v2-refactor-temp/docs/breaking-changes/2026-08-07-knowledge-document-processor-pdf-only.md
A knowledge base's selected document processor (MinerU, Doc2X, Mistral, PaddleOCR,
self-hosted MinerU) is now applied to .pdf files only. Word, PowerPoint and Excel
files (.doc, .docx, .pptx, .xls, .xlsx) added to a knowledge base are read
straight into the index by the built-in reader instead of being sent through the
processor first. They are still accepted — what changes is which component extracts
their text.
The processor and the built-in reader were always alternatives, never both: once a
processor produced its Markdown, indexing read that file instead of the original. So
this is a change of extraction engine, not the removal of a redundant pass. Tables
and layout may be formatted differently from before for these formats. In exchange,
adding an Office document is faster and leaves no sibling .md artifact next to the
source; for the hosted processors (MinerU, Doc2X, Mistral, PaddleOCR) it also spends
no API quota and no longer uploads the document to a third-party service. Files
already indexed keep their existing artifact and are not reprocessed.
Windows on ARM64 users: check any .pptx, .xls or .xlsx you keep in a
knowledge base. The built-in reader ships no ARM64 Windows binary, so on that
platform those three formats fall back to a plain-text reader that produces unusable
output — and unlike before, selecting a document processor no longer routes around
it. Convert those files to PDF and re-add them; PDFs still go through the chosen
processor. .doc and .docx are unaffected (they have working fallback readers).
On every other platform (Windows x64, macOS, Linux) nothing is required. Users who preferred a specific processor's rendering of an Office document can convert it to PDF before adding it, which routes it through the processor as before.
Known regression, Windows on ARM64 only, affecting .pptx / .xls / .xlsx: those
users previously had a working path via a document processor and no longer do. Worth
calling out separately if the release note has a platform-specific section.
Two scoping notes so this is not over-reported:
.doc / .docx are not affected on any platform — they fall back to dedicated
readers, not the plain-text one..ppt is not part of this change. It was never routed through a document
processor (it is absent from the pre-change extension list), so it neither gained
nor lost a processor route here. Its Windows-ARM64 behaviour is pre-existing.