ad

app.py

@@ -1,14 +1,23 @@
-# app.py
-# Flask RAG app (HF Spaces / Static) — dataset sudah ada di Space.
+# app.py (HF Spaces CPU-Optimized)
+# RAG sekolah super hemat CPU:
+# - Default model: 3B instruct (GGUF) + ctx 1024
+# - Retrieval cepat: FAISS top-12 → pilih kalimat pakai lexical overlap (tanpa encode per-kalimat)
+# - Encoder dipakai HANYA untuk query & FAISS (1x per request)
+# - Jawaban final lewat <final>...</final>, stop di </final>, retry kalau kosong/ellipsis
+# - Admin + Auth Postgres tetap sama
+
 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
 from zoneinfo import ZoneInfo
 from pathlib import Path
 
-from flask import Flask, render_template, request, redirect, url_for, session, jsonify, flash
+from flask import (
+    Flask, render_template, request, redirect, url_for, session, jsonify, flash
+)
+
 import numpy as np
 import faiss
 import torch
@@ -20,70 +29,72 @@ load_dotenv()
 # ========= ENV & LOGGING =========
 os.environ.setdefault("KMP_DUPLICATE_LIB_OK", "TRUE")
 os.environ.setdefault("OMP_NUM_THREADS", "1")
-torch.set_num_threads(1)
-torch.set_num_interop_threads(1)
+try:
+    torch.set_num_threads(int(os.environ.get("NUM_THREADS", "3")))  # 3 thread cukup di CPU Spaces
+    torch.set_num_interop_threads(1)
+except Exception:
+    pass
+
 logging.basicConfig(level=logging.INFO, format="%(asctime)s | %(levelname)s | %(message)s")
 log = logging.getLogger("rag-app")
 
-# ========= IMPORT EKSTERNAL =========
-from Guardrail import validate_input           # -> bool (lazy di file)
+# ========= IMPORT EKSTERNAL (wrapper & guardrail) =========
+from Guardrail import validate_input           # -> bool
 from Model import load_model, generate         # -> llama.cpp wrapper
 
-# ========= PATH ROOT PROYEK =========
+# ========= PATH ROOT =========
 BASE_DIR = Path(__file__).resolve().parent
 
-# ========= KONFIGURASI RAG =========
-MODEL_PATH   = str(BASE_DIR / "models" / os.getenv("GGUF_FILENAME", "DeepSeek-R1-Distill-Qwen-7B-Q4_K_M.gguf"))
-CTX_WINDOW   = 4096
-N_GPU_LAYERS = 0  # HF Spaces CPU only
-N_THREADS    = int(os.environ.get("NUM_THREADS", "4"))
+# ========= KONFIG MODEL & RAG (di-tune untuk CPU) =========
+GGUF_DEFAULT = "DeepSeek-R1-Distill-Qwen-7B-Q4_K_M.gguf"  # kecil & cepat; upload ke /models
+MODEL_PATH   = str(BASE_DIR / "models" / os.getenv("GGUF_FILENAME", GGUF_DEFAULT))
+CTX_WINDOW   = int(os.environ.get("CTX_WINDOW", 1024))
+N_GPU_LAYERS = int(os.environ.get("N_GPU_LAYERS", 0))
+N_THREADS    = int(os.environ.get("NUM_THREADS", 3))
 
-ENCODER_NAME   = "intfloat/multilingual-e5-large"
+ENCODER_NAME   = os.environ.get("ENCODER_NAME", "intfloat/multilingual-e5-large")
 ENCODER_DEVICE = torch.device("cpu")
 
