Aleph Alpha distills its parsing harness into Parsy, a single-pass model

Why Aleph Alpha built a parser

Aleph Alpha says its forward-deployed AI solutions hit a performance ceiling in customer projects, and the cause was poor data ingestion quality. Search systems, assistants and extraction pipelines depend on content parsed from datasheets, maintenance manuals, official forms, progress reports, test reports and specifications. Those documents contain nested tables, footnotes, charts and scanned pages.

When answer quality fell short, the company traced most errors to the parsing step. Complex tables were a typical case: the parsers recovered every word and number but lost which value belonged to which unit and which row. Other parsers invented characters that were never on the page or attached footnotes to the wrong passages.

Customers also require that documents stay inside their own environment, so any solution has to be self-hosted, on-premises or in a private cloud. Aleph Alpha says no tool it tested offered both the quality and the sovereignty it needed.

A harness that checks output against the page

The company built an automated harness that mimics a careful human reviewer placing extracted output beside the original page. It rests on three ideas.

• Position is kept. Each page is rendered once, regions, categories and reading order are established, and every extracted element carries a bounding box so the page can be rebuilt at its original coordinates.
• The original is the reference. The reconstruction is screenshotted and compared with the original by a vision model, and a judge model scores each comparison. Pages with low similarity are re-extracted and re-checked, and uncertain pages go to an optional human reviewer. Aleph Alpha says this needs no labeled data.
• Human review is targeted. Every page has a quality score and a note, so reviewers open only pages that need attention. The review interface is also available as an API.

On ParseBench, a public benchmark of 2,078 human-verified pages from real enterprise documents, the harness scored 86.0 in Aleph Alpha's internal evaluation. The company says that would have ranked first on the public leaderboard until the end of August 2026.

The problem was speed and cost. Every page goes through the judge at least once, and difficult pages loop several times, using several models and up to three passes. For customers processing millions of pages a month on their own hardware, that means more GPUs.

Training the student

Aleph Alpha decided to train a model with the single goal of matching the harness in one pass. It says it did not use customer data. Instead it collected public documents, in line with legal requirements, that resemble customer material, from technical datasheets to scanned forms. The harness labeled this corpus, pairing each element with its category, bounding box and reading order position. A specialized vision-language model was trained on that format.

Because a model trained on parser output inherits that parser's systematic errors, the team checked labels after each round with tools that do not share the harness's blind spots:

• an independent layout model whose regions are compared with the labels;
• an ink-coverage check that finds printed text outside every labeled box;
• a second OCR model whose disagreements get flagged.

People sampled the flagged pages to confirm what each detector catches before the team fixed or removed that failure type across the corpus. In one example, labels in long lists and tables had merged several printed rows into one box, and the model learned the same habit until a row-comparison check found and repaired those pages.

The first model version scored about two points below the harness on ParseBench. In the sixth round of training and label cleaning it passed the harness, and it has stayed above it since. Aleph Alpha says recent rounds gain little because the remaining errors are mostly formatting conventions rather than mistakes. It attributes the result to the harness reaching its quality at run time, while the model learns from the cleaned output, so most mistakes never reach its training data.

What it means for customers

According to Aleph Alpha, Parsy costs one to two orders of magnitude less per page than the harness. It now powers extraction work in the company's customer solutions and drives source highlighting in its chat applications. For organizations with very large document volumes and strict self-hosting requirements, the company presents it as a way to keep high parsing quality without the GPU load of a multi-pass system. The figures come from Aleph Alpha's own evaluation.