Update app.py

app.py

@@ -1,28 +1,49 @@
 import gradio as gr
 from transformers import pipeline
+import torch
+import json
 
-# Setting up Coretex - Specifically for CPU Free Tier
-model_id = "Qwen/Qwen2.5-3B-Instruct"
+# Using the 1.5B version for speed and stability on free CPU
+model_id = "Qwen/Qwen2.5-1.5B-Instruct"
 
-# We removed device_map="auto" to fix the error and added a simpler loader
+# Optimized loader for CPU
 pipe = pipeline(
     "text-generation", 
     model=model_id, 
-    model_kwargs={"low_cpu_mem_usage": True}
+    model_kwargs={"torch_dtype": torch.float32, "low_cpu_mem_usage": True}
 )
 
+def load_knowledge():
+    knowledge = ""
+    try:
+        with open("knowledge.jsonl", "r") as f:
+            for line in f:
+                data = json.loads(line)
+                knowledge += f"Role: {data['role']} | Info: {data['context']} | Response: {data['response']}\n"
+    except Exception:
+        knowledge = "No extra knowledge found."
+    return knowledge
+
 def coretex_chat(user_input):
-    messages = [{"role": "user", "content": user_input}]
-    # Coretex is thinking... 
-    # Note: On CPU, this might take 10-20 seconds per answer.
-    response = pipe(messages, max_new_tokens=128)
-    return response[0]['generated_text'][-1]['content']
+    kb = load_knowledge()
+    
+    # Keeping the prompt clean so the 1.5B model stays focused
+    prompt = f"You are Coretex. Use this info:\n{kb}\nUser: {user_input}\nCoretex:"
+
+    response = pipe(prompt, max_new_tokens=128, clean_up_tokenization_spaces=True)
+    
+    # Extract only the new part of the text
+    full_text = response[0]['generated_text']
+    answer = full_text.split("Coretex:")[-1].strip()
+    return answer
 
+# The .queue() is the secret to handling "huge limits" 
+# it lines up users so the CPU never crashes
 demo = gr.Interface(
     fn=coretex_chat, 
     inputs=gr.Textbox(label="Message Coretex"), 
     outputs=gr.Textbox(label="Response"),
-    title="Coretex AI System"
+    title="Coretex AI System (v1.5B)"
 )
 
-demo.launch()
+demo.queue().launch()