-# Dataset sudah ada di Space → path RELATIF
+# Dataset sudah ada di Space → path RELATIF (samakan dengan struktur kamu)
 SUBJECTS: Dict[str, Dict[str, str]] = {
     "ipas": {
-        "index":      str(BASE_DIR / "Rag-Pipeline" / "Vektor Database" / "Ipas" / "IPA_index.index"),
-        "chunks":     str(BASE_DIR / "Dataset"      / "Ipas"            / "Chunk" / "ipas_chunks.json"),
-        "embeddings": str(BASE_DIR / "Dataset"      / "Ipas"            / "Embedd"/ "ipas_embeddings.npy"),
-        "label":      "IPAS",
-        "desc":       "Ilmu Pengetahuan Alam dan Sosial"
+        "index": str(BASE_DIR / "Rag-Pipeline" / "Vektor Database" / "Ipas" / "IPA_index.index"),
+        "chunks": str(BASE_DIR / "Dataset" / "Ipas" / "Chunk" / "ipas_chunks.json"),
+        "embeddings": str(BASE_DIR / "Dataset" / "Ipas" / "Embedd"/ "ipas_embeddings.npy"),
+        "label": "IPAS",
+        "desc": "Ilmu Pengetahuan Alam dan Sosial"
     },
     "penjas": {
-        "index":      str(BASE_DIR / "Rag-Pipeline" / "Vektor Database" / "Penjas" / "PENJAS_index.index"),
-        "chunks":     str(BASE_DIR / "Dataset"      / "Penjas"          / "Chunk"  / "penjas_chunks.json"),
-        "embeddings": str(BASE_DIR / "Dataset"      / "Penjas"          / "Embedd" / "penjas_embeddings.npy"),
-        "label":      "PJOK",
-        "desc":       "Pendidikan Jasmani, Olahraga, dan Kesehatan"
+        "index": str(BASE_DIR / "Rag-Pipeline" / "Vektor Database" / "Penjas" / "PENJAS_index.index"),
+        "chunks": str(BASE_DIR / "Dataset" / "Penjas" / "Chunk" / "penjas_chunks.json"),
+        "embeddings": str(BASE_DIR / "Dataset" / "Penjas" / "Embedd" / "penjas_embeddings.npy"),
+        "label": "PJOK",
+        "desc": "Pendidikan Jasmani, Olahraga, dan Kesehatan"
     },
     "pancasila": {
-        "index":      str(BASE_DIR / "Rag-Pipeline" / "Vektor Database" / "Pancasila" / "PANCASILA_index.index"),
-        "chunks":     str(BASE_DIR / "Dataset"      / "Pancasila"       / "Chunk"     / "pancasila_chunks.json"),
-        "embeddings": str(BASE_DIR / "Dataset"      / "Pancasila"       / "Embedd"    / "pancasila_embeddings.npy"),
-        "label":      "PANCASILA",
-        "desc":       "Pendidikan Pancasila dan Kewarganegaraan"
+        "index": str(BASE_DIR / "Rag-Pipeline" / "Vektor Database" / "Pancasila" / "PANCASILA_index.index"),
+        "chunks": str(BASE_DIR / "Dataset" / "Pancasila" / "Chunk" / "pancasila_chunks.json"),
+        "embeddings": str(BASE_DIR / "Dataset" / "Pancasila" / "Embedd" / "pancasila_embeddings.npy"),
+        "label": "PANCASILA",
+        "desc": "Pendidikan Pancasila dan Kewarganegaraan"
     }
 }
 
-# Threshold dan fallback
-TOP_K_FAISS = 24
-TOP_K_FINAL = 10
-MIN_COSINE  = 0.84
-MIN_HYBRID  = 0.15
-
-FALLBACK_TEXT        = "maap pengetahuan tidak ada dalam database"
-GUARDRAIL_BLOCK_TEXT = "maap, pertanyaan ditolak oleh guardrail"
-ENABLE_PROFILING     = False
+# Threshold & parameter cepat
+TOP_K_FAISS   = int(os.environ.get("TOP_K_FAISS", 12))
+TOP_K_FINAL   = int(os.environ.get("TOP_K_FINAL", 6))
+MIN_COSINE    = float(os.environ.get("MIN_COSINE", 0.80))  # lebih longgar biar jarang fallback
+MIN_LEXICAL   = float(os.environ.get("MIN_LEXICAL", 0.10))
+FALLBACK_TEXT = os.environ.get("FALLBACK_TEXT", "maap pengetahuan tidak ada dalam database")
+GUARDRAIL_BLOCK_TEXT = os.environ.get("GUARDRAIL_BLOCK_TEXT", "maap, pertanyaan ditolak oleh guardrail")
+ENABLE_PROFILING = os.environ.get("ENABLE_PROFILING", "false").lower() == "true"
 
 # ========= APP =========
 app = Flask(__name__)
 app.secret_key = os.environ.get("FLASK_SECRET_KEY", "dev-secret-please-change")
 
 from werkzeug.middleware.proxy_fix import ProxyFix
-
 app.wsgi_app = ProxyFix(app.wsgi_app, x_for=1, x_proto=1, x_host=1)
-# supaya session tersimpan di browser saat lewat proxy/HTTPS (HF Spaces)
 app.config.update(
     SESSION_COOKIE_NAME="session",
     SESSION_COOKIE_SAMESITE="None",
@@ -93,11 +104,10 @@ app.config.update(
     PREFERRED_URL_SCHEME="https",
 )
 
-
-# ========= GLOBAL MODEL =========
+# ========= GLOBALS =========
 ENCODER_TOKENIZER = None
-ENCODER_MODEL = None
-LLM = None
+ENCODER_MODEL     = None
+LLM               = None
 
 @dataclass(frozen=True)
 class SubjectAssets:
@@ -105,18 +115,25 @@ class SubjectAssets:
     texts: List[str]
     embs: np.ndarray
 
