Update app.py

app.py

@@ -6,14 +6,22 @@ import numpy as np
 from sentence_transformers import SentenceTransformer
 from huggingface_hub import InferenceClient, hf_hub_download
 
+# =======================================================
 # 1. KONFIGURASI
+# =======================================================
+print("⏳ Memulai Asisten Kesehatan (Mode Strict)...")
+
 hf_token = os.getenv("HF_TOKEN")
-REPO_ID_DATASET = "gekina/medical_dataset"  # <--- ID DATASET ANDA
+REPO_ID_DATASET = "gekina/medical_dataset"
+
+# Setup Client LLM
 client = InferenceClient(model="Qwen/Qwen2.5-7B-Instruct", token=hf_token)
 
+# =======================================================
 # 2. LOAD DATA
-print("⬇️ Mendownload data...")
+# =======================================================
 try:
+    print(f"⬇️ Download data dari {REPO_ID_DATASET}...")
     path_data = hf_hub_download(repo_id=REPO_ID_DATASET, filename="chunks_data.parquet", repo_type="dataset", token=hf_token)
     path_index = hf_hub_download(repo_id=REPO_ID_DATASET, filename="alodokter_index.faiss", repo_type="dataset", token=hf_token)
     
@@ -26,40 +34,87 @@ except Exception as e:
     df_chunks = pd.DataFrame()
     index = None
 
-# 3. LOGIKA SEARCH
-def cari(query):
+# =======================================================
+# 3. LOGIKA PENCARIAN
+# =======================================================
+def cari_dokumen(query, k=4):
     if index is None: return []
     try:
         q_emb = embedder.encode([f"query: {query}"], normalize_embeddings=True)
-        D, I = index.search(q_emb, 3)
+        D, I = index.search(q_emb, k)
+        
         results = []
         seen = set()
         for idx in I[0]:
             if idx < 0 or idx >= len(df_chunks): continue
             row = df_chunks.iloc[idx]
-            if row['parent_id'] not in seen:
+            p_id = row['parent_id']
+            if p_id not in seen:
                 results.append(row['chunk_text'])
-                seen.add(row['parent_id'])
+                seen.add(p_id)
         return results
     except: return []
 
-# 4. LOGIKA CHAT
-def respon(message, history):
-    docs = cari(message)
-    context = "\n\n".join(docs) if docs else "Tidak ada data."
-    
-    sys_prompt = "Anda Asisten Medis. Jawab berdasarkan KONTEKS. Jangan resepkan obat keras."
-    prompt = f"KONTEKS:\n{context}\n\nPERTANYAAN:\n{message}"
+# =======================================================
+# 4. LOGIKA STRICT RAG (MODIFIKASI UTAMA)
+# =======================================================
+def respon_bot(message, history):
+    # A. Cari Dokumen
+    docs = cari_dokumen(message)
     
-    msg = [{"role": "system", "content": sys_prompt}, {"role": "user", "content": prompt}]
-    
-    partial = ""
-    for chunk in client.chat_completion(msg, max_tokens=1024, stream=True):
-        if chunk.choices[0].delta.content:
-            partial += chunk.choices[0].delta.content
-            yield partial
+    # [FILTER 1] Jika mesin pencari tidak menemukan apa-apa sama sekali
+    if not docs:
+        yield "Mohon maaf, informasi tersebut tidak tersedia di dalam dataset database kami."
+        return
+
+    context_str = "\n\n".join(docs)
+
+    # B. SYSTEM PROMPT (SANGAT KETAT)
+    system_prompt = """Anda adalah Asisten Penjawab Data.
+
+ATURAN MUTLAK (STRICT MODE):
+1. Anda HANYA boleh menjawab berdasarkan informasi yang tertulis di 'KONTEKS DATA' di bawah.
+2. DILARANG menggunakan pengetahuan internal/umum Anda sendiri. Lupakan bahwa Anda adalah AI yang tahu segalanya.
+3. JIKA JAWABAN TIDAK ADA DI KONTEKS: Katakan "Mohon maaf, informasi spesifik mengenai hal ini tidak ditemukan dalam dataset kami." Jangan mencoba menjawab atau mengarang.
+4. JANGAN pernah menyebutkan "Berdasarkan konteks...", langsung saja berikan jawabannya.
+5. Gunakan Bahasa Indonesia yang baik."""
+
+    prompt_final = f"""KONTEKS DATA:\n{context_str}\n\nPERTANYAAN:\n{message}"""
+
+    messages = [
+        {"role": "system", "content": system_prompt},
+        {"role": "user", "content": prompt_final}
+    ]
+
+    try:
+        # [SETTING 2] Temperature 0.1 agar tidak kreatif/halusinasi
+        stream = client.chat_completion(
+            messages, 
+            max_tokens=1024, 
+            stream=True, 
+            temperature=0.1, 
+            top_p=0.9
+        )
+        
+        partial_text = ""
+        for chunk in stream:
+            if chunk.choices[0].delta.content:
+                partial_text += chunk.choices[0].delta.content
+                yield partial_text
+                
+    except Exception as e:
+        yield f"Error: {e}"
+
+# =======================================================
+# 5. UI GRADIO
+# =======================================================
+demo = gr.ChatInterface(
+    fn=respon_bot,
+    title="📚 Chatbot Database Medis (Strict)",
+    description="Bot ini **hanya** menjawab jika datanya ada di dataset. Jika tidak ada, bot akan menolak menjawab.",
+    theme="soft",
+    examples=["Apa obat sakit kepala?", "Cara mengatasi demam"]
+)
 
-# 5. UI
-demo = gr.ChatInterface(fn=respon, title="🏥 Chatbot Medis (RAG)", theme="soft")
 if __name__ == "__main__":
     demo.launch()