8. Enrichment (Settings → Ingest enrichment)#

All off by default; each is opt-in per workspace:

• Canonical language: translate incoming content at ingest (original is preserved); queries in any language find it.
• Image processing (rescue/full): pages/scans that yield no text get sent to the vision model. rescue = only pages that would otherwise contribute nothing; full = also describe images on normal pages. Required for scanned-image uploads.
• Deep page rendering: for PDFs whose pages even image processing couldn't read (vector drawings, fax scans, JPEG2000), a badge appears on the document. With this toggle on, a newly-uploaded flagged document is deep-processed automatically as part of ingest — no button needed. For a document flagged before you turned the toggle on, a Deep process control on the row lets you re-upload the file, which processes it the same way. Either way, it rasterizes just the skipped pages and processes them page-by-page in a separate, cost-isolated pipeline; rescued pages become searchable with "Page N" citations. Requires Anthropic BYOM. Zero data retention: files are processed and discarded, never stored a second time.

BYOM providers (Settings → Model key): OpenAI (one key does chat + embeddings) or Anthropic (required for all vision features; add a separate OpenAI key for embeddings — the form guides you). Keys are stored encrypted and never displayed again.