-# ========= TEKS UTILITAS =========
+# ========= TEKS UTIL =========
 STOPWORDS_ID = {
     "yang","dan","atau","pada","di","ke","dari","itu","ini","adalah","dengan",
     "untuk","serta","sebagai","oleh","dalam","akan","kamu","apa","karena",
     "agar","sehingga","terhadap","dapat","juga","para","diri",
 }
 TOKEN_RE = re.compile(r"[A-Za-zÀ-ÖØ-öø-ÿ]+", re.UNICODE)
+
+@lru_cache(maxsize=4096)
+def _tok_cached(word: str) -> str:
+    return word.lower()
+
 def tok_id(text: str) -> List[str]:
-    return [t.lower() for t in TOKEN_RE.findall(text or "") if t.lower() not in STOPWORDS_ID]
+    return [tw for w in TOKEN_RE.findall(text or "") if (tw:=_tok_cached(w)) not in STOPWORDS_ID]
+
 def lexical_overlap(query: str, sent: str) -> float:
     q = set(tok_id(query)); s = set(tok_id(sent))
-    if not q or not s: return 0.0
+    if not q or not s:
+        return 0.0
     return len(q & s) / max(1, len(q | s))
 
 QUESTION_LIKE_RE = re.compile(r"(^\s*(apa|mengapa|bagaimana|sebutkan|jelaskan)\b|[?]$)", re.IGNORECASE)
@@ -133,9 +150,10 @@ META_PREFIX_RE = re.compile(r"^\s*(?:" + r"|".join(META_PREFIX_PATTERNS) + r")\s
 
 def clean_prefix(t: str) -> str:
     t = (t or "").strip()
-    for _ in range(5):
+    for _ in range(3):
         t2 = META_PREFIX_RE.sub("", t).lstrip()
-        if t2 == t: break
+        if t2 == t:
+            break
         t = t2
     return t
 
@@ -143,23 +161,31 @@ def strip_meta_sentence(s: str) -> str:
     s = clean_prefix(s or "")
     if re.match(r"^\s*(berdasarkan|menurut|merujuk|mengacu|bersumber|dari)\b", s, re.IGNORECASE):
         s = re.sub(r"^\s*[^,.;!?]*[,.;!?]\s*", "", s) or s
-        s = clean_prefix(s)
+    s = clean_prefix(s)
     return s.strip()
 
 SENT_SPLIT_RE = re.compile(r"(?<=[.!?])\s+")
-def split_sentences(text: str) -> List[str]:
+
+def split_sentences_fast(text: str) -> List[str]:
+    # tanpa encoding per-kalimat
     outs = []
     for p in SENT_SPLIT_RE.split(text or ""):
         s = clean_prefix((p or "").strip())
-        if not s: continue
-        if s[-1] not in ".!?": s += "."
-        if QUESTION_LIKE_RE.search(s):  continue
-        if INSTRUCTION_RE.search(s):    continue
-        if len(s.strip()) < 10:         continue
+        if not s:
+            continue
+        if s[-1] not in ".!?":
+            s += "."
+        if QUESTION_LIKE_RE.search(s):
+            continue
+        if INSTRUCTION_RE.search(s):
+            continue
+        if len(s) < 12:
+            continue
         outs.append(s)
     return outs
 
-# ========= MODEL WARMUP (LAZY) =========
+# ========= MODEL WARMUP =========
+
 def warmup_models():
     global ENCODER_TOKENIZER, ENCODER_MODEL, LLM
     if ENCODER_TOKENIZER is None or ENCODER_MODEL is None:
@@ -167,31 +193,35 @@ def warmup_models():
         ENCODER_TOKENIZER = AutoTokenizer.from_pretrained(ENCODER_NAME)
         ENCODER_MODEL = AutoModel.from_pretrained(ENCODER_NAME).to(ENCODER_DEVICE).eval()
     if LLM is None:
-        log.info(f"[INIT] Load LLM: {MODEL_PATH}")
+        log.info(f"[INIT] Load LLM: {MODEL_PATH} | ctx={CTX_WINDOW} | threads={N_THREADS}")
         LLM = load_model(MODEL_PATH, n_ctx=CTX_WINDOW, n_gpu_layers=N_GPU_LAYERS, n_threads=N_THREADS)
 
