real PubMed abstracts. Every result carries a PMID, so nothing here has to
be taken on trust.
Typed questions use keyword matching (a browser cannot run an embedding model). The two
sections below replay the embedding search this project actually ships.
You are the judge
The tool returns a pool, not an answer. Eight candidates for a real question — pick the one you
think the question came from.
Where it fails
Three questions this index gets wrong. Nothing hidden: the right abstract genuinely is not in
the five it returned.