Chat may not be the best UX for LLMs. A prototype is presented that blends recommendation systems, NLP, and LLMs to let users interact primarily through clicks and filters rather than text. Users browse books, filter by 'vibes', and only resort to minimal chat when needed — with the LLM already having context from past behavior. The backend uses approximate nearest neighbors on item embeddings for retrieval, an LTR model for ranking, pre-cached vibe keywords, and a FastAPI/Jinja app with streaming LLM responses via async Python.
1 Impression