feat: Migrate the default LLM from ChatGLM3 to Qwen2-0.5B-Instruct, updating model loading, generation logic, and documentation.

.env.example

@@ -1,5 +1,5 @@
 # Model Configuration
-MODEL_NAME=THUDM/chatglm3-6b
+MODEL_NAME=Qwen/Qwen2-0.5B-Instruct
 EMBEDDING_MODEL=sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
 
 # Device Configuration (auto/cuda/cpu)

QUICKSTART.md

@@ -22,7 +22,7 @@ Aplikasi akan:
 2. Inisialisasi vector database
 3. Launch Gradio interface di `http://localhost:7860`
 
-**Catatan**: Model GLM akan di-download otomatis saat pertama kali digunakan (ukuran ~13GB untuk ChatGLM3-6B).
+**Catatan**: Model Qwen2 akan di-download otomatis saat pertama kali digunakan (ukuran ~1GB, jauh lebih kecil dari model sebelumnya).
 
 ## 📚 Workflow Penggunaan
 

README.md

@@ -32,7 +32,7 @@ short_description: Chat dengan dokumen PDF menggunakan RAG dan GLM Model
 
 RAG ChatBot adalah aplikasi AI yang memungkinkan Anda untuk mengupload dokumen PDF dan melakukan tanya jawab interaktif tentang isi dokumen tersebut. Sistem menggunakan:
 
-- **ChatGLM3-6B**: Model bahasa generatif untuk menghasilkan jawaban
+- **Qwen2-0.5B-Instruct**: Model bahasa generatif yang ringan untuk menghasilkan jawaban
 - **RAG (Retrieval-Augmented Generation)**: Teknik untuk mencari informasi relevan dari dokumen
 - **ChromaDB**: Vector database untuk penyimpanan dan pencarian semantic
 - **Gradio**: Interface web yang modern dan interaktif
@@ -145,7 +145,7 @@ Aplikasi akan berjalan di `http://localhost:7860`
     └────┬─────┘
          │
     ┌────▼─────┐
-    │ChatGLM3  │  (Response Generation)
+    │  Qwen2   │  (Response Generation)
     └──────────┘
 ```
 
@@ -183,7 +183,7 @@ Edit file `.env` untuk mengatur konfigurasi:
 
 ```bash
 # Model
-MODEL_NAME=THUDM/chatglm3-6b
+MODEL_NAME=Qwen/Qwen2-0.5B-Instruct
 EMBEDDING_MODEL=sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
 
 # Device (auto/cuda/cpu)
@@ -227,8 +227,8 @@ Berikut dependencies utama yang digunakan:
 
 ### Model Loading Error
 ```bash
-# Jika model terlalu besar, gunakan quantized version
-MODEL_NAME=THUDM/chatglm3-6b-32k
+# Model saat ini sudah sangat ringan (~1GB)
+# Jika masih ada masalah, pastikan koneksi internet stabil untuk download
 ```
 
 ### PDF Extraction Error
@@ -254,5 +254,5 @@ Untuk pertanyaan dan support, silakan buat issue di repository ini.
 ---
 
 <div align="center">
-Made with ❤️ using Gradio and ChatGLM
+Made with ❤️ using Gradio and Qwen2
 </div>

app.py

@@ -375,7 +375,7 @@ with gr.Blocks(css=CUSTOM_CSS, theme=gr.themes.Soft(), title="RAG ChatBot - GLM"
             
             **Tech Stack:**
             - Framework: Gradio
-            - LLM: ChatGLM3 (Transformers)
+            - LLM: Qwen2 (Transformers)
             - Embeddings: Sentence Transformers
             - Vector DB: ChromaDB
             - PDF Processing: PyPDF2 + pdfplumber
@@ -385,7 +385,7 @@ with gr.Blocks(css=CUSTOM_CSS, theme=gr.themes.Soft(), title="RAG ChatBot - GLM"
                 gr.Markdown("""
                 Untuk mengubah konfigurasi model, edit file `.env`:
                 ```bash
-                MODEL_NAME=THUDM/chatglm3-6b
+                MODEL_NAME=Qwen/Qwen2-0.5B-Instruct
                 DEVICE=auto
                 CHUNK_SIZE=500
                 CHUNK_OVERLAP=50
@@ -401,7 +401,7 @@ if __name__ == "__main__":
     os.makedirs(config.VECTOR_DB_DIR, exist_ok=True)
     
     print("=" * 60)
-    print("🚀 Launching RAG ChatBot dengan GLM")
+    print("🚀 Launching RAG ChatBot dengan Qwen2")
     print("=" * 60)
     print(f"Model: {config.MODEL_NAME}")
     print(f"Device: {config.DEVICE}")

config/model_config.py

@@ -11,7 +11,7 @@ class Config:
     """Configuration class for RAG ChatBot"""
     
     # Model Settings
-    MODEL_NAME = os.getenv("MODEL_NAME", "THUDM/chatglm3-6b")
+    MODEL_NAME = os.getenv("MODEL_NAME", "Qwen/Qwen2-0.5B-Instruct")
     EMBEDDING_MODEL = os.getenv("EMBEDDING_MODEL", "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2")
     
     # Device Configuration

requirements.txt

@@ -11,6 +11,7 @@ langchain>=0.1.0
 langchain-community>=0.0.20
 langchain-text-splitters>=0.0.1
 sentencepiece
+dontenv
 
 # PDF Processing
 PyPDF2>=3.0.0

utils/rag_pipeline.py

@@ -3,7 +3,7 @@ RAG Pipeline for retrieving relevant context and generating responses
 """
 from typing import List, Dict, Optional
 import torch
-from transformers import AutoTokenizer, AutoModel
+from transformers import AutoTokenizer, AutoModelForCausalLM
 from config.model_config import config
 from utils.vector_store import VectorStore
 
