Every chat completion now records token counts, latency and finish reason, and a new admin dashboard surfaces usage over time with token, request and cost charts, a per-model summary and a recent-calls table. Administrators can register chat models, set per-million-token prices and toggle cost calculation, with price changes tracked over time so historical usage keeps the cost that was in effect when each call was made. Reference prices refresh automatically every day from the public Curiosity pricing reference, so the catalog stays current without a rebuild.

Opening a chat now shows what it is chatting about, with an auto-opened preview, a header breadcrumb for the selected chat and a tidied sidebar, making it easier to see and navigate the references behind a conversation.

The assistant now opens with a clear working-mode choice (read-only is pre-selected) presented as selectable cards in the empty state, replaces it with a denser, easier-to-scan chat layout, and streams replies more reliably across freshly started, resumed or server-driven turns. Stopping a turn now also cancels the work in progress on the server, and tool calls only show an expandable view when there is something to inspect.

The model catalog and pricing editor now show each provider's logo instead of a raw provider code, with a fallback for unrecognized providers.

The admin Tasks page now separates your own code tasks (with a New Task button and onboarding for an empty list) from the built-in task types. Custom code tasks, previously not reachable from this page, are now accessible here.

The task editor now has a Run Task button next to History, so an admin can trigger an immediate run without leaving the editor.

A new read-only page under Operate, Observability lists nodes scheduled for delayed deletion, ordered by deletion time with remaining-time labels and paging.

The inspect view now opens agent runs as well as chats, with full run detail. Administrators' diagnostic code snippets are saved per user and searchable, so they persist across devices instead of living only in the browser.

Assistants can now list the references attached to a chat, read the text of a reference (including a specific page of a file), and search or list the items inside an attached folder, space or saved search, all access-checked against the chat user. This is enabled automatically when a chat has attached context.

Chat and agent messages can now carry blob attachments, sent to multimodal models as images or PDF document parts. Files can also be rendered on demand into high-resolution page images (single page or a page range) for use as attachments, with results cached so repeat renders are instant.

Nodes can be scheduled for deletion after a delay rather than removed immediately, with due deletions dispatched automatically in the background and the option to cancel a pending deletion.

The Box data connector now runs on Box's current SDK, refreshing access tokens through Curiosity's OAuth authentication and handling folder listing, events, downloads and user info on the new client.

Curiosity's built-in embedding model now runs as a fully managed encoder with no native runtime dependency, downloading a single weights file. Some unused model options were removed; existing indexes continue to work unchanged.

A new built-in tool lets an agent stop mid-turn and present a multiple-choice question right in the chat, rendered as an inline questionnaire. Your answers are fed back into the conversation so the agent can continue with your input, and an agent can ask several questions at once and collect every answer in one step.

AI Search now supports multiple embedding models side by side, including new built-in and third-party embedding models. A new Vector Indexing settings page lets you create and manage indexes pinned to a specific model, with each index showing its model and whether chunking is enabled. A compact built-in model is the default for new AI Search indexes, offering multilingual retrieval with a smaller download.

When the workspace runs as a Docker container, administrators can now see image builds newer than the running one and update to any of them from a new Software Updates page under Operate, Runtime. The update performs a self-replacing cutover, preserves operator-supplied environment variables and Docker network aliases, and lets the new image's built-in defaults take effect.

The search box now understands inline filters such as @file, @webpage, filetype:, ext:, and src:, turning them into chips that scope your results. These work on both the home search and the results view.

The View Documentation panel gained a sticky search box that filters the listed signatures and descriptions as you type, with multi-term matching, so long documentation lists are far easier to scan.

The data hub node schema list now scrolls reliably with the keyboard and loads instance and edge counts only as rows come into view, so large schema lists stay responsive.

Refined search box styling across the app, and the PDF viewer now shows the correct tooltip for its precise and extended matching modes.

Endpoints and AI tools that fail to compile are now flagged with a red bug icon in the admin Build tree, and the flag updates as a tool starts failing or recovers. A new toggle filters the tree down to just the failing items, and your search term and filter are reapplied when the list refreshes.

The build interface now shows the exceptions thrown by code endpoints and AI tools in paginated Errors accordions alongside usage metrics, including the error message and details for each error, with a matching View Errors button in the AI tool editor.

The active tab across the admin views is now reflected in the page URL, so tabs can be deep-linked and browser back and forward navigate between them.

A round of improvements to live updates addresses message delivery and timeout edge cases and bounds the event queue under pressure, making real-time connections more reliable.

Node label definitions now support an interpolated syntax using and placeholders, with standard .NET numeric, date, and custom format specifiers plus string-case transforms, alongside the existing fallback-list labels.

Queries gain StartAtSimilar methods that seed a query at one or more nodes and immediately expand to their nearest neighbors in a single call.

A new self-tuning rate limiter keeps the workspace process at or below a target CPU usage by smoothly throttling task admission, smoothing out load spikes.

The number of threads used for AI Search embeddings is now a single workspace-wide setting (defaulting to half the available CPU cores) that can be viewed and changed from the AI Search settings page, replacing the previous per-index option.

Removing a user's access to a shared item now also revokes leftover admin relationships, so unsharing fully removes the user's ownership and admin links.

The Curiosity SDK gains generic, strongly typed overloads mirroring the existing node story, including ClearPermissions, Link and Unlink, RestrictAccessToUser and RestrictAccessToTeam, alias management, and a ToNode resolver, plus symmetric remove operations for revoking a single owner, removing a single alias, and demoting a team admin.

Workspace-definition imports that report errors now surface the real import result instead of failing with an opaque error, and the response helpers guard against missing messages.

The import-workspace-definitions command now prints the import's success, errors, and warnings when the import reports problems.

A new in-product AI assistant lets a system administrator design, change and operate the workspace by conversation. It can introspect every configurable surface (endpoints, AI tools, agents, scheduled tasks, indexes, schemas, search settings, access groups), propose concrete changes as a reviewable plan, test them, and apply them only after the admin approves. The assistant is always available from the admin console, follows the surface the admin currently has open, and shows the proposed changes as a clear diff before anything is committed.

The search experience across the home page, search, similar-search, embedded search and neighbors views now uses a richer search box with the same query syntax, a result count, save and clear controls, and inline templates.

Search results gain an opt-in "select all results" mode. Instead of acting only on the rows currently in view, you can run a command (such as adding to a space) against every matching result, with a progress dialog and the option to cancel.

The sign-in, invitation and password-reset forms now use proper form semantics so browser and extension password managers reliably offer to save, generate and autofill credentials scoped to the correct site.

When a chat answer involves a tool call, the conversation now shows a readable breakdown of the tool's inputs, its result (rendered as prose or a formatted data block), and any error message if the call failed.

The unified Build admin view gains a Skills section for creating, editing and deleting skills, with a name, description, category, icon and a markdown body with live preview. Built-in skills are listed read-only, and custom skills are now included in workspace configuration export and import. The Admin Agent can read and propose skills as well.

Tool-call activity in the admin assistant chat no longer disappears the moment a tool finishes; completed steps remain on screen with a done indicator so you can see what the assistant actually did.

Per-signal weights now genuinely affect ranking under reciprocal-rank fusion, and previously inert options take effect: limiting candidates per signal and overall, tracking per-stage timings, and reporting progress. A new option filters results to what a given user is allowed to see.

The query API for integrators gains cancellation support on its async enumeration, single-element terminals (First, Single and their default variants), paging and reversing, additional type and relationship overloads, and async composition helpers, making query pipelines easier to write and cancel cleanly.

Custom OpenAI-compatible hosts now have their API style detected automatically from the host URL, so self-hosted servers that only implement the chat-completions endpoint stop failing on every request. Structured (typed) responses are also passed through correctly on these hosts.

Selecting newer OpenAI models, including GPT-5.2 and the GPT-5 nano and mini variants, now uses the chosen model instead of falling back to a different one.

The built-in AI tools now return well-formed JSON across the board, making their output reliable to parse and chain. Tool output that wraps JSON in a markdown code fence is also unwrapped correctly.

Loading a large graph is significantly faster: per-node lookups during load were replaced with a single streaming pass and the per-type index build was parallelized.

The first day of recording (and any day following a gap with no data) is now measured from the moment the workspace first came online that day, instead of being penalized for the hours before recording started.

Generating a connector from a CSV file now produces code that compiles and runs in many more cases, with correct key and timestamp typing, better header detection, culture-invariant type inference, and recognition of long integers and yes/no boolean columns.

Built-in skills, AI tools and agents are now reliably seeded and updated on every workspace, so newly added or revised built-ins appear after an upgrade rather than only on brand-new workspaces.

Internal node types (skills, agents, agent runs and code history) are now excluded from search results by default.

A new agent capability lets administrators define reusable agents (name, description, system prompt, icon, model, attached tools and output schema), run them against a chat AI provider, and review each run as a self-contained record with its prompts, results, status and timing. The agent editor offers an icon picker, a searchable tools grid, and a Variables panel that detects ${VAR} placeholders in the system prompt so you can supply values for a test run. Agents are included in workspace export and import, and old agent runs are cleaned up under the same retention policy as AI chats.

New Endpoint Access and AI Tool Access views under Identity and Access let an administrator review and change the access model of endpoints and AI tools, including applying a configuration to several selected items at once, without recompiling code.