-# ========= LOAD ASSETS PER-MAPEL =========
+# ========= ASSETS =========
+
 @lru_cache(maxsize=8)
-def load_subject_assets(subject_key: str) -> SubjectAssets:
+def load_subject_assets(subject_key: str) -> "SubjectAssets":
     if subject_key not in SUBJECTS:
         raise ValueError(f"Unknown subject: {subject_key}")
     cfg = SUBJECTS[subject_key]
     log.info(f"[ASSETS] Loading subject={subject_key} | index={cfg['index']}")
-    if not os.path.exists(cfg["index"]): raise FileNotFoundError(cfg["index"])
-    if not os.path.exists(cfg["chunks"]): raise FileNotFoundError(cfg["chunks"])
-    if not os.path.exists(cfg["embeddings"]): raise FileNotFoundError(cfg["embeddings"])
-
+    if not os.path.exists(cfg["index"]):
+        raise FileNotFoundError(cfg["index"])
+    if not os.path.exists(cfg["chunks"]):
+        raise FileNotFoundError(cfg["chunks"])
+    if not os.path.exists(cfg["embeddings"]):
+        raise FileNotFoundError(cfg["embeddings"])
     index = faiss.read_index(cfg["index"])
     with open(cfg["chunks"], "r", encoding="utf-8") as f:
-        texts = [it["text"] for it in json.load(f)]
-    embs = np.load(cfg["embeddings"])
+        texts = [it.get("text", "") for it in json.load(f)]
+    embs = np.load(cfg["embeddings"])  # (N, dim)
     if index.ntotal != len(embs):
         raise RuntimeError(f"Mismatch ntotal({index.ntotal}) vs emb({len(embs)})")
-
     return SubjectAssets(index=index, texts=texts, embs=embs)
 
-# ========= ENCODER & RETRIEVAL =========
+# ========= ENCODER =========
+
 @torch.inference_mode()
+@lru_cache(maxsize=1024)
 def encode_query_exact(text: str) -> np.ndarray:
     toks = ENCODER_TOKENIZER(text, padding=True, truncation=True, return_tensors="pt").to(ENCODER_DEVICE)
     out = ENCODER_MODEL(**toks)
@@ -200,7 +230,10 @@ def encode_query_exact(text: str) -> np.ndarray:
 
 def cosine_sim(a: np.ndarray, b: np.ndarray) -> float:
     a = np.asarray(a).reshape(-1); b = np.asarray(b).reshape(-1)
-    return float(np.dot(a, b) / ((np.linalg.norm(a) * np.linalg.norm(b)) + 1e-12))
+    denom = (np.linalg.norm(a) * np.linalg.norm(b)) + 1e-12
+    return float(np.dot(a, b) / denom)
+
+# ========= RETRIEVAL CEPAT =========
 
 def best_cosine_from_faiss(query: str, subject_key: str) -> float:
     assets = load_subject_assets(subject_key)
@@ -213,53 +246,54 @@ def best_cosine_from_faiss(query: str, subject_key: str) -> float:
             best = max(best, cosine_sim(qv, assets.embs[i]))
     return best
 
-def retrieve_rerank_cosine(query: str, subject_key: str) -> List[str]:
+def retrieve_top_chunks(query: str, subject_key: str) -> List[str]:
     assets = load_subject_assets(subject_key)
     q = encode_query_exact(query)
-    D, idx = assets.index.search(q, TOP_K_FAISS)
+    _, idx = assets.index.search(q, TOP_K_FAISS)
     idxs = [i for i in idx[0] if 0 <= i < len(assets.texts)]
-    if not idxs:
-        return []
-    qv = q.reshape(-1)
-    scores = [cosine_sim(qv, assets.embs[i]) for i in idxs]
-    pairs = sorted(zip(scores, idxs), reverse=True)
-    top_texts = [assets.texts[i] for _, i in pairs[:TOP_K_FINAL]]
-    log.info(f"[RETRIEVE] subject={subject_key} | top={len(top_texts)}")
-    return top_texts
-
-def pick_best_sentences(query: str, chunks: List[str], top_k: int = 5) -> List[str]:
-    if not chunks: return []
-    qv = encode_query_exact(query).reshape(-1)
+    return [assets.texts[i] for i in idxs[:TOP_K_FINAL]]
+
+def pick_best_sentences_fast(query: str, chunks: List[str], top_k: int = 4) -> List[str]:
+    # Tanpa encode per kalimat — hanya lexical overlap + panjang wajar
     cands: List[Tuple[float, str]] = []
     for ch in chunks:
-        for s in split_sentences(ch):
-            sv = encode_query_exact(s).reshape(-1)
-            cos = cosine_sim(qv, sv)
+        for s in split_sentences_fast(ch):
             ovl = lexical_overlap(query, s)
-            penalty = 0.1 if len(s) < 50 else 0.0
-            score = 0.7 * cos + 0.3 * ovl - penalty
-            if score >= MIN_HYBRID:
-                cands.append((score, s))
+            if ovl < MIN_LEXICAL:
+                continue
+            # bonus sedikit kalau kalimat panjang wajar (50–220 char)
+            L = len(s)
+            len_bonus = 0.05 if 50 <= L <= 220 else 0.0
+            score = ovl + len_bonus
+            cands.append((score, s))
     cands.sort(key=lambda x: x[0], reverse=True)
     return [s for _, s in cands[:top_k]]
 