@@ -29,7 +29,7 @@ class RAGPipeline:
         return config.DEVICE
     
     def load_model(self):
-        """Load GLM model and tokenizer"""
+        """Load Qwen2 model and tokenizer"""
         if self.model is not None:
             print("Model already loaded")
             return
@@ -38,16 +38,23 @@ class RAGPipeline:
         print(f"Using device: {self.device}")
         
         try:
+            from transformers import AutoModelForCausalLM
+            
             self.tokenizer = AutoTokenizer.from_pretrained(
                 config.MODEL_NAME,
                 trust_remote_code=True
             )
             
-            self.model = AutoModel.from_pretrained(
+            self.model = AutoModelForCausalLM.from_pretrained(
                 config.MODEL_NAME,
                 trust_remote_code=True,
-                torch_dtype=torch.float16 if self.device == "cuda" else torch.float32
-            ).to(self.device)
+                torch_dtype=torch.float16 if self.device == "cuda" else torch.float32,
+                device_map=self.device if self.device == "cuda" else None
+            )
+            
+            # Move to device if CPU
+            if self.device == "cpu":
+                self.model = self.model.to(self.device)
             
             # Set to evaluation mode
             self.model = self.model.eval()
@@ -113,7 +120,7 @@ class RAGPipeline:
         
         Args:
             query: User query
-            history: Chat history (for ChatGLM format)
+            history: Chat history (list of [user_msg, assistant_msg] pairs)
             use_rag: Whether to use RAG retrieval
             max_length: Maximum response length
             temperature: Sampling temperature
@@ -146,17 +153,58 @@ class RAGPipeline:
         else:
             prompt = query
         
-        # Generate response using ChatGLM
+        # Build messages for Qwen2 chat format
+        messages = []
+        
+        # Add system message
+        messages.append({
+            "role": "system",
+            "content": config.SYSTEM_PROMPT
+        })
+        
+        # Add chat history
+        if history:
+            for user_msg, assistant_msg in history:
+                if user_msg:
+                    messages.append({"role": "user", "content": user_msg})
+                if assistant_msg:
+                    messages.append({"role": "assistant", "content": assistant_msg})
+        
+        # Add current query
+        messages.append({"role": "user", "content": prompt})
+        
+        # Generate response using Qwen2
         try:
-            response, history = self.model.chat(
-                self.tokenizer,
-                prompt,
-                history=history or [],
-                max_length=max_length,
-                temperature=temperature,
-                top_p=top_p
+            # Apply chat template
+            text = self.tokenizer.apply_chat_template(
+                messages,
+                tokenize=False,
+                add_generation_prompt=True
             )
             
+            # Tokenize
+            model_inputs = self.tokenizer([text], return_tensors="pt").to(self.device)
+            
+            # Generate
+            with torch.no_grad():
+                generated_ids = self.model.generate(
+                    model_inputs.input_ids,
+                    max_new_tokens=max_length,
+                    temperature=temperature,
+                    top_p=top_p,
+                    do_sample=True,
+                    pad_token_id=self.tokenizer.pad_token_id,
+                    eos_token_id=self.tokenizer.eos_token_id
+                )
+            
+            # Decode response
+            generated_ids = [
+                output_ids[len(input_ids):] 
+                for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
+            ]
+            
+            response = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+            
             return response, sources
             
         except Exception as e:
@@ -211,17 +259,69 @@ class RAGPipeline:
         else:
             prompt = query
         
-        # Stream response using ChatGLM
+        # Build messages for Qwen2 chat format
+        messages = []
+        
+        # Add system message
+        messages.append({
+            "role": "system",
+            "content": config.SYSTEM_PROMPT
+        })
+        
+        # Add chat history
+        if history:
+            for user_msg, assistant_msg in history:
+                if user_msg:
+                    messages.append({"role": "user", "content": user_msg})
+                if assistant_msg:
+                    messages.append({"role": "assistant", "content": assistant_msg})
+        
+        # Add current query
+        messages.append({"role": "user", "content": prompt})
+        
+        # Stream response using Qwen2
         try:
-            for response, history in self.model.stream_chat(
+            from transformers import TextIteratorStreamer
+            from threading import Thread
+            
+            # Apply chat template
+            text = self.tokenizer.apply_chat_template(
+                messages,
+                tokenize=False,
+                add_generation_prompt=True
+            )
+            
+            # Tokenize
+            model_inputs = self.tokenizer([text], return_tensors="pt").to(self.device)
+            
+            # Setup streamer
+            streamer = TextIteratorStreamer(
                 self.tokenizer,
-                prompt,
-                history=history or [],
-                max_length=max_length,
+                skip_prompt=True,
+                skip_special_tokens=True
+            )
+            
+            # Generation kwargs
+            generation_kwargs = dict(
+                model_inputs,
+                streamer=streamer,
+                max_new_tokens=max_length,
                 temperature=temperature,
-                top_p=top_p
-            ):
-                yield response, sources
+                top_p=top_p,
+                do_sample=True,
+                pad_token_id=self.tokenizer.pad_token_id,
+                eos_token_id=self.tokenizer.eos_token_id
+            )
+            
+            # Start generation in separate thread
+            thread = Thread(target=self.model.generate, kwargs=generation_kwargs)
+            thread.start()
+            
+            # Stream output
+            full_response = ""
+            for new_text in streamer:
+                full_response += new_text
+                yield full_response, sources
                 
         except Exception as e:
             print(f"Error streaming response: {e}")