A new diagnostics view under Operate lets administrators capture performance diagnostics from the running server to help investigate performance issues, with results available to download.

Endpoints can now run in a sandboxed scope with read-only graph access that blocks write transactions and refuses to call endpoints not marked read-only. The endpoint editor warns at save time when an endpoint is flagged read-only but its code contains write operations, and the Shell view gained a Sandbox mode toggle.

A command palette (Ctrl/Cmd+P) fuzzy-searches every endpoint, AI tool, prompt, code task, index and integration from one overlay. Open tabs can be cycled with Ctrl+Tab, are remembered in the page URL and restored when you return, and the endpoint and agent editors can open related items as tabs directly. Editors and metrics now show a clear loading spinner while they load.

A more complete C# formatter (whitespace, line endings, final newline and import organization) backs the format action in the admin code editors and the format API.

The indexes list and other long admin lists are now paginated and only refresh status for items currently on screen, keeping these pages responsive in large workspaces.

The admin sidebar gained a dedicated Back to app button in its footer.

The unused graph dump download settings page and its endpoints were removed.

A search consisting only of a negated term (for example NOT fuel) now correctly removes parents whose children contain the excluded term when related results are merged.

Search result merging now handles files owned by other files, so attachments nested under another file appear in results.

A new /health/replica endpoint reports per-replica write-ahead-log lag for monitoring replication status.

Looking up nodes by type and key now applies the same per-type normalization (such as lower-casing and trimming) that the strongly-typed helpers use, so identifiers resolve consistently across the query API.

Idle keep-alive handling on the live event and websocket streams was improved so the UI shows a steady connection state while an active stream is still delivering data.

You can now attach your own CSS classes to search-result cards and node preview panels, making it easier to brand or fine-tune the look of these surfaces without overriding the default styling.

The admin area has a new navigation rail that groups everything into five clear sections (Home, Operate, Build, Govern, Configure), each with a matching landing page and a focused sidebar. Admin pages now live under stable, predictable addresses, the entry point is a dedicated "Workspace Admin" item, and a new profile menu collects preferences, account, theme, and logout in one place. Creating endpoints, AI tools, prompts, code tasks, indexes, and integrations now opens as an editor tab that turns into the new item's tab on save, with an unsaved-changes marker on the tab while you edit.

Admins can now write a description for each node style from the schema style settings, and that description is carried through as documentation for the matching node in endpoint code, so it shows up as helpful context while writing code against the schema.

PDFs whose embedded text decodes incorrectly (broken or custom font encodings) are now detected and automatically routed through OCR instead, so these documents become searchable rather than indexing meaningless text. Sensitivity is configurable to avoid misfiring on very short files.

A new helper lets AI tools and scheduled tasks turn any file into Markdown with a single call, transparently running OCR for scanned PDFs and images.

New methods add a node (or add-or-update it) together with its owner references atomically, so a node and the teams or users that own it are established in the same operation.

User and team creation calls can now opt into caching so repeated calls with the same inputs reuse the in-flight or completed result instead of contacting the server again.

Workspaces can now run with replicas that stream changes from a primary over the network, keeping standby copies in sync. Replication includes file content, validates transferred data, reports replica lag, and adapts how often replicas catch up to balance freshness against load.

Press Ctrl+B (Cmd+B on macOS) anywhere in the workspace to show or hide the navigation sidebar.

The pin icon now appears filled in when an item is pinned, making it easier to see at a glance what you have pinned.

The date and time facet in search results renders a single Apply control.

Administrators can configure request rate limits per user, with the settings stored in the workspace and a current-usage view to monitor activity. The usage panel now shows a friendly placeholder when no data has been collected yet.

In the endpoint editor, the test button is now disabled while there are unsaved changes, so you do not accidentally test an out-of-date version of an endpoint.

Edge changes on nodes with more than 128 connections, added or removed after their first save, are now persisted reliably across restarts.

Channel file folders that sit at the root of a document library are now handled correctly during Teams synchronization, avoiding misplaced or skipped files.

When an attachment shares a name with an existing one, the newer attachment now replaces the older entry, matching Confluence's own behavior instead of duplicating it.

A new endpoint returns the document for a given file page, letting the UI fetch and display an individual page directly.

A new low-memory mode lets server deployments opt into the same memory-conservative tuning the desktop app uses, independent of the desktop context.

Text indexes can now report aggregated keyword statistics, with unsupported index types handled gracefully.

The query SDK gains relationship traversal and filtering by edge type (related-to / not-related-to via a given edge), string-based property filters, timestamp filters, and a way to start a query at nodes of a given type identified by their keys.

Storing the authentication token now works on macOS when the configuration lives in a hidden dot-directory.

Workspaces can now generate sentence embeddings from an external provider (such as OpenAI, Anthropic, Cohere, Google, or Azure OpenAI) instead of only the built-in local model. When creating an embeddings index an administrator can pick the external model type and supply the provider, endpoint, model name, and API key, and the workspace then uses that model to power similarity and semantic search. Already-encoded text is cached so unchanged content is not re-sent to the provider on every update, keeping indexing fast and reducing API cost.

API keys for connected AI providers (Anthropic, Mistral, and others) are no longer returned in settings responses or shown in the admin configuration screens, so saved credentials cannot be read back after they are entered.

The schema view now defaults to the visual graph tab, giving a clearer overview at a glance before switching to the code or JSON representations.

Users and Teams now have their own dedicated views with consistent filtering, so links and in-app navigation land on the right screen and preserve the selected filter.

When the workspace fails to start because the host does not support cgroup memory limits, the error report now explains how to resolve it through a configuration setting.

The "set your name" prompt now appears only when a name is genuinely missing, pre-fills any existing values, and no longer reappears after you have already completed it in the same session.

The PDF viewer can now default to either precise or extended (broader) match highlighting, with a toggle on the viewer to switch between the two while reading.

Code editors throughout the workspace now use the Monaspace font family for a cleaner, more legible look.

Workspaces can now route static file requests by hostname, so multiple custom domains pointed at one workspace can each serve their own content, with per-folder access rules that restrict which referrers may load the files.

The admin sidebar now keeps a short list of recently visited settings pages and lets you pin the ones you use most so they stay at the top.

Users and Teams are now combined into a single Access Management view with a tabbed layout, and the Data, Search, NLP, Languages, Interface, Audit, Migration, Indexes, Endpoints, and Active Queries screens were rebuilt around a consistent segmented-tab layout for easier navigation. Languages now lives under the NLP section, and data import and schema viewing tools are grouped together under the Data section.

Deleting a user now shows a proper confirmation dialog before the account is removed.

The code editor in the admin console handles its font configuration more reliably.

A new fluent similarity engine lets developers build "find items like this" queries directly against the graph, combining multiple scored signals and filters, with support for rule explanations so you can see why a result matched.

New graph operations let you create users and teams, add or remove users from teams, and restrict access to specific users or teams programmatically.

Password strength now counts any non-alphanumeric character as a symbol, so a wider range of special characters satisfies the complexity requirement.

Upgrading an existing workspace now migrates its underlying storage more reliably and reports its internal metrics without erroring when the database is still starting up.

Renaming a chat now loads the latest name from the server before opening the dialog, so the field reflects the chat's current title even if it was renamed elsewhere.

The workspace now stores its graph, blob, and index data in a single consolidated database instead of several separate ones. Existing deployments are migrated automatically on first startup after upgrading, with the previous databases preserved until migration completes.

Cached sign-in data is now periodically cleared of expired entries, keeping memory steady on long-running workspaces.

Operators can enable additional storage diagnostics at startup to help investigate storage issues.

The SDK gains a method to retrieve the schema of a given node type, returning its name, key field, and the list of fields with their types (including list and dictionary types). A matching public endpoint backs it.

When you open a node preview from the AI chat, the most relevant actions now appear as buttons directly in the preview's title bar instead of being tucked away behind the menu, so common operations are one click away.

After importing or changing graph definitions, the workspace now refreshes its cached endpoints, entry points, and AI tools, so the updated code runs right away instead of an earlier compiled version persisting until a restart.

The assistant can now remember facts you state across conversations. Explicitly stated facts are extracted from your chats and stored as personal memories, and the assistant uses them in later chats. A new preferences screen lets you review, edit, enable, disable, or delete individual memories, and turn the whole feature off. Memory is kept tidy automatically: when it grows past a limit, older entries are consolidated into fewer combined facts.

You can now edit one of your earlier messages in place and regenerate the assistant's response from that revised text, without retyping the whole conversation.

A new menu in the chat content preview header surfaces the same node actions available elsewhere, so you can act on the previewed item directly.

Clicking New Chat while the current chat is already empty no longer creates a redundant duplicate.

The AI prompt template editor no longer pops up code-style autocomplete suggestions, since the content is meant to be written as plain English.

Resolved several issues around the chat archive view, loading archived chats, and the selected assistant not being applied when starting a new chat.

Improved symbol resolution for hover information, smoother scrolling, and more consistent timing for showing errors and documentation in the code editor.

A redesigned audit view browses audit logs as files, similar to the logs browser, with search and filtering by user, node, and audit type so administrators can pinpoint relevant activity faster.

