Three real-world documents parsed end-to-end by the production pipeline — pick one and inspect the raw output next to the source. No signup, no credit card.
DocUnwrap Engineering Team
DocUnwrap Technical Report Series · TR-26-114 · September 2026
This report describes the parsing pipeline behind DocUnwrap. Every page is treated as an independent unit: a router inspects the page image, assigns a model matched to its degradation profile, and attaches a confidence score to the result. Pages that fall below the confidence threshold are escalated to a stronger model before the caller ever sees them. We summarize the default routing policy, the page-level review flags, and the operational rules that keep the pipeline predictable under load.
Not all pages are equally hard. A text-native PDF can be read from its embedded text layer in milliseconds, while a skewed phone photo may need a full vision-language pass. A single-model stack therefore overpays for easy pages and still underperforms on hard ones. DocUnwrap treats routing as the first-class decision, and every downstream field hangs off it.
Table 1: Default routing policy
| Page condition | Route | Notes |
|---|---|---|
| Text-native PDF | Text layer | No GPU inference, near-instant |
| Clean scan | paddleocr-vl | High-accuracy layout model |
| Photo or heavy skew | teleocr | Degradation-tuned re-parse |
| Confidence below bar | glm-ocr | Escalated before delivery |
Callers receive clean Markdown with the heading hierarchy, lists and tables intact, or a per-page JSON record that carries the page text, its confidence, the model that produced it and a needsReview flag — everything a review queue or a RAG indexer needs to act on its own.
Create an account and parse your first 100 pages free — upload a PDF or call the API.