Scanned books

A PDF that is scans of paper carries no text. Ask the agent to read it and the pages are read here, on your machine: three small models — one that divides a page into its parts, one that finds the lines in a part, one that reads what a line says. Nothing is uploaded and no account is involved.

What comes out is kept in the vault’s own .numen/ folder, named after the bytes of the document, so a book that moves keeps its reading and two copies of one book share it.

The models are fetched the first time a reading is asked for, and the corner of the window shows how far that has got.

Everything here is indexing.recognition in the settings file.

What is worth changing

Almost nothing: every number was measured, and the defaults are where the measurement said to put them. These are the three that come up.

page.dpi	what a page is rendered at before it is read. 300. Raising it reads finer print and costs time.
page.threads	how many threads one model may use. 4.
detect.expand	how many pixels a found line is widened by before it is read. 18. The line a detector finds is drawn inside the letters, so without widening the top of every capital and the last letter of every line are cut away.

regions.body says which parts of a page carry what the document says, and regions.head which of them open a section — where a name stands in that list is how deep the section sits. A part the model finds that body does not name is not read at all, which is how page numbers and running heads stay out of your search.

dir names a folder holding the models, and download turned off keeps numen from fetching anything — the two together are how a machine with no network reads a book. runtime names the ONNX Runtime library, for a machine whose own is the one to use.

Putting a reading right

A reading can be corrected afterwards by a model that speaks the ordinary chat request. Nothing does this by default: name nothing under indexing.proofreading and a reading is used exactly as it was read, with no key and no network.

{
  "indexing": {
    "proofreading": {
      "use": "service",
      "service": {
        "base_url": "https://openrouter.ai/api/v1",
        "batch_url": "https://openrouter.ai/api/beta/batches",
        "name": "google/gemini-2.5-flash",
        "key_env": "NUMEN_PROOFREADING_KEY",
        "pages_at_once": 40,
        "letters_apart": 0.30
      }
    }
  }
}

batch_url leaves the pages in a queue and collects them later at half the price. A batch outlives the run that left it, so one left before you closed the application is collected the next time you open it. Leave it empty and pages are asked for one at a time, and waited for.

letters_apart is the safety catch: how far a correction may move a line’s letters and still count as a correction. Over 931 corrections of one book, every correction further apart than 0.30 was damage — text dragged in from the next line, or one corrected word standing in for a whole line — and everything below it was a genuine fix. A correction further apart than this is dropped, and that line is left as it was read.

The complete list is on every setting.