Configuration changes are now recorded in their own dedicated audit log, making it easier to track exactly which configuration files changed and what action was taken, independently of other audit activity.

The admin usage dashboard now tracks messages sent to the model and the number of AI tool invocations, and surfaces active user rate and average messages per user.

AI tools can now be monitored individually, with execution metrics recorded and shown on each tool's settings, matching the format used for endpoints.

Graph search queries and facet generation now enforce per-node-type and per-field access permissions, so users only see results and facets they are allowed to access. Node-type access can be configured and is included in import and export.

Spotter model exceptions can now be unlinked directly from the model editor.

Corrected several system-admin endpoints to require write authorization when they change data or state (including the AI tool test endpoint), and changed ownership-check failures to return a clear "forbidden" response rather than "not found".

Importing workspace definitions now shows a saving indicator while the import runs.

Custom code can now read a node content field as a typed object directly from its stored JSON, with parsed results cached to avoid repeated work.

Default parameters for the embeddings (HNSW) index were tuned for improved behavior out of the box.

Errors are now linked to the most recent user message, and related error records are cleaned up when a message is removed, so error context stays accurate. Cancelled generations no longer leave behind a stray output message.

Audio files now play through a dedicated player with a visual waveform, scrubbing, and playback controls, replacing the bare browser audio element for a clearer listening experience.

Several sidebar entries now behave as direct links, so they open in a new tab and respond more naturally to clicks.

The Manage area now opens on a new landing page that lists every administration section as searchable cards, so admins can find the right settings page by typing instead of scanning the sidebar.

Administrators can now browse and search log files stored on disk directly from the workspace, including opening a full log file, with results ordered for easier reading.

The built-in code editor now shows richer, faster hover information with generated documentation for types and members, and handles HTML content and unusual type names more gracefully.

The AI prompt template editor now uses the full code editor with Markdown support instead of a plain text box, giving syntax highlighting and a consistent editing experience.

Pattern spotters are now always ordered after other recognizers in a pipeline, so their matches apply on top of model output as expected.

System administrators can now create and delete AI assistant templates they do not personally own, instead of being blocked when they are not the owner.

Custom code and shell endpoints can now return a file response with a name and content type, so a script can hand back a generated file for direct download.

The graph now supports exporting nodes (with configurable edge handling) and importing them back, laying the groundwork for moving graph data between workspaces.

Invalid edge counts are now recovered automatically by resetting the affected edge count, with clearer error reporting, improving resilience of the graph store.

Memory use is reduced and bounded across chat, embedding indexes, and background training, so long-lived workspaces stay leaner over time.

AI chats can now be archived and reviewed later from a dedicated archived chats view, where each chat can be previewed, unarchived to bring it back into the active list, or deleted. Send and stop button states are now tracked per chat rather than shared across all chats.

Generated documents and slide decks inside an AI chat can now be downloaded or exported as PDF directly from the chat, with rich markdown rendered in the conversation. Exported files use the title from the document's markdown.

Live previews inside the AI chat view are more reliable, including fixes for missing headers and for sticky chat commands that previously stopped working.

Onboarding tips now behave more sensibly: introductory tips can still appear when tips are disabled, and tips no longer pop up over embedded modal content.

Renaming a chat now reflects the latest chat state immediately, and archived chats are listed in the correct order based on when they were archived.

The prompt template editor now lets you try a template against a chosen AI provider with a sample input message and see the response in place, so templates can be validated before use.

Entity linking settings now provide two distinct search boxes, one for entities and one for node types and fields, making it easier to find what you need when configuring data linking.

Exported workspace definitions now carry accurate last modified timestamps, and the comparison view can be sorted alphabetically or by last updated, with clear indicators showing whether the external or internal version is newer.

A new setting lets administrators prevent users from deleting AI chats.

Commands that control the host machine (open external URLs, shut down, restart, log off, lock, hibernate, sleep, empty the recycle bin) now only run in the desktop application and are rejected elsewhere.

Custom code endpoints, shell scripts, and scheduled tasks can now run graph queries in the context of a specific user, so results respect that user's access permissions.

Cancelling an AI chat now cleanly stops the response, name generation, and summary generation, and chat message content (including thinking, tool calls, and tool results) is preserved correctly. Large background saves stream to disk instead of loading entire payloads into memory.

Tightened file handling, preview security, and input validation across the workspace.

Code execution now reports errors with accurate line numbers and hides the internal scaffolding that wraps your script, so error messages point at the lines you actually wrote. Errors surface inline in the shell rather than as disruptive pop-up alerts.

Save actions across editors, modals, and settings panels now use a standard save button and toast that show "Saving", "Saved", and "Error" states, giving reliable feedback for both manual saves and background operations.

The logs view now renders correctly when a log message spans multiple lines, and very long log messages keep the live connection stable.

A new admin view shows the queries currently running against the workspace, with per minute metrics for executed queries, nodes and edges touched, and run times over a rolling window. You can sort by most recent, longest running, or most nodes/edges touched, enable additional per query diagnostics, and hide very short queries (under half a second) to focus on the slow ones.

Testing an endpoint or AI tool now shows a richer response preview with status, size, and timing, and uses shared, faster JSON formatting. Code execution results in the shell use the same preview panel.

The entity linking settings page is now an interactive grid that maps entity types to targets for data parsing and search query parsing, with configuration modals per cell and bulk save. You can quickly toggle the default link on or off for a cell with shift-click without opening the configuration dialog.

AI prompt template edits are now tracked in history and can be reviewed from a History button. Import and export now use dedicated handling that downloads templates as a zip of code files and imports them back, instead of incorrectly reusing the AI tools flow.

The About view now shows whether the instance is a replica (with its name and identifier) or the primary server, and logs no longer include entries from inactive replicas.

The command search now shows admin specific commands when you are on admin pages.

A new endpoint estimates the token count for chat AI requests, letting clients gauge prompt size before sending.

Chat conversations are now organized through per user, per provider access groups that align with the standard access control model, with existing conversations migrated automatically. This resolves chats that had stopped working after the access control changes.

Code endpoints that recompile no longer block one another, so endpoints stay responsive under load.

Corrected fuzziness handling in search and improved endpoint matching when an endpoint is registered under a sub path.

Uploading CSV files through the updated file upload now works correctly.

Reading large objects from cloud storage now buffers to a disk cache instead of holding everything in memory, reducing memory pressure when streaming big files.

Tightened file handling, preview security, and input validation across the workspace, including rate limiting on identicon generation and an ownership check so users can only access their own saved search queries.

Exporting and importing workspace definitions now supports zipped archives, making it easier to move full configurations between workspaces.

Individual fields and node types can now be restricted per user, and the workspace enforces these permissions everywhere node data is read back through the API. Users only see the fields they are allowed to see, while administrators retain full visibility.

AI tools can now be scoped to specific users or access groups. Tools a user is not permitted to use are hidden from the assistant's tool list and excluded when running a chat prompt, so each person only works with the tools available to them.

Large dropdown lists throughout the workspace are now searchable, so you can type to filter instead of scrolling to find an option.

Archiving a chat now fades and collapses the row before it disappears, giving clearer visual feedback.

The assistant chat can now show additional per-message action buttons and supports customizing what the copy button does, with the action buttons staying in view as you scroll a long message.

The split preview panel can now open at a chosen width rather than always taking half the screen.

Deleting a schema in Data Hub now uses a two-step confirmation that lists the consequences and requires typing the exact schema name before the delete can proceed.

Audit log files are now written with cryptographic signatures and rolling files, making it possible to detect if log history has been altered after the fact.

The administration sidebar now includes a search box to quickly find configuration sections.

Custom code endpoints can now return an explicit HTTP status code (and optional payload and content type) instead of always returning a default success response.

Custom code can now generate and compare text embeddings directly through a new Graph embeddings API, including similarity scoring, reusing the workspace's sentence encoder.

AI tools that return structured (non-string) results now expose a cached JSON schema describing their output, helping the assistant produce correctly shaped responses.

The AI indexing status now reports the embedding runtime state and which embedding models are loaded, making it easier to confirm semantic search is ready.

Creating, updating, deleting, or importing code endpoints, AI tools, and prompt templates now reliably refreshes the generated code cache, so changes take effect immediately, including deletions.

Generated code documentation now renders the comments for the specific method overload instead of mixing in documentation from other overloads.

You can now archive a conversation from the chat list context menu to keep your active chats focused. Archived chats are hidden from the main list while remaining stored.

Chat responses now support resending in-flight chunks, so a streaming answer that is interrupted by a network hiccup or reconnect recovers cleanly instead of leaving a partial or unfinished message.

The document viewer now has a rotate button, so sideways scans and landscape pages can be turned to a readable orientation.

Search facets you are actively filtering on now sort to the top of the facet list, making it easier to see and adjust the filters currently in effect.

Inline snippet references generated by the assistant are now recognised in more forms and render as proper links to the source material.

Automatically generated chat titles are now produced in the language of the chat rather than defaulting to another language.

Expanding a tool call in a chat answer now reliably shows its result every time, and a stray box shadow on chat entries was removed for a cleaner conversation view.

Editors for code indexes, endpoints, scheduled tasks, AI tools, custom AI providers and entity post-processing now disable the save button while the code is invalid, and the unsaved-changes confirmation no longer appears after a successful save or lets you close past genuinely unsaved work.

