Private context images

Images provide visual context for a key. Members can view images attached to keys they can read. Owners, Admins and Editors with shared-content write permission can upload and edit them. Writing at least one active language grants shared-content permission; assignment to a task or knowledge resource does not.

Upload and review

1. In Translations, open the key’s Context section. Save or cancel other context changes first.
2. Choose Image file, enter meaningful Image alternative text, then Upload image.
3. Wait for processing. The worker decodes and re-encodes the file as PNG, removing original metadata. A successful HTTP upload means queued work, not a ready image.
4. Choose Enlarge image to inspect the result. Edit Alternative text or choose Remove from context, then Save context to commit that context change.

PNG, JPEG and WebP are accepted: one frame, at most 10 MiB and 16 megapixels. The sanitized PNG must also fit within 10 MiB. Each key supports at most ten images. SVG, animated images and external URL downloads are unsupported. Images are private: authenticated requests return sanitized bytes with private, no-store; there is no public image URL. Removal changes the reference, not an irreversible erasure of history.

Storage budget

The default total is 1 GiB per workspace, shared by every project, branch and retained AI preview. Workspace storage displays stored bytes, reserved bytes and the current limit. Each pending upload reserves 20 MiB for the original and sanitized files. Once processing succeeds, only the sanitized size remains charged. The same file referenced by several keys, forks, snapshots or previews is charged once.

Removing a live reference does not free a file still needed by a snapshot, fork or valid AI preview. Unreferenced files are eligible for bounded cleanup after a one-hour grace period. Abandoned uploads retain their reservation for up to one hour; cleanup and processing require a running worker. Failed file removal stays charged until cleanup succeeds. A paused worker can extend these times.

Customer Owners and Admins cannot change the limit. Ask the system operator to use the audited quota command. Lowering a limit never removes existing files; it prevents new reservations until usage fits again.

Explicit AI image selection

Only the configured OpenAI and Gemini image-capable models accept image context. Images start unselected, even when text context is selected. In the free assistance preview, open Provider context, inspect the exact image and enable Image (explicit opt-in, high resolution). Check the revised usage estimate before approving Generate. The selection sends sanitized image bytes and alternative text to the chosen provider; it does not send private storage paths or workspace identifiers.

The preview freezes the image hash, size, dimensions and alternative text. Generate, the paid-call boundary and Apply recheck current access and key revisions. Removing or changing context requires a new preview. Expired previews cannot generate or apply. Image input contributes a model-specific conservative token reservation; it is not a price quote or a promise that lower resolution costs less. Requests are limited to 18 MB of estimated inline image payload, within the provider’s total request bound. Reduce the selected images if one cell exceeds this limit; nothing is silently cropped.

Google Basic v2 and DeepL do not receive image input. Generated text remains a suggestion until explicit Apply. See assistance recovery for uncertain paid outcomes.

Errors and recovery

• Storage is full: wait for eligible cleanup or request an operator limit change. Retained snapshots and active previews may legitimately keep the bytes in use.
• Processing failed: verify file type, frame count, input/output size and dimensions. Also check whether the key changed or your write permission was removed while queued.
• Conflict: your draft stays available. Refresh current context and compare before saving or uploading again; stale uploads cannot overwrite newer context.
• Upload outcome unclear: refresh its status. A repeated reservation uses the same request identity; it does not allocate another file. A failed upload requires choosing the file again after inspection.

Native translation file formats do not carry these private binary assets. Exporting keys with image context returns an explicit compatibility error. The common encrypted recovery archive protects the private image store and its database references together.