Skip to main content

PDF documents

Kapa converts PDF documents from your knowledge sources into structured content so they can be found during retrieval and used in answers. This page describes how PDF processing works across all connectors that support it.

What Kapa extracts​

From each document, Kapa extracts:

  • The full text content
  • The document structure, preserving the heading hierarchy
  • Tables
  • Code blocks and equations, preserved as text
  • Lists, including nested lists
  • Hyperlinks, preserved as links
  • Images, such as figures, diagrams, and screenshots (see Image indexing)

Page headers, page footers, and tables of contents are excluded, since they add noise without carrying useful information.

Scanned documents​

Documents with an embedded text layer give the best results. Scanned pages are processed with optical character recognition, which can be less accurate, especially for complex layouts or low-quality scans.

Supported sources​

The following connectors ingest PDF documents:

PDF processing works the same way regardless of which connector a document comes from.

When a PDF document has a URL, citations link to the specific page in the document using a page anchor (for example https://example.com/manual.pdf#page=42), which browsers open directly at that page.

Whether a document has a URL depends on the connector:

  • Web Crawling: crawled PDFs are cited by the URL they were found at
  • File Upload: provide a URL in the optional Linked URL field when uploading
  • S3 Storage: map files to URLs with an index.json file
  • Google Drive: documents are cited by their Google Drive link

Kapa does not serve the document itself. A citation is a regular link, so users need access to wherever the document is hosted to open it. Images extracted from the document are the exception: Kapa serves indexed PDF images itself, so they display in answers even when the document lives in a private location.

Limits​

Documents with more than 7,000 pages are skipped.