Hovering a symbol in the C# code editors now shows formatted documentation for that symbol, served by a dedicated hover lookup instead of the previous workaround.

Diagnostic squiggles, completions and hovers now report the correct line numbers for code that runs against the workspace schema, instead of being offset by the helper code appended behind the scenes.

Administrators can enable additional diagnostics for file content extraction to investigate problem files, and the option to skip a non-processing file now appears after a few seconds instead of several minutes.

A maintenance task re-indexes existing .xlsm files so their content becomes searchable on workspaces that previously skipped them.

Custom AI tools can now accept structured class parameters (not just simple values), and tool definitions now meet the stricter schema requirements of the latest OpenAI APIs, with the newer response API used only when the selected model supports it.

Initializing code indexes is now more dependable at startup, and code index configuration and indexing are more reliable.

Known error conditions when posting chat messages are now handled cleanly instead of surfacing as raw failures.

The AI assistant can now remember information about you across conversations. It can store details you ask it to keep, recall relevant memories when answering, list what it has stored, and forget specific entries on request, with each user's memories kept private to them.

A new option lets you pin the AI chat history panel so it stays visible instead of collapsing, making it easier to jump between past conversations.

The node view's export action now produces a CSV file instead of an Excel/XLS download, giving you a lighter, more portable format that opens anywhere.

A new "Code Integrations" settings area lets admins configure custom data-connector tasks alongside the existing scheduled tasks, including importing and exporting their definitions.

Admins can now use an in-place AI chat assistant directly inside the endpoint, index, and other code editors to draft and revise code, with a side-by-side diff of the proposed changes before applying them.

Admins can define named prompt templates with variables, test them in place, and reference them from AI tools, with full import and export support so templates move between workspaces.

When configuring an AI Assistant template, admins can now choose exactly which tools the assistant is allowed to use.

A new background maintenance task uses the configured AI provider to review stored abbreviations, flag likely errors or noise, and clean them up automatically on workspaces that have abbreviation data.

Importing large CSV files now streams the upload to the server, so big files import reliably without overwhelming the browser, with clearer error messages when something goes wrong.

Nodes can now carry an associated summary that is surfaced in their data, providing a place to attach AI-generated or computed overviews of a node's content.

Code written for endpoints, scheduled tasks, migrations, and AI tools now ships with clearer built-in guidance and documentation, helping the AI assistant generate correct graph, search, and tool code.

Code running in the workspace can now request strongly-typed JSON responses from the AI providers, making it more reliable to drive structured output from a model.

Hovering over a warning or error marker in the code editor no longer has its message replaced by documentation tooltips, so diagnostics stay readable.

The Curiosity SDK and its server endpoints now support a dry-run mode for write operations such as creating users, teams, folders, and uploading files, returning the resulting identifier without persisting any changes so integrations can preview the effect of a call.

Chat AI replies now track how long the model spent reasoning and show a "Thought for N seconds" summary once the thinking section is collapsed, so you can see how much time went into a response at a glance.

The chat side preview can now be dismissed directly, letting you clear an opened document or node and return focus to the conversation.

Tool calls in chat now mark failures explicitly and surface the error message alongside the call, and successful calls show their referenced sources more reliably.

The interactive graph view received layout and rendering improvements for a smoother exploration experience.

Resolved an issue that left the share dialog in a broken state, so adjusting who a node is shared with works consistently.

When you import changes that include tools or endpoints, the workspace now compiles them as part of the import and records whether each compiled cleanly, so updated code is ready to run without a separate step.

Importing entity spotters, patterns, pipelines, or node style settings now clears the relevant caches, so the new definitions apply right away instead of after a restart.

Read-only workspaces no longer attempt to raise notifications, avoiding errors in that mode.

Updating a node's access group now reliably clears cached state so the new sharing settings are reflected immediately.

Conversations can now be renamed from the chat view, and new chats get a meaningful title generated automatically from their content. Empty conversations are cleaned up so the chat list stays tidy.

You can now regenerate an assistant reply from any point in a conversation, re-running the model on the messages up to that turn without starting over.

The in-product code editor now shows richer documentation: type names in signatures, hovers, and autocompletion are clickable and open a documentation panel, interfaces list their concrete implementations, and you can drill into nested types. The documentation also surfaces the types and APIs available in scope for endpoints, AI tools, scheduled tasks, indexes, migrations, queries, and NLP post-processing.

Endpoints, AI tools, and scheduled tasks now keep a history of their code, so admins can review and restore earlier versions from a history view.

A "Convert to C# Connector" command is now available on CSV files in search results, generating connector code directly from an uploaded file. The CSV-to-class generation is more reliable with complex files, handling quoted separators, more date formats, nullable columns for missing data, de-duplicated property names, and headers containing quotes.

The audit view now covers login, code, query, and new-token events in addition to the existing types, and audit events can be written to a rolling log file on disk, with automatic cleanup of old logs. Configuration changes such as audit retention now show a saving confirmation.

The search and NLP configuration areas now use the same persistent sidebar layout as the rest of the admin console for a more consistent experience.

The "Toggle All" action in the entity editor now correctly applies its changes to the underlying data instead of only updating the on-screen state.

When the workspace is unhealthy, the health check now returns a 503 status code so load balancers and orchestrators can detect the condition.

When a workspace customizes the chat context panel, it is rendered even before the first message, and the chat input stays in sync as the context updates.

Editing a user from search results now opens a proper user editor instead of the generic node editor.

Corrected a rendering glitch in date and time facets in search results.

Added translations for Spanish, French, Italian, Portuguese and Chinese, and revised the German translations.

The endpoints admin view now tracks per-endpoint request volume, latency and error rates with inline sparkline charts, and a permissioned admin diagnostics panel shows an endpoint's execution logs while you edit and test it.

A new Schema view shows field-level statistics and value distribution histograms for any node type, so admins can understand the shape of their data at a glance.

The admin code editor now renders formatted documentation and method signatures in hover tooltips and autocomplete suggestions, making it easier to write endpoint, tool and task code.

Custom code and the downloadable front-end template now expose generated helper classes for the workspace's endpoints and AI tools, giving developers typed references instead of hand-written paths and identifiers.

Cron-based schedules for scheduled tasks and apps are now edited through an interactive schedule builder with validation instead of a raw cron text field.

Generating a connector from a CSV file now uses stronger heuristics for date-format detection and primary-key selection, producing more accurate generated code.

The monitoring view supports saving and switching between named views.

Successful user logins are now recorded to the audit log (with originating IP), debounced so repeat logins from the same user within an hour are not duplicated.

Search, data, profile and other admin settings now use a unified save button with explicit saving, saved and error states, and the search settings screens list all available indexes and facets inline so nothing is hidden. The admin sidebar and several settings pages were reorganized, including a dedicated Workspace Definitions page, and large log messages now render and scroll reliably.

Streamed chat responses are batched before being sent to the browser for steadier delivery, and each message now carries its position and total count in metadata.

Hardened OneNote extraction against concurrent updates and missing blobs, corrected recycle-bin cleanup so it captures all affected files, and added guards that make file extraction more reliable.

Corrected how filters behave when disabled so they no longer affect results unexpectedly.

Workspace definition exports now serialize deterministically, so repeated exports of unchanged configuration produce identical files.

The AI assistant can now run against a local or self-hosted Ollama server, with streaming responses and visible thinking, alongside the existing cloud providers. The provider and model selector now lives directly in the chat box, so you can switch the underlying model on the fly, and the assistant template is now chosen separately from the provider. Provider and model names are shown together so it is clear which model a chat is using.

While the assistant is working, the live progress now shows tool names as readable, spaced words (for example "Web Search" instead of "WebSearch").

The short cooldown that could block sending a new chat message right after the previous one has been removed, so consecutive messages go through immediately.

Uploading a definitions archive on the About screen no longer fails because of an incorrect file type, and imports complete as expected.

Endpoint editors now have an "Import code from another endpoint" button that opens a searchable list of endpoints to pull in. Imported endpoints stay linked, so they update automatically when the source changes.

The Logs view now includes a search box to filter log entries by text in addition to the existing log level selector.

The indexes list now surfaces an extra tag column for custom code indexes and includes that information when searching, making it easier to find a specific index.

Custom code endpoints now have access to the incoming HTTP request headers, and code that calls another endpoint can pass headers through to it.

Autocomplete in the code editors now shows the documentation comments for a method alongside its signature, so you get inline guidance while writing endpoint and tool code.

Long-running processes now relay a renamed step and report progress from running code scopes more promptly, so the live status reflects what is happening sooner.

The CLI now builds on the stable 2.0 release of its command-line framework, keeping argument parsing current and reliable.

Audio and video files can be scrubbed and seeked directly in the player. New files are stored in a seek-friendly form by default, and files served from S3-backed storage are delivered with range request support so playback starts quickly and jumping within a file works without downloading the whole thing.

Administrators can mount the workspace definition files (indexes, endpoints, code tasks, NLP models, and similar configuration) as a virtual file system at the /dav/ path and edit them with any WebDAV-capable client, signing in with an administrator account. The mount can be turned on or off from the workspace settings.

