The no-chunking model
Why Vectorless has no chunks, no embeddings and no vector database, and what it costs instead.
No chunking
Fixed-size chunks cut a document mid-sentence and mid-table. Vectorless keeps the document's own sections and returns whole pages. The unit of evidence is a page of the PDF, verbatim, with its page number.
No embeddings
An embedding is computed at ingest, before any question exists, and relevance is a distance between two vectors. Vectorless asks Jev, at query time, whether this section or this page answers this question. The decision is conditioned on the question.
The difference was measured. On FinanceBench at k=5, the evidence was among Vectorless's returned pages for 0.925 of questions, against 0.375 for BGE-small embeddings (2026-09-21). Embeddings could not even serve as a pre-filter: at k=50 they held every gold page for only 0.875 of questions. See Benchmarks.
No vector database
The engine runs on Postgres plus object storage. There are no embedding jobs, no index rebuilds and nothing to re-embed when you change the answer model.
The trade-offs
| Vector RAG | Vectorless | |
|---|---|---|
| Retrieval unit | fixed-size chunk | a page of the document, verbatim |
| Relevance | vector distance, computed at ingest | Jev's decision, at query time |
| Infrastructure | vector database and embedding pipeline | Postgres and object storage |
| Citations | chunk-level | line, page or section (pincites) |
| Query latency | milliseconds (49 ms in our run) | seconds (1.9 s navigation in a benchmark tool; about 3.5–6 s for a full answer from the production API) |
| Cost per query | none beyond compute | about $0.004 of Jev calls, plus the answer model |
| External dependency | the embedding model | Jev, a hosted TypeSafe API |
| Determinism | deterministic | not deterministic; replay returns a stored answer |
Vectorless trades milliseconds for seconds and a hosted dependency for evidence that is found more often and cited more exactly. Whether that trade is right depends on the documents and the questions. Multi-hop questions have not been measured.