Enterprise document processing faces a trade‑off: running a heavyweight parser on every page is wasteful, while relying only on a fast, cheap parser can miss critical tables or figures. The solution is to start cheap and escalate to a heavier parser only when the page signals that the cheap parse missed the answer.

The article builds an escalation cascade that applies a series of deterministic checks to the cheap parser’s output. Each check—pre‑parsing metadata, parsing‑time outputs that flag flattened tables or opaque figures, retrieval scoring that catches drifting anchors, and generation flags that spot unresolved content—is cheaper and more reliable than the next. If any check flags insufficiency, the pipeline routes that page to a deeper parser; otherwise it stays on the baseline.

Two recurring examples from the Attention paper illustrate the need: Table 3 (a flat‑parsed grid where each cell becomes its own line) and Figure 1 (a diagram that Py Mu PDF returns empty). Free signals such as character count, image count, and a flat‑table fingerprint detect these cases before any expensive parser runs, and the second part of the series walks through the full escalation.

Baseline parsers are Py Mu PDF for native PDFs (≈5 seconds per document, free) and a free OCR engine for scans. A Vision LLM on the same page can cost 10,000 times more and take ten seconds. By applying the cascade on a 200‑page report, the system avoids paying for heavy parsing on pages that are plain text, invoking it only for the few that truly need it.