+# ========= PROMPT =========
+
 def build_prompt(user_query: str, sentences: List[str]) -> str:
     block = "\n".join(f"- {clean_prefix(s)}" for s in sentences)
     system = (
-        "- Gunakan HANYA daftar kalimat fakta berikut sebagai sumber.\n"
-        "- Jika tidak ada kalimat yang menjawab, balas: maap pengetahuan tidak ada dalam database\n"
-        "- Jawab TEPAT 1 kalimat, ringkas, Bahasa Indonesia baku.\n"
-        "- DILARANG menulis frasa meta seperti 'berdasarkan', 'menurut', 'merujuk', atau 'bersumber'."
+        "Kamu asisten RAG.\n"
+        f"- Jika tidak ada kalimat yang relevan, tulis persis: {FALLBACK_TEXT}\n"
+        "- Jawab TEPAT 1 kalimat, ringkas, Bahasa Indonesia baku (≥ 6 kata).\n"
+        "- Tanpa frasa meta (berdasarkan/menurut/merujuk/mengacu/bersumber).\n"
+        "- Tulis jawaban final di dalam tag <final>Jawaban.</final> dan jangan menulis apa pun setelah </final>."
+    )
+    fewshot = (
+        "Contoh format: \n"
+        "KALIMAT SUMBER:\n- Air memuai saat dipanaskan.\n"
+        "PERTANYAAN: Apa yang terjadi pada air saat dipanaskan?\n"
+        "<final>Air akan memuai ketika dipanaskan.</final>\n"
     )
-    return f"""{system}
-KALIMAT SUMBER:
-{block}
-PERTANYAAN:
-{user_query}
-JAWAB (1 kalimat saja):
-"""
-
-@lru_cache(maxsize=512)
+    return (
+        f"{system}\n\n{fewshot}\n"
+        f"KALIMAT SUMBER:\n{block}\n\n"
+        f"PERTANYAAN: {user_query}\n"
+        f"TULIS JAWABAN DI DALAM <final>...</final> SAJA:"
+    )
+
+@lru_cache(maxsize=1024)
 def validate_input_cached(q: str) -> bool:
     try:
         return validate_input(q)
@@ -270,7 +304,7 @@ def validate_input_cached(q: str) -> bool:
 # ========= AUTH (POSTGRES) =========
 from werkzeug.security import generate_password_hash, check_password_hash
 from sqlalchemy import create_engine, Column, Integer, String, Text, Boolean, func, or_
-from sqlalchemy.orm import sessionmaker, scoped_session, declarative_base
+from sqlalchemy.orm import sessionmaker, scoped_session, declarative_base, Session
 
 POSTGRES_URL = os.environ.get("POSTGRES_URL")
 if not POSTGRES_URL:
@@ -282,25 +316,26 @@ Base = declarative_base()
 
 class User(Base):
     __tablename__ = "users"
-    id        = Column(Integer, primary_key=True)
-    username  = Column(String(50), unique=True, nullable=False, index=True)
-    email     = Column(String(120), unique=True, nullable=False, index=True)
-    password  = Column(Text, nullable=False)
+    id = Column(Integer, primary_key=True)
+    username = Column(String(50), unique=True, nullable=False, index=True)
+    email = Column(String(120), unique=True, nullable=False, index=True)
+    password = Column(Text, nullable=False)
     is_active = Column(Boolean, default=True, nullable=False)
     is_admin  = Column(Boolean, default=False, nullable=False)
 
 class ChatHistory(Base):
     __tablename__ = "chat_history"
-    id          = Column(Integer, primary_key=True)
-    user_id     = Column(Integer, nullable=False, index=True)
+    id = Column(Integer, primary_key=True)
+    user_id = Column(Integer, nullable=False, index=True)
     subject_key = Column(String(50), nullable=False, index=True)
-    role        = Column(String(10), nullable=False)
-    message     = Column(Text, nullable=False)
-    timestamp   = Column(Integer, server_default=func.extract("epoch", func.now()))
+    role = Column(String(10), nullable=False)
+    message = Column(Text, nullable=False)
+    timestamp = Column(Integer, server_default=func.extract("epoch", func.now()))
 
 Base.metadata.create_all(bind=engine)
 
 JKT_TZ = ZoneInfo("Asia/Jakarta")
+
 @app.template_filter("fmt_ts")
 def fmt_ts(epoch_int: int):
     try:
@@ -313,14 +348,15 @@ 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"))
@@ -328,10 +364,8 @@ 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
 
