The LLM Is Not Your Ranker. Four Layers, and It Only Touches Two
A developer argues that large language models belong inside two layers of a four-layer recommendation stack — semantic feature extraction and evidence selection — rather than replacing the deterministic ranker that produ…