Release overview — June 2025

The Curiosity Team

• 2025-06-30

June was about the AI assistant and the search box: a reworked chat in Curiosity Workspace, better retrieval behind it, and a search box that remembers what you look for. Four releases landed this month.

A reworked AI chat

• Overhauled chat view: a spinner while a message is in flight, a clear error if it fails, and a chat list that refreshes once the reply arrives. Switching conversations no longer flashes a blocking progress bar.
• Better context retrieval: improved retrieval and system prompts for more relevant answers. Conversations continue even when no supporting context is found, and older chats still surface their referenced files.
• Answers in your language: answers, summaries and key points follow each user's preferred language, and generated chat titles follow the chat's.
• GPT-4.1 is available as a chat model. The o1 and o1-mini options were retired, and existing chats fall back to a supported model.

Steadier conversations

• Clearer messages when the provider is unconfigured, a request is rate limited, the context grows too long or the service can't be reached, with guidance on what to do next.
• Streamed answers recover from partial and oversized chunks and reconnect instead of breaking the live update.
• If reading a file takes too long, the assistant says it timed out, and extraction failures give a more specific error.

A search box that remembers

• Autocomplete from your history: suggestions come from your own previous searches, ranked by how often you ran them. They appear when you click into an empty search box, and a setting turns them off.
• Preview pane: a resizable side preview next to the results lets you inspect an item without leaving the result set.
• Clearer facets: active filters are highlighted, showing the selected value or a compact "+N" count. A type facet now flags when results of another type were grouped under it.
• Header commands on search results, and global commands that appear everywhere.

Better matching

• Accent-insensitive file search: "resume" now matches "résumé".
• N-gram matching across an entire indexed thread improves recall on partial-word queries over file content.
• Better handling of abbreviations and full file paths, and a scoring fix so vector similarity matches rank correctly.
• Searches run on their own thread pool, so heavy queries are less likely to block other work.

For administrators

• Customize the assistant's chat box text, button labels, tooltips and placeholders, and turn off onboarding tips for a workspace.
• The server logs a warning when its TLS certificate is within two months of expiring, and a startup setting names the workspace before the graph loads.
• Custom OpenAI-compatible hosts given as a full URL are used as-is.
• The CLI honors system proxy settings, and users it syncs are active on creation.

The little things

• Summaries are no longer forced to be overly short.
• On-demand extraction runs even when background extraction is paused.
• PDF text detection and Markdown conversion run under a time limit.
• Recurring Google Calendar events import more accurately.
• More interface strings translated into German, Spanish, French, Italian, Portuguese and Chinese.

This overview covers builds v25.6.58642 → v25.7.59324. For the full, build-by-build detail see the changelog.