---
title: "Discussion on \"Docling Drops Pages and Footnotes Without a Word. Three Numbers Catch It.\" | Hashnode"
description: "Discussion on \"Docling Drops Pages and Footnotes Without a Word. Three Numbers Catch It.\". A docling-parse threading race returns fewer pages with no error, and table footnotes never reach chunks. Three numbers catch both, and the fix for each."
url: "https://hashnode.com/posts/docling-drops-pages-and-footnotes-without-a-word-three-numbers-catch-it/6a9f7beb132656abfddc160b"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Docling Drops Pages and Footnotes Without a Word. Three Numbers Catch It.](https://lijibuilds.hashnode.dev/docling-drops-pages-and-footnotes-without-a-word-three-numbers-catch-it?utm_source=hashnode&utm_medium=feed)

A docling-parse threading race returns fewer pages with no error, and table footnotes never reach chunks. Three numbers catch both, and the fix for each.

Sep 8 · 13 min read · lijibuilds.hashnode.dev  
Read the article

-   [#docling](https://hashnode.com/tag/docling)
-   [#rag](https://hashnode.com/tag/rag)
-   [#pdf](https://hashnode.com/tag/pdf)
-   [#document-parsing](https://hashnode.com/tag/document-parsing)
-   [#debugging](https://hashnode.com/tag/debugging)

## Written by

[Liji Alex](https://hashnode.com/@lijia)

@lijia · Hyderabad, India

  
[View profile](https://hashnode.com/@lijia)

Engineering AI Systems

## Responses 1

Markdown supported · ⌘↵ to post

[Ahmet Özel](https://hashnode.com/@ahmetozel)

[Sep 11](https://hashnode.com/posts/docling-drops-pages-and-footnotes-without-a-word-three-numbers-catch-it/6a9f7beb132656abfddc160b/comment/6aa39afa417493cb6a584da8)

Silent partial parses are the worst class of RAG bug, because every downstream check still passes. The tree looks complete, the vector store fills, recall on your eval set looks fine, and the only signal arrives months later as a question that returns nothing. In a hospital corpus that is worse still, since the missing item is exactly the rule nobody asks about until they need it.

The check that saved me was blunt: run a second, dumber extractor and compare page count and total character count, then alert on the delta rather than trying to validate the tree itself. Two parsers disagreeing is a far easier signal than one parser being wrong. Curious which of your three numbers you would keep if you could only afford to run one.