Documentation

The document compiler, documented.

nautris lowers any document (PDF, Office, HTML, images, scans, e-mail) into one canonical, deterministic, lossless representation, the CDOM, and generates every output from that single tree. These pages cover how it works, how to run it, and how to talk to it.

On this page

What it is

nautris is to documents what a compiler is to source code. Instead of every parser reinventing structure and every output drifting from the last, all inputs are compiled into a single source of truth, the Canonical Document Object Model: a flat tree of typed blocks joined by relationship edges, preserving hierarchy, tables, images, reading order, metadata, and page coordinates. Each output format (JSON, RAG chunks, knowledge graph, LLM context, fine-tune JSONL) is just another exporter reading from that tree.

The core path is generative-model-free: deterministic passes, no LLM in the loop, and the same bytes always compile to the same tree. Everything runs inside your own infrastructure.

Performance is measured, not promised: the published benchmarks report real throughput and accuracy numbers, including 100% scores on tables, reading order, and headings.

Architecture

Readers lower each format into the CDOM; deterministic passes clean and normalize it; exporters generate artifacts from it. Adding a format is a plugin, never an N×M rewrite:

every format in, one canonical tree, every output an exporter

Deployment topology

One image, two roles: the API enqueues, worker containers extract, and everything shares one volume. The perimeter is yours; nothing crosses it outbound:

the API runs read-only; workers run unprivileged with resource caps

Where to next