Update app.py

app.py

@@ -1,4 +1,3 @@
-
 import os, re, json, pickle, hashlib, requests
 from pathlib import Path
 import gradio as gr
@@ -8,31 +7,35 @@ from sklearn.neighbors import NearestNeighbors
 from sentence_transformers import SentenceTransformer
 
 # =================== Config ===================
-DATA_PATH = Path(os.getenv("DATA_PATH", "IPLM_QnA_Chatbot.jsonl"))  # default filename
+DATA_PATH = Path(os.getenv("DATA_PATH", "IPLM_QnA_Chatbot.jsonl"))  # nama file JSONL kamu
 CACHE_EMB = Path("embeddings.pkl")
 CACHE_META = Path("meta.json")
 
-# Embedding model for retrieval
+# Embedding model untuk retrieval (cepat & akurat)
 EMB_MODEL = os.getenv("EMB_MODEL", "sentence-transformers/all-MiniLM-L6-v2")
 
-# LLM endpoint (HF Inference API / TGI-compatible / OpenAI-compatible route)
+# LLM endpoint (Inference API / TGI-compatible / OpenAI-compatible route)
 HF_CHAT_URL = os.getenv("HF_CHAT_URL", "https://api-inference.huggingface.co/v1/chat/completions")
 HF_TOKEN    = os.getenv("HF_TOKEN", "")
-LLM_MODEL   = os.getenv("LLM_MODEL", "meta-llama/Meta-Llama-3.1-8B-Instruct")
+LLM_MODEL   = os.getenv("LLM_MODEL", "Qwen/Qwen2.5-7B-Instruct")  # default publik (non-gated)
 
 TOP_K_DEFAULT = int(os.getenv("TOP_K_DEFAULT", "4"))
 TEMPERATURE_DEFAULT = float(os.getenv("TEMPERATURE_DEFAULT", "0.3"))
 MAX_TOKENS = int(os.getenv("MAX_TOKENS", "512"))
 
-SYSTEM_PROMPT = os.getenv("SYSTEM_PROMPT", 
-    "You are an Indonesian librarian assistant. Jawab ringkas, akurat, dan sopan. "
-    "Gunakan HANYA informasi dari konteks yang diberikan. Jika konteks tidak memuat jawabannya, "
-    "katakan bahwa data tidak tersedia di basis pengetahuan."
+SYSTEM_PROMPT = os.getenv(
+    "SYSTEM_PROMPT",
+    (
+        "You are an Indonesian librarian assistant. Jawab ringkas, akurat, dan sopan. "
+        "Gunakan HANYA informasi dari konteks yang diberikan. Jika konteks tidak memuat jawabannya, "
+        "katakan bahwa data tidak tersedia di basis pengetahuan."
+    ),
 )
 
 # =================== Utils ===================
 def norm(s: str) -> str:
-    if s is None: return ""
+    if s is None:
+        return ""
     s = str(s).strip()
     s = re.sub(r"\s+", " ", s)
     return s
@@ -50,20 +53,21 @@ def load_jsonl(path: Path) -> list:
     with path.open("r", encoding="utf-8") as f:
         for line in f:
             line = line.strip()
-            if not line: continue
+            if not line:
+                continue
             obj = json.loads(line)
-            # support various key names
+            # dukung berbagai nama key
             q = obj.get("question") or obj.get("pertanyaan") or obj.get("q")
-            a = obj.get("answer")   or obj.get("jawaban")    or obj.get("a")
+            a = obj.get("answer") or obj.get("jawaban") or obj.get("a")
             if q and a:
                 rows.append({"question": norm(q), "answer": norm(a)})
     if not rows:
         raise ValueError("JSONL kosong atau tidak mengandung pasangan 'question'/'answer'.")
-    # drop dup by question
+    # buang duplikat berdasarkan question
     seen = set()
     uniq = []
     for r in rows:
-        if r["question"] in seen: 
+        if r["question"] in seen:
             continue
         seen.add(r["question"])
         uniq.append(r)
@@ -79,32 +83,46 @@ class FAQIndex:
 