Result lists can now be shown in a table layout with columns, in addition to the existing card and grid views, making it easier to scan and compare many results at once.

The node editor now supports viewing and editing list, dictionary, and table fields directly.

Resolved errors that could prevent chat messages from rendering, and a case where an opened conversation would not load.

The bundled Llama model integration and the Setapp distribution have been removed.

Administrators can upload TrueType or OpenType font files from the file indexing settings so that documents which rely on specific fonts render correctly during processing and preview.

Endpoints and AI tools now show which imported workspace definitions reference them, with links to the related items, and importing definitions handles endpoints and AI tools more reliably. Importing workspace definitions has also been made more reliable, including more dependable file uploads, and change tracking now records which user made each change.

Users are now presented in a table, and access groups surface more detail about their membership and permissions.

A new action copies all custom node style definitions from a customized front end back into the workspace, with progress shown while it runs.

Additional rendering options improve font handling and fidelity when converting Word documents, including support for embedding PostScript fonts and resolving ambiguous fonts.

Tools can now be called and tested reliably even when no chat is associated, and testing tools that take array parameters works correctly. Search facet counts can now also be retrieved on the server.

Addressed a condition that could interrupt a graph task, and improved background monitoring to recover from unexpected errors instead of stopping.

Integrations can now specify the timestamp on a graph edge when creating or updating it.

You can now paste an image from your clipboard directly into the assistant chat box (for example a screenshot), and it is attached to the conversation automatically. Pasting multiple images in a row works too, each kept as its own file.

The AI Assistant tools settings now include export and import buttons. Export downloads all configured AI tools as a single archive, and import recreates them from that archive, so you can move tool configurations between workspaces or back them up.

The NLP entity post-processing settings gained matching export and import buttons, so post-processing definitions can be downloaded as an archive and restored or moved between workspaces.

The endpoints administration view now shows when each code endpoint was last edited and which user made the change, so admins can track recent configuration changes at a glance.

The system log view now has a log-level selector, letting admins narrow the live log to a minimum severity (for example warnings and above) instead of seeing every message.

Creating new AI tools from the default template works again, and exporting configuration archives now downloads AI tool and entity post-processing definitions reliably.

The size threshold above which images are dropped from a PDF during rendering has been raised substantially, so larger embedded images are now preserved instead of being stripped out.

Assistant replies are now structured into distinct parts, so reasoning, tool calls, and the final answer render as separate, clearly delineated blocks in the chat view instead of one undifferentiated message. This applies across the supported model providers.

The send and stop controls are now combined into a single state-aware button that prevents firing off messages in quick succession, so a reply cannot be sent again while one is still streaming.

When an AI tool call fails, the chat now marks it clearly and shows the error details when the entry is expanded, making it obvious when a tool did not complete.

Expanded translations for the AI-related interface across German, Spanish, French, Italian, Portuguese, and Chinese.

Administrators can download the full logs for a connector run and benefit from a refreshed log view that renders multi-line messages and appends new rows without visual flicker.

Importing and exporting AI tool definitions now reads access mode and access group settings correctly, and schema exports are emitted in a stable, predictable order so repeated exports stay consistent.

AI tools can now accept array-typed parameters, with validation that any non-required parameter declares a default value.

Search requests can now have their target, sort, and exclude queries set on the server, giving more control over how results are scoped and ordered without relying solely on client-supplied queries.

Startup now reports progress through its final phase, uses asynchronous I/O during initialization for smoother boots, and adds monitoring of the root drive plus a heartbeat-delay metric so operators can spot disk and responsiveness issues sooner.

Uploading files now reports live progress, limits how many files transfer in parallel, and lets you retry a file by dropping it onto the upload area again after a failure.

Inline source snippets attached to a chat answer only appear under the message they belong to, instead of every message showing the snippets gathered from the whole conversation.

Cached search data is now cleared automatically when you close the page or start a new search, reducing memory use during long sessions.

Administrators can now add custom code per entity type to merge, modify, or ignore extracted entities, managed from a new Post Process Entities settings page. The rules apply during document indexing, surface compile errors clearly, and are included in workspace export and import.

The integrations view now reports total elapsed time for connector runs, renders run details more clearly, and exposes connector runs as a facet so they can be filtered and inspected.

The endpoints administration view now retains status and info messages relayed from an endpoint while it runs, making it easier to follow long-running endpoint activity.

The toggle-all control in the entity linking editor now works correctly.

A notebook now continues importing its remaining pages when one child page is missing, and a user needing re-authentication is skipped so other users keep syncing instead of failing the whole run.

Asking the assistant to summarize a file that has not finished indexing now returns a message telling you to come back once processing completes, instead of a generic "nothing to summarize" reply.

Downloading the local link certificate now produces a valid file again.

Operators can now allow the workspace to start against a renamed external S3 storage bucket through a configuration setting, instead of being blocked at startup.

Older per-day usage records are now rolled up into weekly and monthly figures, keeping long-term usage reporting accurate without unbounded growth.

Stability and memory use are improved for long-running workspaces, including cached custom code, search queries, real-time connections, and background jobs.

Image-to-text now has two independent controls: classic on-device OCR for text in images and scanned PDFs, and a new on-device AI visual model that describes the contents of images. You can enable either or both, and turning on captioning re-indexes images and text-free PDFs so they become searchable by what they depict.

You can force OCR on a specific file even when OCR is not enabled globally, so a single scanned document or image can be re-processed for text without changing your workspace-wide indexing settings.

Search results now offer a "View JSON" command that opens the full node content in a read-only code viewer, in addition to copying it to the clipboard. It works for a single result or a multi-selection.

The evaluation-license banner now shows a countdown progress bar and stays out of the way after the first, longer appearance, making it less intrusive while still reminding you the workspace runs on an evaluation-only license.

When downloads are disabled for a workspace, the right-click menu on image previews is now blocked so images cannot be saved through the context menu.

AI tools can now be restricted to specific access groups instead of being available to everyone. Each tool carries an access mode, and access-controlled tools are only offered to members of the groups you assign, giving administrators fine-grained control over who can invoke which tools.

The AI-powered search and embeddings settings now live on their own "AI Indexing" page, split out from the File Indexing page, so semantic-search configuration is easier to find and manage.

Zooming or panning one chart in the monitoring view now applies the same time range to the other charts, and the charts gained a shared crosshair, making it easier to correlate metrics across the same moment in time.

The code editor used to write AI tools now offers completions tailored to the tool authoring context, so the suggested types and imports match what tool code actually has available.

Running a local development or evaluation instance no longer requires a license file; a built-in evaluation license unlocks the full feature set for development and evaluation use, accompanied by a recurring non-production reminder.

File content extraction can now run inside a Landlock sandbox on Linux, restricting the extractor process to only the folders it needs to read and its temporary working directory. This hardens the workspace against malicious or malformed documents and falls back to normal extraction if sandboxing is unavailable.

Resetting the facets in a search view fully clears the selection, including facets that were active before you started refining, so you get a clean result set instead of one that still carries hidden filters.

Custom left and right footer areas in the AI chat view are now only built when present, avoiding errors when a workspace does not configure them.

The known-links shown next to a result (parent folder, archive, notebook, file) now carry stable style classes, making them easier to target with custom workspace styling.

The administration view that lists past task runs groups them into more accurate time buckets, including thousands separators on the week counts.

When search results are merged across related nodes, type facet counts are no longer inflated by double-counting the same item, so the numbers shown next to each facet match the actual results.

A search can now be limited to chosen node-type fields, so queries only look in the indexes you care about instead of every searchable field.

Updating, reimporting, or deleting a custom code endpoint now invalidates its cached definition right away, so requests stop serving the previous version.

Script-based AI tools can now invoke a workspace endpoint (including running it on the primary from a replica) or call another AI tool from within their own code, making it easier to compose behavior.

The server-side search request now exposes the same helpers used by the UI for setting related, value, numeric, and time facets, so custom code can construct faceted searches the same way the front end does.

Search results now show how your query was understood (the "parsed as" explanation), including when only a single result type is returned, so it is easier to tell why you got the results you did. Hosting surfaces can also add their own extra content to this explanation line.

The type and property filter toggles now stay visible when a result set narrows to a single type or an already-active filter, and related filters behave correctly again.

The start and end dates of the time range filter are now applied correctly using the chosen range.

Comparing exported definitions against the current workspace now uses a cleaner side-by-side layout with a per-file list, clearer status icons and colors for added, removed, changed and unchanged files, and toggle buttons to show or hide each status so you can focus on just the differences that matter.

A lightweight helper process now watches the workspace and automatically restarts it if it stops responding, improving uptime without manual intervention.

Custom code in endpoints, indexes and tasks can now invoke a workspace AI tool directly and receive its result. Note that the existing endpoint-runner method has been renamed to RunEndpointAsync; update any custom code that called the previous name.

When sharing an item you can now add or remove its administrators, and changing how an item is shared now immediately refreshes the view for every other user who can access it instead of leaving them with stale information.

Custom code indexes are now required to check for cancellation (via the cancellation token) and will fail validation if they do not, so indexes can always be stopped cleanly.

A new wait-for command polls a workspace until it reports ready (with a configurable maximum timeout), which is useful for scripting deployments and startup sequencing.

