Skip to content

Document Processing

Uploaded documents pass through text extraction and then structure extraction before they can be used to generate questions.

TypeExtensionsHow text is obtained
PDF.pdfEmbedded text is read first; pages without usable text are read with optical character recognition
Word.docxText is read directly from the document
PowerPoint.pptxText is read directly from the slides
Plain text.txtRead as-is
Image.png, .jpg, .jpegOptical character recognition

A PDF whose pages are scans of paper is handled by character recognition rather than by asking an AI model to read the images. Recognition is configured for Urdu, Arabic, and English by default; other languages depend on the language packs installed for your deployment.

Text extraction and structure extraction are tracked separately, so a document can have finished one and not the other.

Text extractionMeaning
noneNot started
queuedWaiting for a worker
processingBeing read now
doneText is available
failedCould not be read
Structure extractionMeaning
noneNot started
queuedWaiting for a worker
processingBeing analysed now
readyA structure tree is available
failedCould not be produced

Large documents are routed to a separate, deliberately slower processing lane so that one long textbook cannot block every short document in the workspace.

Structure extraction produces a tree of entries — chapters, sections, and topics — each with an optional summary and page range. You can edit any entry’s title, summary, and page range.

Entry originMeaning
AIProduced by structure extraction
EditedChanged by a person after extraction

Each regeneration snapshots the previous tree, so an earlier version can be restored.

Once text extraction finishes, document text is split into chunks and indexed so the platform can find passages by meaning rather than by exact wording. This index is what grounds question generation and the AI assistant in your own material.