Retrieval that selects,not generates.
No chunking. No embeddings. No vector DB. Vectorless reads a document's own contents, selects the pages that answer a question, returns them verbatim and pins each claim in the answer to a line on its page.
// 1. Install
npm i vectorless
// 2. Evidence pages for a question
const { sections } = await client.query(doc.id, "FY2018 capex?")Contents placed, pages selected, claims pinned.
Every choice in the pipeline is a pick from a closed set, made by a System One model that writes no prose. The only generated text is the answer.
Select, don't generate
The document's own contents
Ingest finds the contents page, places each entry on its real page and splits long sections. Jev, a System One model, makes each decision; ingest makes no generative call.
Sections, then pages
Jev ranks the sections against the question, skims the candidate pages, reads the best ones and follows "see Note N". The evidence pages come back verbatim.
Pincites
Each claim in the answer is pinned to a line on its page. Jev picks the quote from the page's own text, so no generated text can reach a citation.
No vector DB to run
One Go binary on Postgres and object storage. No chunks, no embeddings, no index. Jev is a hosted TypeSafe API; without a key the engine falls back to a generative model.
If a decision can be written as a closed set of options, it goes to the judge.