PDFs, text, markdown, repositories and profile data all become plain text. Scanned images are rejected out loud rather than silently producing nothing.

Every skill, date and project must quote the document word for word. If the quote is not found in the source, the fact is thrown away instead of stored.

The same thing named five ways becomes one node. Spelling is handled by rules; genuine abbreviations are judged by a model, because similarity alone would merge deep learning into machine learning.

Links no single document stated get proposed across the whole graph — and every one carries a sentence explaining itself. A connection that cannot be explained is not shown.