Chemical name resolution remains a critical task in scientific data extraction because unresolved or incorrectly matched names directly increase hallucination rates in downstream LLM workflows. This presentation describes our journey from pure vector search using Titan embeddings and Amazon S3 Vectors to a hybrid chemical name resolution system that combines deterministic token matching, semantic retrieval, and LLM-based chemical reasoning. Using real patent-extraction examples, we show where vector similarity performs well, where it fails on subtle chemical differences, and how a hybrid architecture can re-rank candidates, explain structural equivalence, and improve confidence in extracted chemical entities. The presentation will focus on practical lessons for building reliable AI-assisted indexing pipelines (Phoenix) where chemical precision matters more than generic semantic similarity.
Video Stream
