost:7860")
+    uvicorn.run(fastapi_app, host="0.0.0.0", port=7860)