Update app.py

app.py

@@ -18,23 +18,27 @@ from langchain_community.embeddings import HuggingFaceEmbeddings
 # -------------------------
 
 # You can change this to any instruct model you like
-HF_MODEL_ID = "mistralai/Mistral-7B-Instruct-v0.2"
+HF_MODEL_ID = "ibm-granite/granite-3.1-8b-instruct"
+
+# Option B: Llama 3.3 (Current industry standard)
+# HF_MODEL_ID = "meta-llama/Llama-3.3-70B-Instruct"
 
 # If you use a private token, set it as HF_TOKEN env var in Spaces
 client = InferenceClient(HF_MODEL_ID, token=os.getenv("HF_TOKEN"))
 
 
 def get_llm():
-    """Simple callable LLM wrapper compatible with LangChain's RetrievalQA."""
+    token = os.getenv("HF_TOKEN")
 
     def _llm(prompt: str) -> str:
+        # We add wait_for_model to wake up the API if it's sleeping
         response = client.text_generation(
             prompt,
-            max_new_tokens=256,
-            temperature=0.5,
+            max_new_tokens=512,
+            temperature=0.1,
+            wait_for_model=True 
         )
         return response
-
     return _llm