Rag clean up + missing specs

Address Sourcery feedback: fix CUDA memory check and remove unused quantization config

Use HF_TOKEN from environment for authenticated downloads

Add src/models module for HuggingFace model loading