📌 Release Overview

• 2026.32

• Unique hosted: August 5, 2026

Chat Platform

🚀 What’s New (Business/End-User Focus)

User EXPERIMENTAL Informative notifications can now be capped per day and / or week: The number of informative notifications a user receives can now be capped per day and / or per week.

Admin Experimental Service users API key scoping: Service users can now be synced into the platform's user system, where they're clearly identifiable and shown separately in Access Management. Using the Unique API, personal API keys - including for synced service users - can now be generated, regenerated, or revoked directly. Read more in the Unique API Release Notes.

🔧 Improvements

User EXPERIMENTAL Notifications in app are now delivered real time: The notifications in app are now delivered real time.

User Model selection in chat is now grouped by provider: The dropdown for the model selection in chat now groups the models by provider.

Admin ROI metrics in the detailed chat interactions report: The detailed chat interactions CSV export now includes per-interaction metrics for ROI and time-saved reporting: tool count, referenced sources count, subagent and MCP tool names, prompt and answer length, loop iterations, model name, total response time, artifacts count and filetypes, output size, cost in USD, and total token usage.

🐞 Bug Fixes

User No more duplicate "ghost" text when pasting into the chat input: Pasting text into the chat prompt could leave a faint duplicate of the pasted content behind the real text, with the cursor visibly out of line with what you typed. The input's text and its highlight layer now stay in sync while typing, pasting, and scrolling, so what you see is exactly what you're about to send.

Admin Email address changes in Zitadel now sync to Unique: When a user's email was changed in Zitadel, the update never reached their Unique user profile. The platform now reacts to Zitadel email-changed events.

RAG

🚀 What’s New (Business/End-User Focus)

Admin Detailed per-file ingestion report: Data administrators can now generate a downloadable CSV from the Analytics page for a selected date range. The report maps Unique content IDs to source file IDs and includes file paths, source details, size, page count, ingestion status, timestamps, metadata, and failure reasons - making it easier to reconcile source inventories and troubleshoot failed or missing ingestions across connectors.

User BETA New Knowledge Base UI refresh: The redesigned Knowledge Base browsing experience, introduced as Experimental in 2026.28, is now available as Beta. Folders and files are shown in a single sortable list with a side detail pane, making it faster to scan large folders and jump straight to a file’s details. Highlights include:

• Sortable list view with sortable columns (name, type, ingestion status, last updated) replacing the stacked column navigation.

• List view filters to quickly narrow down folder contents.

• Detail pane with a redesigned folder header, ingestion status card, and folder analytics shown alongside the list.

• Drag-and-drop upload directly onto the list, with clearer upload and ingestion progress.

• Consistent List and Column views with a unified top bar (New Folder, Upload, Search, view toggle) and the same side panel behavior in both views.

🔧 Improvements

User Speed up Select Files with Advanced KB filters: Fixed slow loading of the Select Files / folder tree when a space uses Advanced KB scope filters. The file context window now loads much faster for filtered spaces.

Agentic Table

🚀 What’s New (Business/End-User Focus)

User EXPERIMENTAL Automatic assignee email notifications: Agentic Table (DDQ) sheets now send an automatic email the first time a user is pulled into a sheet (or re-activated when their open-row count goes from 0 → 1), replacing the manual "Notify assignees" button. Emails are consolidated per recipient per sheet (never one per question) and include the assigner's name, the sheet name, and a deep link. Group assignees are expanded to their individual members — including users who are eligible only through group membership — so no valid assignee is missed, while the assigner and the "All Company" root group are excluded.

User Images in Agentic Table cells: Agentic Table cells can now render images, using the same content-id markdown notation already supported in chat messages.

ADMIN Per-row dynamic metadata filters in Agentic Table: Column search filters can now use <row.metadata.KEY> tokens that resolve from each row's metadata at run time, instead of a single static filter shared across the whole space. This lets each sheet (or even each row) search a different slice of the knowledge base — e.g. "only documents valid as of this sheet's cutoff date" — set per sheet via the existing metadata API, with optional per-row overrides. Works for both custom and subagent columns; missing/unparseable values fail open (the condition is dropped) to avoid silent zero-result cells. More details can be found in the documentation.

🐞 Bug Fixes

User Column widths persist after cell update: Custom column widths in an Agentic Table are now preserved when a cell is updated. Previously, editing a cell triggered a table layout reset that discarded users' resized columns.

User Feedback export includes the original question: When exporting Agentic Table feedback to CSV, each row now carries the original question that produced the answer (previously the source-question columns were empty for in-cell feedback), so exported feedback can be mapped back to what was asked and used to refine the sub-agent.

Unique AI

🚀 What’s New (Business/End-User Focus)

User BETA End user can now enable and edit their context memory: Unique AI Chat can now remember your working style and recurring context across sessions, no more re-explaining yourself in every new conversation. Enable it once; it travels with you wherever Context Memory is supported. The memory is shared across spaces where it is activated. See here for more details.

ADMIN New Models Added to Unique AI Approved List: GPT-5.6 Terra, GPT-5.6 Luna and Claude Opus 5 have been added to the approved model list. These models are now appearing in the Unique AI dropdown, provided they are enabled for your organization.

🔧 Improvements

Admin Configurable token budget for Internal Search: The Internal Search limit now applies to the combined result of all search strings rather than each one individually, and the default is unified at 200 chunks. RAG research showed that 200 chunks is enough to reliably surface the golden chunk, which strikes the right balance in the trade-off between retrieval accuracy on one side and response time and token cost on the other.

🐞 Bug Fixes

User Web search blocks unsafe URLs individually instead of failing the batch: URL safety validation now handles unexpected errors per URL, so a single problematic link is blocked on its own while all other URLs in the same request continue to be crawled. Previously an unforeseen failure while checking one URL — for example a malformed address — aborted the entire search batch, causing the assistant to lose otherwise valid web results.

Legacy Modules

🔧 Improvements

Admin Additional LLM options available on more AI modules: Space administrators can now set additionalLLMOptions on Chat with GPT, Document Summarizer V2, the Translator V2 translation services, and the assistant's module-selection step, passing any parameter supported by the model or provider straight through to the LLM request. The fields input, messages, model, stream and tools are ignored, since the platform continues to own those.

🐞 Bug Fixes

User Stop button no longer stays visible after a document summary finishes: Short documents that produce a single summary chunk were published without a terminal timestamp, so the chat kept treating the assistant message as still streaming and left the Stop button on screen. The final summary is now marked as complete, and the Stop button disappears as soon as the summary appears.

Admin Document translation no longer leaves temporary files behind: The translator downloaded each source document into a temporary file that was never removed, so disk usage grew with every translation whether it succeeded or failed. Temporary files and directories are now cleaned up automatically on both the success and failure paths.

Language Model

ADMIN Claude Opus 5

Opus 5 is now available through LiteLLM. Two deployment modes are possible:

• litellm:anthropic-claude-opus-5: Directly through anthropic API

• litellm:vertex-claude-opus-5: Through VertexAI (processing restriction to EU)

ADMIN Model retirement

The following models are deprecated and not further available after August 5, 2026:

• litellm:anthropic-claude-opus-4-1