-
 # ========= ROUTES =========
 @app.route("/")
 def root():
@@ -341,38 +375,30 @@ def root():
 def auth_login():
     if request.method == "POST":
         identity = (
-            request.form.get("identity")
-            or request.form.get("email")
-            or request.form.get("username")
-            or ""
+            request.form.get("identity") or request.form.get("email") or request.form.get("username") or ""
         ).strip().lower()
         pw_input = (request.form.get("password") or "").strip()
-
         if not identity or not pw_input:
             flash("Mohon isi email/username dan password.", "error")
             return render_template("login.html"), 400
-
         s = db()
         try:
             user = (
                 s.query(User)
-                 .filter(or_(func.lower(User.username) == identity,
-                             func.lower(User.email) == identity))
-                 .first()
+                .filter(or_(func.lower(User.username) == identity, func.lower(User.email) == identity))
+                .first()
             )
             log.info(f"[LOGIN] identity='{identity}' found={bool(user)} active={getattr(user,'is_active',None)}")
             ok = bool(user and user.is_active and check_password_hash(user.password, pw_input))
         finally:
             s.close()
-
         if not ok:
             flash("Identitas atau password salah.", "error")
             return render_template("login.html"), 401
-
         session["logged_in"] = True
-        session["user_id"]   = user.id
-        session["username"]  = user.username
-        session["is_admin"]  = bool(user.is_admin)
+        session["user_id"] = user.id
+        session["username"] = user.username
+        session["is_admin"] = bool(user.is_admin)
         log.info(f"[LOGIN] OK user_id={user.id}; session set.")
         return redirect(url_for("subjects"))
     return render_template("login.html")
@@ -386,7 +412,6 @@ def whoami():
         "is_admin": session.get("is_admin"),
     }
 
-
 @app.route("/auth/register", methods=["GET", "POST"])
 def auth_register():
     if request.method == "POST":
@@ -407,9 +432,8 @@ def auth_register():
         try:
             existed = (
                 s.query(User)
-                 .filter(or_(func.lower(User.username) == username,
-                             func.lower(User.email) == email))
-                 .first()
+                .filter(or_(func.lower(User.username) == username, func.lower(User.email) == email))
+                .first()
             )
             if existed:
                 flash("Username/Email sudah terpakai.", "error")
@@ -431,14 +455,12 @@ def auth_logout():
 def about():
     return render_template("about.html")
 
-
 @app.route("/subjects")
 @login_required
 def subjects():
     log.info(f"[SESSION DEBUG] logged_in={session.get('logged_in')} user_id={session.get('user_id')}")
     return render_template("home.html", subjects=SUBJECTS)
 
-
 @app.route("/chat/<subject_key>")
 @login_required
 def chat_subject(subject_key: str):
@@ -446,39 +468,41 @@ def chat_subject(subject_key: str):
         return redirect(url_for("subjects"))
     session["subject_selected"] = subject_key
     label = SUBJECTS[subject_key]["label"]
-
     s = db()
     try:
         uid = session.get("user_id")
         rows = (
             s.query(ChatHistory)
-             .filter_by(user_id=uid, subject_key=subject_key)
-             .order_by(ChatHistory.id.asc())
-             .all()
+            .filter_by(user_id=uid, subject_key=subject_key)
+            .order_by(ChatHistory.id.asc())
+            .all()
         )
         history = [{"role": r.role, "message": r.message} for r in rows]
     finally:
         s.close()
-
     return render_template("chat.html", subject=subject_key, subject_label=label, history=history)
 
 @app.route("/health")
 def health():
-    return jsonify({"ok": True, "encoder_loaded": ENCODER_MODEL is not None, "llm_loaded": LLM is not None})
+    return jsonify({
+        "ok": True,
+        "encoder_loaded": ENCODER_MODEL is not None,
+        "llm_loaded": LLM is not None,
+        "model_path": MODEL_PATH,
+        "ctx_window": CTX_WINDOW,
+        "threads": N_THREADS,
+    })
 
 @app.route("/ask/<subject_key>", methods=["POST"])
 @login_required
 def ask(subject_key: str):
     if subject_key not in SUBJECTS:
         return jsonify({"ok": False, "error": "invalid subject"}), 400
-
-    # pastikan model siap saat request (lazy)
     warmup_models()
-
     t0 = time.perf_counter()
-    data  = request.get_json(silent=True) or {}
-    query = (data.get("message") or "").strip()
 
+    data = request.get_json(silent=True) or {}
+    query = (data.get("message") or "").strip()
     if not query:
         return jsonify({"ok": False, "error": "empty query"}), 400
     if not validate_input_cached(query):
@@ -495,71 +519,112 @@ def ask(subject_key: str):
     if best < MIN_COSINE:
         return jsonify({"ok": True, "answer": FALLBACK_TEXT})
 
