Release 2026.22
5 min read
📌 Release Overview
2026.22
Unique hosted:
Chat Platform
Platform upgrades and security
Admin Modernised runtime across the whole product:
Release 22 completes our migration to Node.js 24 on Debian 13 (Trixie) base images for every backend service, and ships the Next.js 15 upgrade on the chat app. The result: faster cold starts, smaller container images, and the latest LTS runtime guarantees.
Admin LiteLLM (GA): LiteLLM is now Generally Available. LiteLLM acts as the AI Gateway connecting the Unique platform to approved AI model providers beyond the standard Azure OpenAI setup, including models hosted on Google Cloud Vertex AI and other enterprise AI providers. For Unique SaaS deployments, Unique handles provider setup, compliance controls, and operations; the client reviews and accepts the relevant sub-processors, regions, and contractual terms. Self-hosted clients are responsible for their own provider onboarding, compliance review, and ongoing operations. Full documentation is available here:
🐞 Bug Fixes
User Multi-File Paste: Fixed an issue where pasting multiple files into a chat only uploaded the first file; all pasted files now reach the upload flow correctly.
User Navigation Tooltip Truncation: Fixed the navigation tooltips so long space and chat names are shown in full instead of being cut off vertically.
Admin Ingestion Config Modal Scrolling: Fixed the Configure File Ingestion modal so the state-select dropdown in "UI Ingestion State Messages" is now fully visible without manual scrolling.
RAG
🔧 Improvements
User Search Improved combined search error handling so failed search requests return clearer, safer errors and are easier to diagnose.
Admin Experimental Hyperlink Extraction alongside Microsoft Document Intelligence: When ingesting PDF documents, Unique now extracts embedded hyperlinks — including their anchor text and destination URLs — in addition to the standard text extracted by Microsoft Document Intelligence (MDI). Previously, MDI returned only visible anchor text; the underlying URLs were silently dropped during ingestion. With this feature enabled, hyperlinks found in PDFs (e.g. links to fund pages, regulatory documents, or external data sources) are preserved and merged back into the ingested content, making navigable references available for search and AI chat responses.
Note: This feature is behind the FEATURE_FLAG_ENABLE_MDI_PDF_HYPERLINK_EXTRACTION_UN_19716 feature flag, is disabled by default, and can be activated by upon request.
🐞 Bug Fixes
User Word Ingestion: Fixed .DOCX ingestion to ensure tables within Word fallback content are preserved during PDF conversion when using the “Ingest Word as PDF” option.
Agentic Ingestion
🚀 What’s New (Business/End-User Focus)
Admin User BETA Agentic Image Content Extraction: Agentic Image Content Extraction is now available as a Beta feature. It enriches the standard Microsoft Document Intelligence PDF ingestion flow by extracting searchable text from charts, diagrams, figures, and other visual content inside PDFs.
This feature was first introduced as Experimental in Release 2026.16. More details are available in the documentation:
for Knowledge Base users, see Agentic Image Content Extraction in Knowledge Base
for Admins configuring the feature, see Agentic Image Content Extraction
for Developers configuring ingestion programmatically, see Agentic Image Content Extraction Configuration
for Infrastructure Admins
Note: Agentic PDF Document Extraction via CUSTOM_SINGLE_PAGE_API / Custom API is deprecated and will be removed on 31.07.2026 when Image Content Extraction goes GA. New configurations should use Image Content Extraction with pdfReadMode: DOC_INTELLIGENCE_DEFAULT and the new Image Content Extraction ingestion configuration.
Admin User BETA Agentic Metadata Extraction: Agentic Metadata Extraction is now available as a Beta feature. It allows to extract custom Metadata automatically with a LLM based on a given JSON schema.
More details are available in the documentation:
for Admins configuring the feature, see Agentic Metadata-Extraction .
for Infrastructure Admins:
Agentic Table
🚀 What’s New (Business/End-User Focus)
User BETA Commenting in the Agentic Table: Users can now leave inline comments on cells in the Agentic Table, with threaded replies, @mentions, resolve/re-open and a dedicated side panel for triage. This feature is turned off by default but can be enabled by setting the FEATURE_FLAG_AGENTIC_TABLE_COMMENTS_UN_17606 feature flag to true.
More information can be found in our public documentation: Unique AI Table
🔧 Improvements
Admin Strip reference markers from excel export: Excel exports of an Agentic Table can now be configured to remove inline reference markers (e.g. [1], [2]) from the answer column, so the exported spreadsheet contains clean answers. The behaviour is controlled by a new per-space strip_references_from_export flag (default false, so existing exports are unchanged). Configuration details can be found in the newly added xlsx generator config section.
🐞 Bug Fixes
User "View Files (n)" count mismatch: Fixed an issue where the View Files (n) button on an Agentic Table sheet showed a higher count than the number of files actually listed in the dropdown (e.g. "View Files (5)" opening to only 3 files). The count and the listed files are now consistent.
User Row ordering error after adding 200+ rows: Fixed a sequence error that could occur in the Agentic Table after adding 200+ rows in a single batch, causing rows to appear in the wrong order. Large batches now keep the original row order.
User Re-run skipped when no files are attached to the sheet: Fixed an issue where clicking Re-run on an Agentic Table sheet with no question files and no source files attached silently did nothing instead of reprocessing every row. Re-run now always reprocesses every row, regardless of which (if any) files are currently attached.
User Code-style background on tickers, indices and emails: Inline text such as ticker symbols, index names and email addresses previously rendered with a heavy, dark code-block background that overpowered the surrounding answer. The inline-code background has been lightened so these items remain distinct but are now comfortable to read.
Unique AI
🚀 What’s New (Business/End-User Focus)
Admin User BETA Skills Capability for Unique AI: Unique AI now supports the Skills capability, allowing task-specific instructions to be loaded only when relevant instead of permanently expanding the space prompt. Users can invoke skills automatically, through the chat input menu, or with /skill-name; see Skills in Unique AI Chat . Administrators can create SKILL.md files, upload them to the Knowledge Base, and configure which skills are available per space; see Skills . The main Unique AI Chat and Space Management pages now link to the new guidance. Required Feature Flag for Skills activation FEATURE_FLAG_ENABLE_SKILLS_UN_20042 .
Selection of skills
Triggering of skills in chat
Skills activation in space configuration
Admin User BETA Web Search Query Review: The new Web Search Query Review (Elicitation) feature introduces a user approval step before any web search query is executed. When enabled, all generated search queries must be explicitly approved by the end user, helping prevent unintended exposure of sensitive or proprietary information while increasing transparency and control over external tool usage. Documentation Web Search Query Review (Elicitation) .
🔧 Improvements
User Improved Handling for Non-Ingested Uploaded Files: Fixed an issue where uploaded files configured with SKIP_INGESTION or SKIP_EXCEL_INGESTION were completely invisible to the LLM. Previously, these files were filtered out before reaching the model context, preventing the assistant from acknowledging their existence or informing users that the files were uploaded but not directly searchable. The system now properly exposes links of non-ingested uploads to the LLM, improving transparency.