SDS Knowledge Base -- POC
Ask a question about one of 3 sample Safety Data Sheets (Acetone / AK Scientific, Camphor-10-sulfonic acid / Sigma-Aldrich, Acid:Phenol:Chloroform / Thermo Fisher) by catalog number, CAS number, or product name.
This is a proof of concept for a design meant to scale to an 80,000-document SDS catalog: parse each SDS into its 16 canonical GHS sections, resolve a question to (product, section) via exact identifier + keyword-topic routing when possible (deterministic, e.g. "disposal" -> Section 13, no embeddings involved), and fall back to semantic search only for open-ended questions.
Try the example questions below, or ask your own -- e.g. "what PPE do I need for product 21365" or "disposal considerations for X5851".
Note on the retrieval model: this build uses local TF-IDF vectors rather than a neural embedding model (a network-access constraint in the build environment, not a design choice) -- see the main project README for the production swap-in path to real sentence embeddings + a proper vector index (sqlite-vec / pgvector) at full catalog scale.