The built-in AI search tool now combines keyword and similarity search and returns the most relevant passages from each matching file rather than whole documents. Answers cite those passages directly, so the AI assistant grounds its responses in specific quotes from your content and links each claim back to where it came from.

Sharing now stages all of your changes (who can access an item, public visibility, and related options) and applies them only when you click Apply, so you can review adjustments before they take effect instead of each toggle committing immediately.

You can now be prompted to confirm before signing out, preventing accidental logouts that interrupt your work.

Search results now show their surrounding context, including the parent folder and parent file (for example the archive or OneNote notebook a result lives in), and rendering of email and folder results has been refined for readability.

The workspace now provides only the pairing certificate for connecting the desktop app, with the app itself obtained from the release page, and adds a dedicated settings area for managing the connection.

Individual chat messages can now be flagged to be excluded from indexing, so transient or sensitive messages do not surface in search results.

Outlook template files (.oft) are now recognized and processed, so they are extracted and indexed like other Outlook items.

Imports now remove stray OneNote artifacts such as backup files and recycle-bin entries that were previously picked up, and correct mislabeled notebook sections, giving a tidier set of indexed notes.

PST email archive tasks now track when a user's authentication has expired, so the task surfaces accounts that need to re-authenticate instead of stopping without explanation.

The integration SDK can now send multiple log messages in a single call (automatically splitting oversized batches), and adds a rate limiter, making custom connectors that produce high log volumes faster and more resilient.

Administrators can now define, schedule and run data migration tasks against the workspace from a dedicated migrations screen. Each task can be written in code, run on demand or in the background, show live progress, be cancelled mid run, and have its status reset, so larger one-off data changes can be applied and tracked safely.

The AI chat experience exposes new hooks for embedding and tailoring it inside a workspace, including setting the current chat, attaching metadata to a chat request, and customizing how message bubbles and context are rendered.

The "Leave" and "Hide" actions on spaces and saved searches now use the workspace's configured names for those areas instead of generic labels.

The Teams item now appears correctly in the administration sidebar.

Facets behave correctly when items are pinned while a search is active, and locked facets are handled more predictably.

Custom code endpoints can now opt into request monitoring individually, so administrators can collect metrics only on the endpoints they care about.

A simple text index can now be added to any enum field, making those values searchable.

Triggering an AI chat through an endpoint now validates the conversation state up front (rejecting duplicate or too-frequent runs), can restrict which tools the assistant may use, and returns the generated message along with clearer errors when something goes wrong.

Result scores are now summed correctly under certain sort modes, target sort queries sort as expected, and empty search highlights no longer appear.

Running migrations are cancelled gracefully when the workspace shuts down.

A new optional Local Link integration lets you open a file in its native application, or reveal a file or folder in your operating system's file browser, directly from search results. When an administrator enables it, the workspace hands off signed links to a small companion app (available for both macOS and Windows) so the action happens securely on your own machine.

Search filters and facets now continue to show even when the current query matches nothing, so you can adjust your filters instead of being left with an empty screen.

Typing into contact, user, and group pickers when sharing no longer records those lookups in your search history.

When downloads are disabled for a workspace, the built-in video player now hides its download control and right-click download option.

The pop-up search box opened with the keyboard shortcut can now be turned off for deployments that prefer not to use it.

Clearing or adjusting a time-range filter now takes effect right away, and tooltip and locking behavior on the filters panel is smoother.

The scheduled tasks view in the admin console now lists tasks alphabetically by name, making them easier to find.

The sample code provided when configuring a custom chat AI provider has been brought back in line with the current provider interface, so it builds without manual edits.

Indexing Microsoft OneNote content now handles notebooks that have been deleted at the source; the corresponding folder and its contents are removed from the workspace.

Enumerating large query results now streams instead of holding everything in memory, reducing memory use when working with big collections.

Working with facets is steadier across the board: saved filters apply correctly, pinning behaves consistently, facets render properly on small screens, and the facet panel locks and unlocks as expected. Loading a saved set of facets now replaces the active selection cleanly instead of stacking on top of it.

Versioned front-end assets are now cached for longer, so returning to the workspace loads faster while still picking up new releases immediately.

Date-based time facets apply the correct range, selecting text in the assistant view works inside nested embedded documents, sidebar favorites update immediately, and lazily loaded results scroll more smoothly across changing item heights.

When reviewing differences between built-in and customized definitions, the compare panel is now resizable and labels missing content more clearly.

A new metric reports the number of users active in the last minute, giving operators a finer-grained view of live usage.

Searches that include common stopwords now produce more accurate results, and matched terms are highlighted correctly even when a stopword sits between them. The query highlighting endpoint also returns its result with the correct text encoding.

Exported index definitions no longer include an internal identifier, making exports more portable between workspaces.

Administrators can now compare a workspace definitions backup file against the current workspace in a side-by-side diff view, so you can review exactly which indexes, endpoints, code tasks and NLP models would change before importing. The comparison highlights added, removed and modified definitions per file so you can confirm an import does what you expect.

The share and permissions dialogs now only offer access groups when teams are enabled for the workspace, so users on single-user setups no longer see group options that do not apply to them.

The workspace now watches for approaching high-memory situations and proactively pauses indexing and trims caches to stay responsive under heavy load. Loading large workspaces from disk is also faster on server deployments thanks to larger read-ahead during startup.

The lock-facets and clear-facets controls now sit directly in the facets bar alongside the filters they act on, making them easier to find while refining a search.

Labels shown across the interface now resolve through an aggregated lookup, reducing the number of round-trips when many labels are displayed.

Adjusting the file content extraction settings now shows a warning that the file content index must be rebuilt before existing files reflect the new settings, so the effect of the change is clear.

The facets toggle now appears correctly on narrow and mobile-width screens, and scrolling through content is smoother.

The workspace icon setting is always available now, with a note clarifying that when a custom front end overrides the logo, this setting controls only the favicon and the startup and shutdown icons.

When the workspace is configured to use external (S3-compatible) blob storage, connectivity and the required bucket permissions are now verified before the graph storage starts, so a misconfiguration surfaces immediately rather than failing later. The check now requests only the permissions the workspace actually uses.

Live (websocket) connections now authenticate correctly for workspaces using a custom authentication flow.

Results that reference other people now return only basic profile fields (name, username, email) instead of the full user record, tightening what is shared in search responses.

Duplicate child hits are no longer repeated under a search result, so nested matches appear once each.

Permissions now propagate through chains of nested access groups, so a user who belongs to a group that is itself a member of another group correctly receives the inherited access.

Files uploaded to a space, a user, or an access group now use scoped storage paths, preventing different uploads that share a filename from overwriting one another.

OneNote content is now ingested with its section-group hierarchy, and missing or unmatched section identifiers and deep links are handled gracefully instead of failing the sync.

Removing a user now clears that user's cached spaces, so connected data sources reflect the change immediately.

Entity linking settings now round-trip correctly through workspace definition export and import.

File upload methods accept a new option to suppress per-upload log messages, useful for bulk ingestion where the upload logging would otherwise be noisy.

When you open a page using only active filters with no typed query, results and their facets now render instead of showing an empty page.

Opening one of your saved searches no longer re-records it in history, and drilling into related facets no longer clutters history with throwaway entries.

Lists of related links can now expand inline instead of only opening a popover, making it easier to scan everything connected to an item.

Node styling and custom renderers now load for regular (non-admin) users, so items render with their intended appearance on first view.

Autocomplete suggestions behave correctly after the search box is dismissed, the Leave and Hide actions use the correct wording for spaces and saved searches, and the multi-action bar stays on top of other content with a scrollable overflow menu.

Editing a schema now lets administrators change the name of the key field, with validation that blocks conflicts with existing fields.

Editing a spotter model and saving from the overview now keeps the configured exception list.

Admin-only routes are now blocked for non-admin users in the browser, and the workspace metrics endpoints require system-admin access.

When uploads or account-preference editing are turned off in the workspace settings, those operations are now rejected for regular users at the server, not just hidden in the UI.

Related-node facets now respect each user's access rights, so restricted items no longer appear through facet values.

Tightened input validation, file handling, and access checks across the workspace, including validation of uploaded images before storage.

System administrators can again see connector run and AI tool history that had stopped appearing after a visibility change.

Calling access-controlled and admin-only code endpoints with an API token now works correctly.

Internal scheduling checks are faster, reducing background overhead.

Certain internal node types are now visible only to administrators rather than appearing in regular users' results.

You can now export a complete snapshot of your workspace's configuration (indexes, endpoints, code tasks, NLP models and patterns, AI providers and schema) to a single portable zip file, and import it back into the same or another workspace. The export deliberately contains definitions only and no document data, making it suitable for backing up and moving a workspace's setup between environments. This is available to administrators from the workspace settings and from the command-line tool, with clear, itemized error and warning messages when something in the imported file cannot be applied.

Saved searches and facet queries can now define an exclude clause, so results matching the excluded query are removed from the result set, letting you narrow searches by saying what you do not want to see.

When an AI chat request fails, the chat now surfaces a clear error and includes the underlying details so the problem can be diagnosed rather than failing silently.

Autocomplete now appears correctly for saved searches, quoted tokens render properly, the search box makes fewer autocomplete requests as you type, and the space placeholder and quoted-token display were corrected.

