Klaus Knowledge · online tier
Simple English Wikipedia as semantic vectors + a wikilink graph, in one sovereign Neon store. Your question is embedded, matched by meaning, and answered by quoting a real sentence with its source — or, when nothing is close enough, by honestly saying so. It never generates prose, so it cannot hallucinate.
How it works. POST /klaus/ask → embed the query (bge-base, 768-d) →
pgvector KNN over the corpus → if the top match clears the similarity floor and a
sentence in it actually overlaps your question, quote that sentence + source; otherwise refuse.
Then the graph offers related topics to walk into via wikilink edges. One Postgres holds
the vectors, the graph, and the dedup metadata — one record, never
duplicated.
Honest limits. Retrieval can miss (ask "atom" and it may refuse, or surface a neighbour like Elsa Einstein for "Einstein") — but a shown answer is always a real sentence from a real article, never invented. The online tier isn't byte-deterministic; the offline lexical tier keeps that guarantee.