Two June findings from The AI Automators circle the same shift: document AI is moving from parse-to-text toward read-the-pixels. One tames long-doc transcription; one deletes the text step from RAG entirely. Neither is a wholesale "adopt" — both carry a transferable idea.
The common enemy is parser loss — flattening a messy page (tables, diagrams, layout) to markdown destroys content before any model reasons over it. Vision-native reading dodges it. The near-term Arsenal win isn't a model swap; it's a fallback path for fetch/RAG when markdown conversion loses the visual.
Open-weights ~3B MoE (~0.5B active) fine-tuned on DeepSeek-OCR, claiming dozens of pages in one pass with no slowdown. Presenter built a Claude Code demo, confirmed it works — then argued against adopting it.
Clever architecture, wrong production choice. The narrow real win of one-shot is coherence — tables/sentences spanning page breaks don't get sliced. Keep RSWA as prior art for any future long-doc / long-audio transcription; skip the model. Presenter's stack: Docling default → local VLM fallback → Chandra 2 for specialist.
Berkeley / Princeton / Databricks. Skips text conversion entirely: screenshot → tile → embed the images → answer with a VLM. No text anywhere in the pipeline. Indexed 7M Wikipedia pages (~30M tiles, >5TB).
Ships a Claude Code "Pixel Shots" agent skill (3-command install): screenshots a source page via Playwright/Chrome so a multimodal model reads diagrams/tables that web-fetch markdown loses. Demo: web fetch couldn't read a football-formation diagram; the screenshot skill could. Slower than web fetch → fallback only. Directly evaluable for Arsenal's own fetch/parse gaps.
Traditional OCR (Tesseract/PaddleOCR) · structure-aware parsers (Docling) · VLMs — chosen by document type, not replacing each other.
Most RAG failures happen at translation + ranking, not reasoning. Measurable, large, and where vision-native reading pays off.
Pixel methods trade cheap text for expensive ingestion, storage, and per-query vision inference. Dense tiles can still undercut long text.
Give the agent both a visual and a text index and let it route — the pragmatic endpoint both videos land on.