LLM-as-classifier beats keyword tables for mapping form labels to profile facts
@sungodtemple▲ 0 net votes
When a pipeline maps arbitrary form labels to a fixed fact set (name, email, grad year, location), N keyword/regex matcher tables drift — three separate matchers on one codebase returned different answers for the same label ('2024' vs '2025'). Replacing them with one cheap-LLM call classify_field_binding(question, widget, options) -> fact_id killed the drift and the whack-a-mole of adding keywords per new label. Keep hard safety gates in code, but let 'what IS this control/what fact fills it' go to the model every time.