     def build(self, rows: list, force=False):
         self.rows = rows
-        # try load cache
+        # coba muat cache
         if not force and CACHE_EMB.exists() and CACHE_META.exists():
             try:
                 meta = json.loads(CACHE_META.read_text(encoding="utf-8"))
                 if meta.get("hash") == dataset_hash(rows) and meta.get("emb_model") == EMB_MODEL:
                     cached = pickle.loads(CACHE_EMB.read_bytes())
                     self.emb = cached["emb"]
-                    self.nn  = cached["nn"]
+                    self.nn = cached["nn"]
                     if self.model is None:
                         self.model = SentenceTransformer(EMB_MODEL)
                     return
             except Exception:
                 pass
-        # build fresh
+        # bangun baru
         self.model = SentenceTransformer(EMB_MODEL)
-        # encode "Q: ...\nA: ..." for better grounding
+        # encode gabungan Q+A → lebih “grounded”
         qas = [f"Q: {r['question']}\nA: {r['answer']}" for r in rows]
-        self.emb = self.model.encode(qas, normalize_embeddings=True, convert_to_numpy=True, show_progress_bar=False)
-        self.nn  = NearestNeighbors(n_neighbors=min(10, len(qas)), metric="cosine").fit(self.emb)
+        self.emb = self.model.encode(
+            qas,
+            normalize_embeddings=True,
+            convert_to_numpy=True,
+            show_progress_bar=False,
+        )
+        self.nn = NearestNeighbors(
+            n_neighbors=min(10, len(qas)), metric="cosine"
+        ).fit(self.emb)
         CACHE_EMB.write_bytes(pickle.dumps({"emb": self.emb, "nn": self.nn}))
-        CACHE_META.write_text(json.dumps({"hash": dataset_hash(rows), "emb_model": EMB_MODEL}, ensure_ascii=False))
+        CACHE_META.write_text(
+            json.dumps({"hash": dataset_hash(rows), "emb_model": EMB_MODEL}, ensure_ascii=False)
+        )
 
     def retrieve(self, query: str, top_k: int = TOP_K_DEFAULT):
         if not query.strip():
             return []
-        q_vec = self.model.encode([query], normalize_embeddings=True, convert_to_numpy=True, show_progress_bar=False)
+        q_vec = self.model.encode(
+            [query],
+            normalize_embeddings=True,
+            convert_to_numpy=True,
+            show_progress_bar=False,
+        )
         dists, idxs = self.nn.kneighbors(q_vec, n_neighbors=min(top_k, len(self.rows)))
         sims = 1.0 - dists[0]
         out = []
@@ -116,7 +134,7 @@ class FAQIndex:
 # =================== LLM Caller ===================
 def call_hf_chat(messages, temperature=TEMPERATURE_DEFAULT, max_tokens=MAX_TOKENS):
     if not HF_TOKEN:
-        # allow non-LLM fallback with a clear message
+        # izinkan fallback non-LLM agar Space tetap hidup
         return "⚠️ HF_TOKEN belum diatur. Buka Settings → Secrets dan tambahkan HF_TOKEN agar LLM aktif."
     headers = {"Authorization": f"Bearer {HF_TOKEN}"}
     payload = {
@@ -124,7 +142,7 @@ def call_hf_chat(messages, temperature=TEMPERATURE_DEFAULT, max_tokens=MAX_TOKEN
         "messages": messages,
         "temperature": float(temperature),
         "max_tokens": int(max_tokens),
-        "stream": False
+        "stream": False,
     }
     r = requests.post(HF_CHAT_URL, headers=headers, json=payload, timeout=90)
     try:
@@ -148,7 +166,14 @@ def rag_answer(user_msg, top_k=TOP_K_DEFAULT, temperature=TEMPERATURE_DEFAULT):
     context = build_context(hits)
     messages = [
         {"role": "system", "content": SYSTEM_PROMPT},
-        {"role": "user", "content": f"KONTEKS:\n{context}\n\nPERTANYAAN:\n{user_msg}\n\nInstruksi: Jawab berbasis KONTEKS. Jika tidak ada di konteks, jawab 'Data tidak tersedia.' "}
+        {
+            "role": "user",
+            "content": (
+                f"KONTEKS:\n{context}\n\n"
+                f"PERTANYAAN:\n{user_msg}\n\n"
+                "Instruksi: Jawab berbasis KONTEKS. Jika tidak ada di konteks, jawab 'Data tidak tersedia.' "
+            ),
+        },
     ]
     out = call_hf_chat(messages, temperature=float(temperature), max_tokens=MAX_TOKENS)
     bullets = "\n".join([f"- ({h['score']:.2f}) {h['question']}" for h in hits])
@@ -165,8 +190,10 @@ def upload_jsonl(file_obj):
         return gr.update(value="Tidak ada file.")
     tmp = Path(file_obj.name)
     tmp.replace(DATA_PATH)
-    if CACHE_EMB.exists(): CACHE_EMB.unlink()
-    if CACHE_META.exists(): CACHE_META.unlink()
+    if CACHE_EMB.exists():
+        CACHE_EMB.unlink()
+    if CACHE_META.exists():
+        CACHE_META.unlink()
     global rows, faq
     rows = load_jsonl(DATA_PATH)
     faq = FAQIndex()
@@ -175,19 +202,26 @@ def upload_jsonl(file_obj):
 
