What should a PDF parser preserve before the data enters a RAG pipeline?

[ComPDF-Youna]

A PDF parser can return text and still leave important questions unanswered: which page did this block come from, did the table structure survive, and can a reviewer trace a value back to its source position?

This discussion is about what to check before indexing parsed documents for RAG. DocSlight provides local and cloud modes, CLI and Python SDK entry points, and Markdown / JSON outputs. Those interfaces do not by themselves establish the accuracy of a particular document.

For your own sample, compare the parsed output with the original PDF:

• Can you trace important values to the correct source page?
• Are table headers and rows still associated correctly?
• Does the reading order match the document?
• Are skipped or failed pages visible rather than silently omitted?

Installation check — October 30: A clean macOS Intel / Python 3.12 check has not yet reached a successful local parse. The default installation of 0.1.5 entered a PyMuPDF source build. A separate binary-only installation of 0.1.5 could not resolve paddlepaddle>=3.3.0 for that environment. Allowing version fallback installed 0.1.4, whose local parse command then reported missing local dependencies. No parsing accuracy result is being claimed here. Python 3.9 is outside the documented supported range of 3.10–3.13.

Repository and current setup instructions: https://github.com/ComPDFKit/docslight

Which of these checks matters most in your workflow, and what document layout makes it difficult?

For launch context, DocSlight also has a Product Hunt listing: https://www.producthunt.com/products/docslight