Keyword search misses the paper you need
Two papers can describe the same reaction with different vocabulary. Researchers end up running many searches and reading abstracts by hand.
Structure is the real query
Chemists often know the molecule, not the words. A text box alone is the wrong interface.
Batch ingestion
Papers are parsed, cleaned, split into passages, and stored with their metadata. New papers are added in batches without rebuilding the query path.
Hybrid retrieval
bge-m3 embeddings in a FAISS index, combined with keyword matching and a reranking step, served through FastAPI.
Cited answers
A language model writes a short answer from the retrieved passages. Each claim links back to the paper it came from.
No copying of source text
Answers are checked for long verbatim overlap with the source passages and rewritten if they copy too much, so the tool summarises rather than republishes.
Structure search
A molecule editor in the browser lets you draw a structure and search for papers that mention it.
Evaluation
A small evaluation harness checks retrieval quality and whether answers stay faithful to the cited passages.
Code is private
Chemister is not open source. I'm happy to walk through the design and a live demo.