The time facet chart no longer loses its sizing, and facet sizing was corrected so charts render at the right dimensions.

Confluence pages and files now sync and convert to readable content more reliably.

Confluence can now be connected as a multi-user data source so its files and pages are shared across the workspace rather than tied to a single user's connection.

User names are now kept up to date from the SAML login flow, so accounts reflect the latest name provided by the identity provider.

Users with viewer access can no longer edit saved searches.

A new setting controls how many results are tracked for related-facet computation, giving you control over the trade-off between facet accuracy and performance.

A new endpoint returns the highlight pattern for a search query, enabling consistent term highlighting on the client.

The command-line tool now imports and exports full workspace definitions as a zip file, replacing the previous custom-code-only export. The commands validate that the path is a zip file and report per-item errors and warnings on import.

The breakdown of how a query was interpreted can now be shown in a compact or detailed form (or left on automatic), making it clearer how each search term contributed to the results.

After you trigger an action on selected results, the selection is now cleared automatically so the next action starts from a clean state.

Recent search history once again appears in the search box suggestions, and autocomplete settings were tidied up for more consistent behavior.

Items that fall outside the main facet groups are now grouped together correctly under "others".

The logs view no longer slows the interface down while it is open.

Administrators get a new control on the search box that opens a detailed, step by step breakdown of how a search was executed and how long each stage took, useful for diagnosing slow or unexpected results.

Custom API endpoints can now be reliably imported and exported from the admin console, including their access group assignments, and the access group is shown in the endpoint header.

The license creation screen now includes the replicas capability, and the extra feature toggles default to enabled when drafting a new license.

Chat requests that are cancelled, hit a model with no configured AI provider, or exceed the model's context length are now closed out cleanly and surface a clear message, so the chat stays responsive.

Text summarization can now be asked to produce its output in a specific language rather than always following the user's preferred language.

Signing in through a custom authentication flow now works on read-only replica servers.

Confluence attachments are now checked directly for availability so removed attachments are cleaned up correctly, and the OneNote recycle bin folder is no longer brought into indexed content.

Storage loading now reports clearer progress, and pushing tracked changes to a connected Git repository runs in the background so it no longer blocks other work or times out.

When creating users through the SDK you can now choose to match and de-duplicate users by username instead of email.

The git sync command can now present a client certificate for mutual TLS, configured through the new CURIOSITY_CLI_MUTUAL_TLS_CERTIFICATE_PATH and CURIOSITY_CLI_MUTUAL_TLS_PRIVATE_KEY_PATH environment variables.

A new verbose option on git sync adds detailed diagnostic logging to help troubleshoot synchronization issues.

Several fixes make pushing tracked changes more reliable, including handling a missing certificate on the websocket connection.

Pinning and reordering search facets now sticks reliably, and the saved order is applied per search context and preserved across reloads.

Result action menus now only show commands that actually apply to the selected items, and the filter tooltip text was simplified to "Select this filter to see all results of this type."

The email composer no longer saves spurious drafts right after a draft is opened, reducing accidental draft churn while you start editing.

Deleting a user or deleting a schema and all of its data now requires typing "DELETE" before the confirm button enables, guarding against accidental, irreversible deletions.

AI provider configuration now lives on a dedicated administration page rather than being mixed into other settings, making it easier to find and manage.

The connector run logs view now loads in pages and can show timestamps, so long-running connector histories stay readable instead of loading everything at once.

The endpoint editor now shows the right option as selected between the default pooling mode (for long-running calls) and run-in-sync-with-request mode.

Single-user connector syncs (including Box, Confluence, Jira, Notion, Salesforce, HubSpot, Evernote, pCloud and Outlook Calendar) now pause for an owner whose authentication has expired and prompt to re-authenticate, rather than retrying repeatedly.

Imported Confluence attachments now resolve to the correct page-relative URLs, and an attachment is re-imported when its link changes.

File syncing now detects and handles file size mismatches and non-unique files returned by the server, so imports complete cleanly.

Connected OneNote content from OneDrive, SharePoint and Teams is now extracted and rendered on demand, so you can preview a OneNote page directly in the workspace and have it ingested in the background. Source detection and file-type handling for OneNote were reworked end to end for more reliable previews.

Result cards now show a dedicated "More actions" button that gathers the commands that don't fit in the header, making secondary actions easier to find.

Searching within the filters list now offers a "Show all filters" button to clear the search and bring every filter back into view, and the filter bar refreshes immediately when you pin or unpin a filter.

A new setting lets a workspace always display how a search query was interpreted, instead of only when something was rewritten.

Workspaces can now configure a sign-out URL so signing out redirects to a chosen page instead of the default sign-in screen.

When an instant preview cannot be generated or authentication has expired, the message is clearer and the workspace offers to open the file directly.

The empty-state prompts and onboarding text now use your workspace's configured names for spaces, so renamed terminology is applied consistently.

Links inside Confluence pages now open in a new tab and clicking links inside embedded Confluence content works as expected.

References to "ChatGPT" were standardized to "OpenAI" across provider selection and related screens.

The admin usage view now orders results chronologically and adds breakdowns of searches per user over time and the most frequent query terms, giving administrators a clearer picture of how the workspace is being used.

Confluence synchronization is better at detecting when there is nothing new to fetch and stopping early, resolves file links correctly, and refreshes content more dependably.

When a model emits reasoning steps, the assistant now shows a brief "Reasoning..." status and clears it once the answer begins, instead of surfacing stray reasoning output.

People and contacts now render as a compact avatar-and-name bubble, giving a consistent way to show who a contact or user is across the workspace.

After running an action on a set of selected search results, the items are now properly deselected so the next selection starts clean.

Confluence syncing now imports content completely, correctly downloads attachments and blobs as the right user, refreshes authentication during long runs, and treats an unreachable Confluence server as a transient condition rather than ending the run.

When you open or upload a file, instant text extraction and PDF preview now wait gracefully when the OCR queue is full, and harmless cases (a file that is already extracted) are handled quietly rather than reported as errors.

The connector SDK now lets you configure the underlying client when listening to websocket messages, so you can adjust connection settings such as headers or proxies for your environment.

Selecting a result now clears selections in other search areas, so only one set of results stays selected at a time and bulk actions apply only to what you picked.

Result and Space commands now use your workspace's configured term for a Space (for example "New ..." and "Remove from ...") instead of a hard-coded label.

Action-bar command buttons now stretch to the full width and left-align their labels for a tidier list, and the Clear and Apply buttons in the facet panel were reordered for a more natural left-to-right flow.

New endpoints now start with sensible defaults, running asynchronously and open to all users, so you spend less time adjusting settings right after creating one.

The query builder can now export a search to a portable text form and rebuild it later, making it possible to persist, share, or reload a configured query.

Scheduled syncs for email (IMAP), Google Calendar and Drive, Microsoft 365 and SharePoint, contacts, and pCloud now recognize expired or invalid credentials, blocked or forbidden access, dropped connections, and cancelled operations, and signal that a connection needs re-authentication instead of failing the whole sync. Missing files and unsupported server features are handled cleanly, and unexpected errors are logged with enough context to identify the affected item.

The match counter in the in-document search bar now truncates with an ellipsis and shows the full count on hover, so result summaries stay readable in languages with longer wording.

Documents and files with no title are now labelled "(No Name)" instead of appearing blank, making them easier to recognise in lists and headers.

When a user no longer qualifies as a system administrator on sign-in, their administrator role is now removed promptly, keeping access in sync with the current authorization state.

OneNote, OneDrive and SharePoint synchronization is more reliable, with better handling of folders, notebooks and transient errors, and it now checks whether notes can be indexed before attempting to index them so connector runs complete more often.

When a chat AI request is blocked by Azure OpenAI's content policy, the workspace now reports it as a content moderation result rather than a generic failure.

The "save search" button now uses the workspace's configured name for spaces in its tooltip instead of a hard-coded label, so the wording matches what you see elsewhere in the interface.

When the live updates connection drops and re-establishes, the workspace now shows a clearer "Server Events reconnected" notice.

Search can now exclude specific node types from result merging, both when merging related hits and when expanding them, giving more precise control over how grouped results appear. Some stability fixes to result merging accompany this change.

Your past AI chats now appear in the chat list and search even if the AI provider that created them has since been disconnected or deleted.

The Confluence connector handles base URLs more reliably and now correctly flags a connection as needing re-authentication when its credentials or access scopes are no longer valid, instead of continuing without warning.

Domains parsed from email addresses now strip stray quote characters, avoiding duplicate or malformed organization entries.

Custom code endpoints now relay progress status to callers by default when invoking other endpoints, so long-running operations report progress end to end. The progress-reporting call available inside custom code is now named RelayStatusAsync.

Workspace availability, deployment identity, and health endpoints are now reachable at a fixed root location regardless of the configured base path, making external monitoring and readiness probes simpler to wire up.

The filter panel now lets you pin your most-used filters so they stay at the top, lock a filter so its selection is preserved across searches, and hide filters you do not need. Your pinned, locked, and hidden choices are remembered between sessions, and the filter modal has a cleaner, more compact layout.

Search areas can now present their own tailored set of filters and filter settings, so different parts of the workspace can show the filters that make sense for that content.

Replies generated in the email composer no longer show internal placeholder markers and now display the formatted text as intended.

