A four-stage pipeline runs the moment a file hits the upload endpoint: extract the text, ask a model for structured metadata, write to a searchable store, surface it for review. Upload returns in under a second; the heavy work runs on a queue worker while the user keeps working. Every supported file type has its own extraction path, each guarded so a missing library degrades gracefully: digital PDFs read directly, Word walks the document tree preserving tables, Excel keeps column alignment, and scanned images go through Tesseract OCR in Romanian and English at once. Extracted text goes to a Chat Completions call with a locked 12-field schema, returning the same fields in the same shape every time, with temperature low so re-uploads produce identical metadata.

A four-stage pipeline runs the moment a file hits the upload endpoint: extract the text, ask a model for structured metadata, write to a searchable store, surface it for review. Upload returns in under a second; the heavy work runs on a queue worker while the user keeps working. Every supported file type has its own extraction path, each guarded so a missing library degrades gracefully: digital PDFs read directly, Word walks the document tree preserving tables, Excel keeps column alignment, and scanned images go through Tesseract OCR in Romanian and English at once. Extracted text goes to a Chat Completions call with a locked 12-field schema, returning the same fields in the same shape every time, with temperature low so re-uploads produce identical metadata.