Use Groq provider for Llama (free tier)

backend/rag_service.py

@@ -9,8 +9,8 @@ from backend.chat_service import save_message, load_chat
 from backend.retrieval_service import retrieve_chunks
 
 MAX_HISTORY_MESSAGES = 20
-# Llama 3.2 3B - try :hf-inference to use HF's serverless (free tier)
-DEFAULT_MODEL = "meta-llama/Llama-3.2-3B-Instruct:hf-inference"
+# Groq has free tier - Llama 3.2 3B. Or use :fastest for auto provider.
+DEFAULT_MODEL = "meta-llama/Llama-3.2-3B-Instruct:groq"
 TOP_K = 5
 
 _client: OpenAI | None = None