 # =================== UI ===================
 with gr.Blocks(title="RAG + LLM (JSONL)") as demo:
-    gr.Markdown("# 📚 RAG + LLM — dari JSONL Q&A\n"
-                "Masukkan pertanyaan → sistem mengambil Q&A paling relevan → LLM merangkum/menjawab berdasarkan konteks.")
+    gr.Markdown(
+        "# 📚 RAG + LLM — dari JSONL Q&A\n"
+        "Masukkan pertanyaan → sistem mengambil Q&A paling relevan → LLM merangkum/menjawab berdasarkan konteks."
+    )
     with gr.Row():
         with gr.Column(scale=2):
             chat = gr.ChatInterface(
                 fn=lambda msg, hist, k, t: rag_answer(msg, top_k=int(k), temperature=float(t)),
                 additional_inputs=[
                     gr.Slider(1, 10, value=TOP_K_DEFAULT, step=1, label="Top-K dokumen"),
-                    gr.Slider(0.0, 1.0, value=TEMPERATURE_DEFAULT, step=0.05, label="Temperatur")
+                    gr.Slider(0.0, 1.0, value=TEMPERATURE_DEFAULT, step=0.05, label="Temperatur"),
                 ],
                 title="Asisten Perpustakaan (RAG)",
                 description="Jawab *berdasarkan konteks* dari dokumen JSONL Anda.",
-                examples=["Apa itu IPLM?", "Bagaimana perhitungan TGM?", "Apa saja tahap pengolahan data?"]
+                # Harus list-of-lists karena ada additional_inputs
+                examples=[
+                    ["Apa itu IPLM?", TOP_K_DEFAULT, TEMPERATURE_DEFAULT],
+                    ["Bagaimana perhitungan TGM?", TOP_K_DEFAULT, TEMPERATURE_DEFAULT],
+                    ["Apa saja tahap pengolahan data?", TOP_K_DEFAULT, TEMPERATURE_DEFAULT],
+                ],
             )
         with gr.Column(scale=1):
             gr.Markdown("### 🔄 Perbarui Basis Data")
@@ -195,5 +229,6 @@ with gr.Blocks(title="RAG + LLM (JSONL)") as demo:
             out = gr.Textbox(label="Status", interactive=False)
             uploader.change(fn=upload_jsonl, inputs=uploader, outputs=out)
             gr.Markdown("Set **HF_TOKEN** di Settings → Secrets untuk mengaktifkan LLM.")
+
 if __name__ == "__main__":
     demo.launch()