Increase token limits for SmolLM2-360M: better quality answers

Complete Ollama removal: Switch to local transformers with SmolLM2-135M for CPU inference

Switch to meta-llama/Llama-3.2-1B-Instruct - available via HF Inference API

Fix HuggingFace model name format: use HuggingFaceTB/SmolLM2-360M instead of Ollama format

Fix HuggingFace Spaces compatibility: standardize InferenceClient API and improve backend auto-detection

Make web search truly optional with ENABLE_WEB_SEARCH flag and graceful fallback

Switch to SmolLM2-1.7B for CPU-friendly inference on HF Spaces

Add HuggingFace Inference API as LLM backend with auto-detection

Remove invalid gradio_demo import from rag_utils

Fix duckduckgo_search import for HF Spaces compatibility

Implement code structure updates and remove redundant sections