Add-to-Space commands and dialogs now follow your workspace's configured term for "Space" everywhere, instead of using the fixed wording in some places.

Operators can now enable additional diagnostic logging for the custom authentication flow at startup, making it easier to troubleshoot single-sign-on and identity mapping issues.

When several users connect to the same Confluence instance, update checks are now tracked per user, so one person's sync no longer suppresses or interferes with another's.

Failed document translations now report which part of the file caused the problem and give more specific messages for moderation, token-limit, and connection failures.

Tightened input validation, file handling, and access checks across the workspace, including stronger session token signing keys for new workspaces.

Stored relationships now use a more compact timestamp representation, reducing the memory footprint of workspaces with very large graphs.

The embeddings index is now optimized when saved, improving similarity-search performance.

Resolved an issue affecting cleanup of removed items during Outlook desktop synchronization.

Pressing Ctrl+C while serving a front end now stops the server cleanly instead of terminating abruptly.

The command help now uses up-to-date wording for the workspace URL and for where to create a Library token (Manage > Tokens > Library).

The assistant can now show source documents and generated files directly inline within a chat message, rendered as compact cards. Clicking a card opens the item in the side preview, and right-clicking it brings up the same command menu available on search results.

Tightened input validation, file handling, and access checks across the workspace. Links inside chat messages are now captured and opened consistently.

A single AI tool definition can now expose more than one callable function, so admins can group related capabilities in one tool. Tool parameters can also declare default values, which are applied automatically when the model omits them.

New workspaces now ship with ready-to-use built-in AI tools, including a document search tool and an image generation tool, so the assistant has useful capabilities without manual setup. Tools can also store generated files and attach result snippets that surface back to the user.

The Confluence data connector now maintains a cleaner folder hierarchy for synced pages, ensuring each item has a single correct parent folder, along with refreshed styling for displaying Confluence content.

Tool calls made through the Amazon Bedrock AI provider now use consistent identifiers, fixing mismatches that could occur during multi-step tool interactions.

A new command syncs a local git repository with a workspace through the git tracker, letting you connect and push repository state to a Curiosity Workspace from the command line.

The chat assistant can now call tools while it answers, so it can take actions and pull in live data instead of only generating text. The conversation view shows which tools were invoked during a response, and custom tools you define are validated against the expected structure before they run.

Your workspace can now host its tools over the Model Context Protocol, letting external MCP clients discover and call them. This makes the workspace usable as a tool provider for other AI applications.

A new file translation capability lets the assistant translate document content, with clearer error reporting when a translation cannot be completed.

Search result rendering can now be given a maximum content height, so embedders can control how tall individual result cards grow.

Search experiences can now post-process facet results before they are shown, allowing custom filtering or relabeling of facet values.

Whether a result is pinned is now carried through with the search result, so result rendering can reflect and act on the pinned state.

Reconnecting after a dropped connection now refreshes the conversation so no messages are lost, and duplicate assistant replies no longer appear. Empty streamed fragments are ignored, giving cleaner responses.

Administrators can configure a custom authentication flow that reads identity details (email, username, name, admin status, department, company) from an incoming bearer token and maps them onto a workspace user, automatically creating or updating that user. This supports integrating the workspace behind external identity providers and reverse proxies. Custom header naming was harmonized and authentication checks were reordered for more predictable behavior.

The login flow now reuses an existing valid session before prompting for credentials, with more reliable cookie handling and secure-cookie detection.

The custom AI provider editor gained a full-screen editing mode and several usability refinements when configuring providers.

A changed default assistant is no longer overwritten on the next start, and assistant configuration (including names) now saves correctly.

Text normalization for indexing now handles diacritics and unusual character sequences correctly, so documents with accented or irregular text index reliably.

Generated answers, summaries, and key-point extractions now respond in each user's preferred language, and the assistant grounds its claims more strictly by quoting the relevant source snippets and citing them inline, reducing made-up references.

When related results of one type are grouped under another type, the affected type facet now shows an indicator and a hint explaining that filtering by that type alone reveals the grouped results. Filtering to a specific type no longer hides the results that were merged into it.

Search now applies n-gram matching across an entire indexed thread, improving recall on partial-word and multi-part queries over file content.

File content indexes can now normalize diacritics, so searches match regardless of accents (for example "resume" matching "résumé"). An accuracy issue in how diacritics were normalized during indexing was also fixed.

Search and facet computation now complete more dependably under load, and diagnostics for operators run in a separate helper process so they no longer affect the running workspace.

Text detection and Markdown conversion now run under a time limit, so a single difficult document no longer holds up content extraction.

Recent-search suggestions now appear when you click into an empty search box rather than automatically, and they can be turned off entirely via a setting.

A new option lets a workspace suppress the onboarding tips and tooltips for users who do not want them.

Fixed the AI chat so the search and follow-up buttons show the right state when global search is disabled or when the selected assistant template already has its own knowledge source.

Switching the UI language reliably reloads the workspace and rebuilds the navigation to reflect the new language, and the German translations were refined.

The server now logs a clear warning when its TLS certificate is within two months of expiring, giving operators advance notice to renew.

A startup setting lets an operator define the workspace name shown before the graph has finished loading.

Users created through the CLI access-control mapping are now activated on creation, so people synced from directory sources can sign in without a separate activation step.

The way the AI assistant gathers and uses context from your workspace was overhauled for more relevant answers, with improved retrieval prompts and system prompts. Conversations now continue even when no supporting context is found, older chats still surface their referenced files, and the assistant only consults the embeddings index when one is actually in use.

Opening the search box now shows your recent searches before you type, not just while typing, making it quicker to repeat a previous query. Hosts embedding the search box can also supply their own autocomplete suggestions.

New options let you hide result icons and hide the search button where they are not wanted, and pinned results have cleaner styling.

Search results can now show header commands, and custom commands render consistently across result banners and previews. Global commands can be registered to appear everywhere.

Facets can be configured to skip the pop-up modal, clearing an inline filter now also applies the change immediately, the secondary facets bar no longer appears when it should not, and the time facet no longer occasionally gets squished to the top of the panel.

Several display and interaction issues in the chat view were resolved, including an error when a conversation had no messages and inconsistent behavior of the in-chat search button.

If reading a file for the AI assistant takes too long you now get a clear message that it timed out, and file extraction failures surface a more specific error instead of a generic one. PDF extraction also reports more detailed progress along the way.

Searches now run on their own dedicated thread pool so heavy queries are less likely to block other work, with added timing metrics to track waiting and throughput.

Improved handling of abbreviations and of full file paths in search, plus a fix to result scoring for vector similarity search so the most relevant matches rank correctly.

Fixed handling of recurrence end dates and end conditions when syncing recurring calendar events, so recurring events import more accurately.

The search box now suggests completions drawn from your own previous searches, ranked by how often you have run them, so frequent queries are one keystroke away.

Search areas can now show a resizable side preview next to the results list, letting you inspect a selected item without leaving the result set. The pane appears when there is something to preview and collapses back to full width when there is not.

Active filters are now highlighted and read more clearly, showing the selected value inline or a compact "+N" count when several values are active, and file facets (source, kind, extension, size) are ordered consistently.

The assistant chat box text, search button labels, tooltips, and message placeholders can now be customized to match a workspace's wording.

Commands shown in the action bar and on item view and preview can now supply their own button rendering for a more tailored look.

Additional interface strings are now translated across German, Spanish, French, Italian, Portuguese, and Chinese.

Messages shown when the AI chat provider is unconfigured, a request is rate limited, the context grows too long, or the service cannot be reached are now clearer and less alarming, with concrete guidance on what to do next.

Streamed AI responses now handle partial and oversized message chunks gracefully, recovering and reconnecting instead of breaking the live update, so long answers and progress arrive smoothly.

Progress for multi-item AI tasks such as summarization, key-topic explanation, and translation is now reported at a steady cadence rather than in bursts, and always finishes cleanly at completion.

Requesting an immediate file extraction now starts a worker even when background extraction is paused, so on-demand extracts no longer wait.

Asking for a chat summary no longer fails when an assistant template is missing, avoiding an interrupted response.

The command-line tool now routes its connections through the configured system proxy, so it works in environments that require one.

The AI chat view has been overhauled across the web interface and the server: sending a message now shows a spinner while it is in flight, surfaces a clear error message if a message fails, and refreshes your chat list automatically once a reply arrives. Loading existing chats and messages no longer flashes a blocking progress bar, so switching between conversations feels smoother. Chats with files now report finer-grained progress while the assistant evaluates the relevant context, so long lookups stay responsive.

Code areas now render correctly when you start a chat directly from the home screen. Labels that include an image size their text to fit instead of being clipped, and the floating command action bar reliably stays above other content.

GPT-4.1 is now available as a chat model (with a much larger context window), and the older o1 and o1-mini options have been retired. Existing chats fall back cleanly to a supported model.

When you point the OpenAI integration at a custom host that already includes the full URL (with its scheme), that address is now used as-is instead of being rewritten, making it easier to connect to compatible third-party providers.

Automatically generated chat titles now use the chat's language, falling back to your preferred account language, instead of defaulting to English.

The built-in summarization no longer forces overly short summaries; the model now chooses an appropriate length for the content.