Module 1 · Reading real-world documents · Lesson 2
Parsing: text layers to vision models
How text-layer extraction, OCR, layout-model pipelines and vision-language models each turn a page into text, how their failures differ, what the benchmarks do and don't tell you, and how to route each page to the cheapest parser that is accurate enough.
By the end of this lesson you can
- Explain how text-layer extraction, classical OCR, layout-model pipelines and end-to-end vision-language models work, and the characteristic failure of each
- Compare current tools (Docling, marker, MinerU, olmOCR, PaddleOCR-VL, Mistral OCR, Reducto, LlamaParse, Unstructured and Claude's PDF input) on how they work, licensing and deployment
- Read OmniDocBench, olmOCR-Bench and ParseBench results critically, including what each metric rewards and who built the benchmark
- Compute per-page and per-case parsing cost for several designs and implement a routing policy with escalation