Comparison
Retrieval, re-reading, and retention
Three ways to put a long document in front of a model. The figures below come from our controlled study on the standard legal RAG benchmark; the full methodology is published.
| RAG (retrieval) | Whole-document re-reading | Retention (Revyzor) | |
|---|---|---|---|
| What the model sees | Fragments chosen by a search engine before the model ever reads them | The whole document, read section by section on every question | The whole document, read once; the working state is retained |
| Exact clause accuracy | Baseline. On the longest agreements in our study: 0 of 31 | 3x baseline; 10x on merger agreements | Same accuracy as the structured read: it is the same read, retained |
| Cost per follow-up question | One retrieval round trip, but against the wrong evidence | A full re-read: ~5.7x the wall-clock of a single-shot answer, every time | About 2.6 MB appended per follow-up, and a reload at roughly 7x less GPU time than the re-read; the read is never repeated |
| When the answer is challenged | No record of what the model considered; the chunks are gone | A prompt log at best; the working state was discarded | The retained working state, cryptographically signed, reloadable for inspection |
| Session continuity | None: every question starts from scratch | None: every question starts from scratch | Suspend Friday, resume Monday from the exact same state |
The pattern: whole-document reading is the best clause-finder in the study, and its only real cost is the repeated read. Retention keeps the accuracy and removes the repetition. The retained state then does double duty as the compliance record: signed at capture, reloadable for inspection, held under an explicit retention policy.
How does this compare to Harvey or Legora?
It doesn’t compete with them; it sits underneath that kind of tool. Harvey, Legora, and most legal AI products are applications: a legal workflow and interface built on top of frontier models running in a US cloud, with retrieval deciding which fragments of your documents the model sees. Revyzor is infrastructure: a plug-in for the engine that runs an open-weights model on your own servers, in your own jurisdiction, under your own keys.
The practical difference: with an application, your client documents travel to someone else’s cloud and the model answers from retrieved fragments. With a Revyzor-backed stack, the documents never leave your infrastructure, the model reads them whole, and every read is retained as a signed, inspectable record. A Harvey-style application could, in principle, be built on top of it.