-    chunks = retrieve_rerank_cosine(query, subject_key)
+    chunks = retrieve_top_chunks(query, subject_key)
     if not chunks:
         return jsonify({"ok": True, "answer": FALLBACK_TEXT})
-    sentences = pick_best_sentences(query, chunks, top_k=5)
+
+    sentences = pick_best_sentences_fast(query, chunks, top_k=4)
     if not sentences:
         return jsonify({"ok": True, "answer": FALLBACK_TEXT})
 
     prompt = build_prompt(query, sentences)
 
     try:
-        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()
+        # PASS-1: deterministik & singkat
+        raw_answer = generate(
+            LLM,
+            prompt,
+            max_tokens=int(os.environ.get("MAX_TOKENS", 72)),
+            temperature=float(os.environ.get("TEMP", 0.0)),
+            top_p=1.0,
+            stop=["</final>"]
+        ) or ""
+        raw_answer = raw_answer.strip()
+        log.info(f"[LLM] Raw answer repr (pass1): {repr(raw_answer)}")
+
+        text = re.sub(r"<think\b[^>]*>.*?</think>", "", raw_answer, flags=re.DOTALL | re.IGNORECASE).strip()
+        text = re.sub(r"</?think\b[^>]*>", "", text, flags=re.IGNORECASE).strip()
+        m_final = re.search(r"<final>\s*(.+)$", text, flags=re.IGNORECASE | re.DOTALL)
+        cleaned = (m_final.group(1).strip() if m_final else re.sub(r"<[^>]+>", "", text).strip())
+
+        def _is_bad(s: str) -> bool:
+            s2 = s.strip()
+            return (len(re.sub(r"[^A-Za-zÀ-ÖØ-öø-ÿ]+", "", s2)) < 3) or (s2 in {"...", ".", "..", "…"}) or (len(s2.split()) < 6)
+
+        if _is_bad(cleaned):
+            prompt_retry = (
+                prompt
+                + "\n\nULANGI DENGAN TAAT FORMAT: Tulis satu kalimat faktual tanpa placeholder/ellipsis, minimal 6 kata, mulai huruf kapital dan akhiri titik. Tulis hanya di dalam <final>...</final>."
+            )
+            raw_answer2 = generate(
+                LLM,
+                prompt_retry,
+                max_tokens=int(os.environ.get("MAX_TOKENS", 72)),
+                temperature=0.2,
+                top_p=1.0,
+                stop=["</final>"]
+            ) or ""
+            raw_answer2 = raw_answer2.strip()
+            log.info(f"[LLM] Raw answer repr (pass2): {repr(raw_answer2)}")
+            text2 = re.sub(r"<think\b[^>]*>.*?</think>", "", raw_answer2, flags=re.DOTALL | re.IGNORECASE).strip()
+            text2 = re.sub(r"</?think\b[^>]*>", "", text2, flags=re.IGNORECASE).strip()
+            m_final2 = re.search(r"<final>\s*(.+)$", text2, flags=re.IGNORECASE | re.DOTALL)
+            cleaned2 = (m_final2.group(1).strip() if m_final2 else re.sub(r"<[^>]+>", "", text2).strip())
+            cleaned = cleaned2 or cleaned
+
+        answer = cleaned
+
     except Exception as e:
         log.exception(f"[LLM] generate error: {e}")
         return jsonify({"ok": True, "answer": FALLBACK_TEXT})
 
+    # Ambil 1 kalimat pertama saja
     m = re.search(r"(.+?[.!?])(\s|$)", answer)
     answer = (m.group(1) if m else answer).strip()
     answer = strip_meta_sentence(answer)
 
-    # === Simpan ke history ===
+    # Simpan history
     try:
         s = db()
         uid = session.get("user_id")
         s.add_all([
             ChatHistory(user_id=uid, subject_key=subject_key, role="user", message=query),
-            ChatHistory(user_id=uid, subject_key=subject_key, role="bot", message=answer)
+            ChatHistory(user_id=uid, subject_key=subject_key, role="bot",  message=answer),
         ])
         s.commit()
     except Exception as e:
         log.exception(f"[DB] gagal simpan chat history: {e}")
     finally:
-        s.close()
+        try:
+            s.close()
+        except Exception:
+            pass
 
     if not answer or len(answer) < 2:
         answer = FALLBACK_TEXT
 
     if ENABLE_PROFILING:
-        log.info({"latency_total": time.perf_counter() - t0, "subject": subject_key, "faiss_best": best})
+        log.info({
+            "latency_total": time.perf_counter() - t0,
+            "subject": subject_key,
+            "faiss_best": best,
+        })
 
     return jsonify({"ok": True, "answer": answer})
 
-# ===== Admin views & delete actions (tetap) =====
-from sqlalchemy.orm import Session
+# ===== Admin =====
 @app.route("/admin")
 @admin_required
 def admin_dashboard():
     s = db()
     try:
-        total_users   = s.query(func.count(User.id)).scalar() or 0
-        total_active  = s.query(func.count(User.id)).filter(User.is_active.is_(True)).scalar() or 0
-        total_admins  = s.query(func.count(User.id)).filter(User.is_admin.is_(True)).scalar() or 0
-        total_msgs    = s.query(func.count(ChatHistory.id)).scalar() or 0
+        total_users  = s.query(func.count(User.id)).scalar() or 0
+        total_active = s.query(func.count(User.id)).filter(User.is_active.is_(True)).scalar() or 0
+        total_admins = s.query(func.count(User.id)).filter(User.is_admin.is_(True)).scalar() or 0
+        total_msgs   = s.query(func.count(ChatHistory.id)).scalar() or 0
     finally:
         s.close()
-    return render_template("admin_dashboard.html",
-                           total_users=total_users,
-                           total_active=total_active,
-                           total_admins=total_admins,
-                           total_msgs=total_msgs)
+    return render_template("admin_dashboard.html", total_users=total_users, total_active=total_active, total_admins=total_admins, total_msgs=total_msgs)
 
