Nace launches Drex for document processing
Nace has added its Drex decision model to its document engine, where it checks every parsed page against the source image, and says the combination now tops the company's own parsing leaderboard.
Nace announced Drex for Document Processing, which adds its Drex decision model to the company's document engine. Drex makes narrow checks: whether an OCR line matches the page, whether a table kept its structure, and whether a ticked box survived the scan.
The pipeline has three steps. The engine first converts any of 30 file types into layout-aware Markdown. Drex then compares that output with the page image, and an agentic correction step fixes the places where the two disagree. If a correction comes back empty or much shorter than the original, the original reading is kept. Nace says each Drex check returns a calibrated probability in about 50 milliseconds and costs about $2 per million answers. That lets the engine save its more expensive reading for pages where something looks wrong.
On Nace's Parse Index, the engine with Drex scored 0.834 in an 11 November run. The engine alone scored 0.820, as did LlamaParse Agentic, and GPT-6 Astra (low) scored 0.813. Nace lists its own cost at $7.00 per 1,000 pages. There are two caveats: the leaderboard is Nace's own, and the other parsers were not re-run, so their scores come from a 15 September run.
Nace reports these results for the Drex version:
- 95.7 on OmniDocBench, 5.4 points ahead of the next-best parser
- 91.9 on MDPBench, which tests phone photos of documents, 8.3 points ahead
- Gains over the engine alone on 3 of 6 benchmarks, and results within noise on 2 more
- Page corrections 28–38% faster than rewriting every page
Drex for Document Processing is available through Nace's Console, a REST API, a CLI, and Python and TypeScript SDKs. Nace did not publish separate pricing for it.