Update app.py

app.py

@@ -1,32 +1,32 @@
 from fastapi import FastAPI, Request, HTTPException, Depends
 from fastapi.middleware.cors import CORSMiddleware
+from fastapi.responses import JSONResponse
 from llama_cpp import Llama
 import uvicorn
+import os
 
 app = FastAPI()
 
-# 1. ALLOW EXTERNAL CONNECTIONS (CORS)
-# This allows your GitHub Pages or local HTML file to talk to the HF Space
+# 1. CORS CONFIGURATION
+# This is crucial so your website can actually read the response
 app.add_middleware(
     CORSMiddleware,
-    allow_origins=["*"],
+    allow_origins=["*"], 
     allow_credentials=True,
     allow_methods=["*"],
     allow_headers=["*"],
 )
 
-# Configuration
 MY_API_KEY = "my-secret-key-456"
-MODEL_PATH = "./model.gguf"
 
-# 2. INITIALIZE MISTRAL (Optimized for CPU)
-# We use 4 threads to utilize the full power of the HF Free Tier CPU
+# 2. LOAD MISTRAL (Optimized for HF CPU)
+# We use 4 threads and a higher batch size for faster 'first thought'
 llm = Llama(
-    model_path=MODEL_PATH,
-    n_ctx=2048,           # Context window
-    n_threads=4,          # Parallel CPU cores
-    n_batch=512,          # Prompt processing speed
-    verbose=True          # Logs performance to the HF console
+    model_path="./model.gguf", 
+    n_ctx=2048, 
+    n_threads=4,
+    n_batch=512,
+    verbose=True
 )
 
 def verify_key(request: Request):
@@ -36,59 +36,55 @@ def verify_key(request: Request):
 
 @app.get("/")
 def home():
-    return {
-        "status": "Online",
-        "engine": "Mistral-7B-Q8_0",
-        "brand": "ChatMPT Official"
-    }
+    return {"status": "Online", "mode": "Mistral-7B-Q8", "brand": "ChatMPT"}
 
 @app.post("/v1/chat")
 async def chat(request: Request, _ = Depends(verify_key)):
     try:
         body = await request.json()
-        user_input = body.get("prompt", "")
+        user_input = body.get("prompt", "").strip()
         
         if not user_input:
-            return {"reply": "System: No prompt received."}
+            return JSONResponse(content={"reply": "Please enter a message!"})
 
-        # 3. MISTRAL INSTRUCT FORMAT
-        # llama-cpp-python adds the <s> automatically. 
-        # [INST] tags guide the model's logic.
-        system_msg = "You are ChatMPT, a professional AI. Be concise and never repeat your own sentences."
+        # --- MISTRAL INSTRUCT FORMAT ---
+        # llama-cpp adds <s> automatically, so we just wrap in [INST]
+        system_msg = "You are ChatMPT, a helpful AI. Be direct and concise."
         prompt = f"[INST] {system_msg}\n\n{user_input} [/INST]"
 
-        # 4. ANTI-LOOP & SPEED SAMPLING
+        print(f"\n--- USER PROMPT: {user_input} ---")
+
+        # --- GENERATION SETTINGS ---
+        # We lower repeat_penalty slightly so it doesn't get 'scared' to talk
         response = llm(
-            prompt,
-            max_tokens=256,
-            stop=["</s>", "[INST]"], # Hard stops for Mistral
-            temperature=0.75,         # Balanced creativity
-            
-            # --- THE "ANTI-LOOP" ARMOR ---
-            repeat_penalty=1.25,     # Heavily discourages repeating phrases
-            frequency_penalty=0.5,   # Discourages using the same word too often
-            presence_penalty=0.5,    # Discourages returning to the same topics
-            
-            # --- MIROSTAT 2.0 (High Quality Control) ---
-            mirostat_mode=2,         # 2 is the modern standard for CPU stability
-            mirostat_tau=5.0,        # Higher = more interesting, Lower = more focused
-            mirostat_eta=0.1         # Learning rate for the AI's "internal mood"
+            prompt, 
+            max_tokens=512, 
+            stop=["</s>", "[INST]", "[/INST]"], 
+            temperature=0.7,
+            repeat_penalty=1.1, 
+            mirostat_mode=2
         )
 
         final_reply = response["choices"][0]["text"].strip()
 
-        # 5. BRAND PROTECTION (Safety Filter)
-        # Force-corrects common hallucinations
-        hallucinations = ["ChatGPT", "Chat GPT", "OpenAI", "ChatPapers", "ChatPBT"]
+        # --- DEBUG LOGGING ---
+        # This will show up in your Hugging Face Logs
+        print(f"--- AI REPLY: {final_reply} ---")
+
+        # --- FALLBACK IF EMPTY ---
+        if not final_reply or len(final_reply) < 2:
+            final_reply = "I'm here! Could you please rephrase that?"
+
+        # Identity Safety
+        hallucinations = ["ChatGPT", "Chat GPT", "OpenAI", "ChatPapers"]
         for item in hallucinations:
             final_reply = final_reply.replace(item, "ChatMPT")
 
-        return {"reply": final_reply}
+        return JSONResponse(content={"reply": final_reply})
 
     except Exception as e:
-        print(f"Error: {str(e)}")
-        return {"reply": f"System Error: The engine is under heavy load. Please try again."}
+        print(f"SERVER ERROR: {str(e)}")
+        return JSONResponse(status_code=500, content={"reply": "System Error: Brain is overloaded."})
 
 if __name__ == "__main__":
-    # Standard Hugging Face port
     uvicorn.run(app, host="0.0.0.0", port=7860)