Introduction
A document retrieval engine with no chunking, no embeddings and no vector database.
Status (2026-10-11)
The hosted Vectorless service is closing. The code is open source and no longer actively developed. These docs describe how the engine works and how to run it yourself. Contact: halleluyaholudele@gmail.com.
Vectorless is a retrieval engine written in Go. It ships as a single binary and runs on Postgres plus object storage. It does not chunk documents, compute embeddings or use a vector database.
It reads a document's own table of contents, places every entry on its page, and answers a question by selecting the sections and then the pages that hold the answer. Those pages come back verbatim. A generative model writes an answer over them, and each claim in the answer is pinned to a line on its page.
Every selection in that pipeline is made by Jev, TypeSafe's System One model. Jev returns a probability or a choice and writes no prose. The only generated text in the pipeline is the final answer.
The rule
If a decision can be written as a closed set of options, it goes to the judge. Select, don't generate.
What it measured
On FinanceBench (40 questions over 10-K filings, k=5, run on 2026-09-21), the evidence was among the returned pages for 0.925 of questions, against 0.375 for BGE-small embeddings and 0.200 for BM25. That is retrieval, not answer accuracy. The Benchmarks page has every figure with its date, its report and what it does not show.
Where to go next
How it works
Ingest, navigation, verbatim evidence pages, pincites and replay.
Jev and System One models
What Jev decides, what it costs, and what runs without a TypeSafe key.
Benchmarks
Dated figures linked to the engine's evaluation reports, with their caveats.
Getting Started
Run an engine locally and ask your first question.
API Reference
The HTTP routes the engine serves.
SDKs
Clients for TypeScript, Python and Go.