Release overview — December 2025
The Curiosity Team
We closed 2025 by making Curiosity Workspace see more and show its work: images become searchable by what they depict, AI chat replies split reasoning and tool calls from the answer, and administrators decide who can use each AI tool. Here are the December highlights.
Images you can search
- Captioning and OCR are now separate controls. On-device OCR reads text in images and scanned PDFs, and a new on-device AI visual model describes what an image contains. Turning on captioning re-indexes images and text-free PDFs.
- Force OCR on a single file, even when OCR is off workspace-wide.
- PDFs keep more of their images when rendering: the size threshold above which images are dropped is substantially higher.
A clearer AI chat
- Structured replies: reasoning, tool calls and the final answer render as separate blocks, across the supported model providers.
- A single state-aware send and stop button prevents sending again while a reply is still streaming.
- Failed tool calls are marked clearly, and source snippets appear only under the message they belong to.
AI tools under control
- Access control for AI tools: restrict a tool to specific access groups, and only their members are offered it.
- AI tools accept array parameters, and the tool code editor offers completions tailored to tool authoring.
- AI search and embeddings settings move to their own AI Indexing page.
For administrators
- Custom entity post-processing: add code per entity type to merge, modify or ignore extracted entities, from a new Post Process Entities page.
- Synchronized monitoring charts: zooming or panning one chart applies the same time range to the others, with a shared crosshair.
- Full connector run logs can be downloaded, and the integrations view reports total elapsed time per run.
- Linux deployments monitor the root drive and heartbeat delay.
"Local development without a license file"
A built-in evaluation license now unlocks the full feature set for local development and evaluation instances, with a recurring non-production reminder.
Safer extraction
File content extraction can run inside a Landlock sandbox on Linux, restricted to the folders it needs to read and its temporary working directory. If sandboxing is unavailable, extraction falls back to the normal path.
The little things
- A View JSON command opens a result's full node content in a read-only viewer.
- File uploads show live progress again, and a failed file can be retried.
- Image previews block right-click saving when downloads are disabled.
- Expanded AI interface translations in six languages.
This overview covers builds v25.12.63066 → v25.12.63342. For the full, build-by-build detail see the changelog.
Read next
Articles on context graphs, enterprise search and industrial AI
