updated

app.py

@@ -44,23 +44,37 @@ import json
 from backend.services.codingo_chatbot import get_response as _codingo_get_response
 
 
-def get_chatbot_response(query: str) -> str:
-    """Proxy to the codingo_chatbot implementation.
-
-    This function exists to preserve the original public API of
-    ``app.get_chatbot_response`` while redirecting calls to the new
-    implementation.  It catches any exceptions and returns a user
-    friendly message, ensuring the Flask route never raises.
-    """
-    try:
-        return _codingo_get_response(query)
-    except Exception as exc:
-        print(f"Chatbot error: {exc}", file=sys.stderr)
-        return (
-            "I'm having trouble processing your request. Please try again or ask "
-            "about Codingo's features, job matching, or how to use the platform."
+def get_chatbot_response(user_input: str) -> str:
+    from llama_cpp import Llama
+
+    # Load model once
+    global llm
+    if 'llm' not in globals():
+        llm = Llama(
+            model_path="/tmp/llama_models/tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf",
+            n_ctx=2048,
+            n_threads=8,
+            n_gpu_layers=20  # GPU acceleration if available
         )
 
+    # Format prompt in TinyLlama's chat style
+    prompt = f"<|system|>\nYou are LUNA, a helpful assistant for the Codingo recruitment platform.\n<|user|>\n{user_input}\n<|assistant|>\n"
+
+    # Generate response with safe parameters
+    output = llm(
+        prompt,
+        max_tokens=256,
+        temperature=0.3,     # lower temperature for stability
+        top_p=0.9,
+        repeat_penalty=1.1,
+        stop=["</s>"]
+    )
+
+    reply = output["choices"][0]["text"].strip()
+    if not reply:
+        reply = "I'm here to help you with Codingo. Could you please rephrase your question?"
+    return reply
+
 # Initialize Flask app
 app = Flask(
     __name__,