sg

app.py

@@ -1,7 +1,7 @@
 # app.py
 # Flask RAG app (HF Spaces / Static) — dataset sudah ada di Space.
 import os, json, re, time, logging
-from functools import lru_cache
+from functools import lru_cache, wraps
 from typing import Dict, List, Tuple
 from dataclasses import dataclass
 from datetime import datetime
@@ -252,13 +252,10 @@ def build_prompt(user_query: str, sentences: List[str]) -> str:
         "- DILARANG menulis frasa meta seperti 'berdasarkan', 'menurut', 'merujuk', atau 'bersumber'."
     )
     return f"""{system}
-
 KALIMAT SUMBER:
 {block}
-
 PERTANYAAN:
 {user_query}
-
 JAWAB (1 kalimat saja):
 """
 
@@ -316,14 +313,15 @@ def db():
     return SessionLocal()
 
 def login_required(view_func):
+    @wraps(view_func)
     def wrapper(*args, **kwargs):
         if not session.get("logged_in"):
             return redirect(url_for("auth_login"))
         return view_func(*args, **kwargs)
-    wrapper.__name__ = view_func.__name__
     return wrapper
 
 def admin_required(view_func):
+    @wraps(view_func)
     def wrapper(*args, **kwargs):
         if not session.get("logged_in"):
             return redirect(url_for("auth_login"))
@@ -331,7 +329,6 @@ def admin_required(view_func):
             flash("Hanya admin yang boleh mengakses halaman itu.", "error")
             return redirect(url_for("subjects"))
         return view_func(*args, **kwargs)
-    wrapper.__name__ = view_func.__name__
     return wrapper
 
 
@@ -508,17 +505,39 @@ def ask(subject_key: str):
     prompt = build_prompt(query, sentences)
 
     try:
-        answer = generate(
+        raw_answer = generate(
             LLM, prompt,
             max_tokens=64, temperature=0.2, top_p=1.0,
             stop=["\n\n", "\n###", "###", "\nUser:",
                   "Berdasarkan", "berdasarkan", "Menurut", "menurut",
                   "Merujuk", "merujuk", "Mengacu", "mengacu", "Bersumber", "bersumber"]
-        ).strip()
+        )
+        if raw_answer is None:
+            raw_answer = ""
+        raw_answer = raw_answer.strip()
+
+        # Log raw model output for debugging
+        log.info(f"[LLM] Raw answer repr: {repr(raw_answer)}")
+
+        # Hati-hati: banyak model menyisipkan <think>...</think>
+        # Hapus block <think>...</think> dengan aman (case-insensitive)
+        cleaned = re.sub(r"<think\b[^>]*>.*?</think>", "", raw_answer, flags=re.DOTALL | re.IGNORECASE).strip()
+        # Jika masih ada sisa tag tunggal seperti </think> atau <think>, hapus saja tag-tag tersebut
+        cleaned = re.sub(r"</?think\b[^>]*>", "", cleaned, flags=re.IGNORECASE).strip()
+
+        # Jika pembersihan menghapus semuanya, coba fallback: keluarkan teks yang bukan tag dari raw_answer
+        if not cleaned:
+            non_tag = re.sub(r"<[^>]+>", "", raw_answer).strip()
+            if non_tag:
+                cleaned = non_tag
+
+        answer = cleaned
+
     except Exception as e:
         log.exception(f"[LLM] generate error: {e}")
         return jsonify({"ok": True, "answer": FALLBACK_TEXT})
 
+    # Ambil 1 kalimat pertama (jika model mengeluarkan beberapa kalimat)
     m = re.search(r"(.+?[.!?])(\s|$)", answer)
     answer = (m.group(1) if m else answer).strip()
     answer = strip_meta_sentence(answer)