ONE UPLOAD, ONE PIPELINE
1
Upload. A user attaches a file to the conversation — a PDF contract, a DOCX report, a scanned form, a screenshot.
2
Extract & detect. Inside your environment, text is parsed and images go through OCR and visual detection where configured. Detection runs across everything extracted — Elif Aksoy, VF-C-88231, the signature on page four.
3
Review & tokenize. Detected items appear in the review window where configured; confirmed content is tokenized — [PERSON_81af3c], [CID_4a2e97] — and only then does the tokenized content reach the external model.
4
Answer, restored. The model's answer about the document comes back through the restore pipeline — the user reads it with real values in place, in-app.

Documents#

PDF and Office documents are parsed inside your environment. Body text, tables, headers, and embedded images are all inputs to detection — a phone number in a footnote and an ID card photographed into an appendix are treated with the same seriousness as the main text. Tokenization preserves the document's structure, so the model can still summarize, compare, and extract from it usefully, as described in Context Preservation.

Images and scans#

Standalone images and scanned pages go through OCR for embedded text and through visual detection for non-text elements — faces, signatures, identity-document regions, stamps — where those classes are configured. The full modality list and its boundaries live in Visual Detection.

The honest boundaries#

Supported file formats and visual detection classes depend on the configured models and policy — they should be confirmed per deployment rather than assumed. And file handling follows the same conservative failure semantics as everything else: content in governed classes that cannot be processed is held or blocked per policy, never passed through unexamined.