Update app.py

app.py

@@ -66,10 +66,8 @@ class LangChainAgentWrapper:
     def __init__(self):
         print("Initializing LangChainAgentWrapper...")
         
-        # Using a newer, more capable instruction-tuned model.
-        # This model is generally better at following the ReAct prompt format used by LangChain agents.
+        # We will keep using the gemma-2b-it model, but load it in 4-bit
         model_id = "google/gemma-2b-it" 
-        # model_id = "bigcode/starcoderbase-1b" # You can still use starcoder if you prefer
         
         try:
             hf_auth_token = os.getenv("HF_TOKEN")
@@ -78,28 +76,41 @@ class LangChainAgentWrapper:
             else:
                 print("HF_TOKEN secret found.")
 
+            # --- NEW: 4-Bit Quantization Configuration ---
+            # Create a configuration for loading the model in 4-bit precision.
+            # This makes the model faster and use less memory.
+            print("Creating 4-bit quantization config...")
+            quantization_config = transformers.BitsAndBytesConfig(
+                load_in_4bit=True,
+                bnb_4bit_quant_type="nf4",
+                bnb_4bit_compute_dtype="bfloat16" # Use bfloat16 for faster computation
+            )
+            print("Quantization config created.")
+            # --- END NEW ---
+
             # Create the Hugging Face pipeline
-            print(f"Loading model pipeline for: {model_id}")
+            print(f"Loading model pipeline for: {model_id} with quantization")
             llm_pipeline = transformers.pipeline(
                 "text-generation",
                 model=model_id,
-                model_kwargs={"torch_dtype": "auto"}, # Use "auto" for dtype
-                device_map="auto", # Requires accelerate
+                model_kwargs={"torch_dtype": "auto"},
+                device_map="auto",
                 token=hf_auth_token,
+                quantization_config=quantization_config # <<< --- PASS THE NEW CONFIG HERE
             )
             print("Model pipeline loaded successfully.")
             
             # Wrap the pipeline in a LangChain LLM object
             self.llm = HuggingFacePipeline(pipeline=llm_pipeline)
 
-            # Define the list of LangChain tools
+            # Define the list of LangChain tools (this part is unchanged)
             self.tools = [
                 Tool(
                     name="get_current_time_in_timezone",
                     func=get_current_time_in_timezone_func,
                     description=get_current_time_in_timezone_func.__doc__
                 ),
-                search_tool, # This is already a LangChain Tool instance
+                search_tool,
                 Tool(
                     name="safe_calculator",
                     func=safe_calculator_func,
@@ -108,8 +119,7 @@ class LangChainAgentWrapper:
             ]
             print(f"Tools prepared for agent: {[tool.name for tool in self.tools]}")
             
-            # Create the ReAct agent prompt from a template
-            # The prompt is crucial for teaching the agent how to think and use tools.
+            # Create the ReAct agent prompt from a template (this part is unchanged)
             react_prompt = PromptTemplate.from_template(
                 """
                 You are a helpful assistant. Answer the following questions as best you can.
@@ -135,10 +145,10 @@ class LangChainAgentWrapper:
                 """
             )
             
-            # Create the agent
+            # Create the agent (this part is unchanged)
             agent = create_react_agent(self.llm, self.tools, react_prompt)
             
-            # Create the agent executor, which runs the agent loop
+            # Create the agent executor (this part is unchanged)
             self.agent_executor = AgentExecutor(agent=agent, tools=self.tools, verbose=True, handle_parsing_errors=True)
             print("LangChain agent created successfully.")