OCR for images and scanned PDF pages
OCR is independent of visual labels. You can leaveintelligence.enabled off
while extracting text. Google Cloud Vision is the supported OCR provider; the
Moondream and Rekognition image-label adapters do not become document OCR engines.
media-studio:search-extract as above. JPEG, PNG, WebP and AVIF are supported;
AVIF is temporarily decoded into a provider-compatible JPEG when needed. Image OCR
reads the managed original under the text-extraction source cap; it does not use
the separate visual-tagging rendition selector. It never uploads a publicly
accessible source URL. Source dimensions, bytes and UTF-8 provider output are
validated. Provider responses request text without per-character geometry, with a
256KiB response ceiling. Only the bounded text projection is retained.
PDF processing first extracts native text. Only textless pages are rasterized for
OCR, one page at a time. Pages with native text do not incur OCR requests. A blank
page can still produce an empty OCR result. The page cap, character budget and
elapsed-time admission budget can leave the document partially indexed. The last
admitted page may finish after the admission deadline, within its own decoder and
HTTP timeouts. No full-document coverage is claimed for limited runs.
Access/current-version checks run before each OCR page request and again before
publication. A completed current-version extraction with the same processing
profile is reused without a second OCR call. Changing queue routing does not
invalidate that result. Changing processing limits does. Private temporary files
are closed and removed after success or failure. Raw OCR payloads, page images and
all-language copies are not stored.
Run one explicit extraction
media-content queue. Existing/imported files are not automatically
backfilled. Failed or ambiguous paid operations are not retried automatically.