 @app.route("/admin/users")
 @admin_required
@@ -571,59 +636,39 @@ def admin_users():
     try:
         base = s.query(User)
         if q:
-            base = base.filter(or_(
-                func.lower(User.username).like(f"%{q}%"),
-                func.lower(User.email).like(f"%{q}%")
-            ))
+            base = base.filter(or_(func.lower(User.username).like(f"%{q}%"), func.lower(User.email).like(f"%{q}%")))
         total = base.count()
-        users = (base
-                 .order_by(User.id.asc())
-                 .offset((page - 1) * per_page)
-                 .limit(per_page)
-                 .all())
+        users = base.order_by(User.id.asc()).offset((page - 1) * per_page).limit(per_page).all()
         user_ids = [u.id for u in users] or [-1]
-        counts = dict(s.query(ChatHistory.user_id, func.count(ChatHistory.id))
-                        .filter(ChatHistory.user_id.in_(user_ids))
-                        .group_by(ChatHistory.user_id)
-                        .all())
+        counts = dict(s.query(ChatHistory.user_id, func.count(ChatHistory.id)).filter(ChatHistory.user_id.in_(user_ids)).group_by(ChatHistory.user_id).all())
     finally:
         s.close()
-    return render_template("admin_users.html",
-                           users=users, counts=counts,
-                           q=q, page=page, per_page=per_page, total=total)
+    return render_template("admin_users.html", users=users, counts=counts, q=q, page=page, per_page=per_page, total=total)
 
 @app.route("/admin/history")
 @admin_required
 def admin_history():
-    q          = (request.args.get("q") or "").strip().lower()
-    username   = (request.args.get("username") or "").strip().lower()
-    subject    = (request.args.get("subject") or "").strip().lower()
-    role       = (request.args.get("role") or "").strip().lower()
-    page       = max(int(request.args.get("page", 1)), 1)
-    per_page   = min(max(int(request.args.get("per_page", 30)), 5), 200)
-
+    q        = (request.args.get("q") or "").strip().lower()
+    username = (request.args.get("username") or "").strip().lower()
+    subject  = (request.args.get("subject") or "").strip().lower()
+    role     = (request.args.get("role") or "").strip().lower()
+    page = max(int(request.args.get("page", 1)), 1)
+    per_page = min(max(int(request.args.get("per_page", 30)), 5), 200)
     s = db()
     try:
         base = (s.query(ChatHistory, User).join(User, User.id == ChatHistory.user_id))
         if q:
             base = base.filter(func.lower(ChatHistory.message).like(f"%{q}%"))
         if username:
-            base = base.filter(or_(
-                func.lower(User.username) == username,
-                func.lower(User.email) == username
-            ))
+            base = base.filter(or_(func.lower(User.username) == username, func.lower(User.email) == username))
         if subject:
             base = base.filter(func.lower(ChatHistory.subject_key) == subject)
         if role in ("user", "bot"):
             base = base.filter(ChatHistory.role == role)
         total = base.count()
-        rows = (base.order_by(ChatHistory.id.desc())
-                .offset((page - 1) * per_page)
-                .limit(per_page)
-                .all())
+        rows = base.order_by(ChatHistory.id.desc()).offset((page - 1) * per_page).limit(per_page).all()
     finally:
         s.close()
-
     items = [{
         "id": r.ChatHistory.id,
         "username": r.User.username,
@@ -633,11 +678,7 @@ def admin_history():
         "message": r.ChatHistory.message,
         "timestamp": r.ChatHistory.timestamp,
     } for r in rows]
-
-    return render_template("admin_history.html",
-                           items=items, subjects=SUBJECTS,
-                           q=q, username=username, subject=subject, role=role,
-                           page=page, per_page=per_page, total=total)
+    return render_template("admin_history.html", items=items, subjects=SUBJECTS, q=q, username=username, subject=subject, role=role, page=page, per_page=per_page, total=total)
 
 def _is_last_admin(s: Session) -> bool:
     return (s.query(func.count(User.id)).filter(User.is_admin.is_(True)).scalar() or 0) <= 1
@@ -708,4 +749,4 @@ def admin_delete_chat(chat_id: int):
 # ========= ENTRY =========
 if __name__ == "__main__":
     port = int(os.environ.get("PORT", 7860))
-    app.run(host="0.0.0.0", port=port, debug=False)
+    app.run(host="0.0.0.0", port=port, debug=False)