Failure feedback in the agent UI - when a NextGen step fails or has to retry, the agent now surfaces what happened inline. Failed steps appear with a recovering (orange) or terminal (red) state, and a "What happened?" accordion explains the failure in plain language. The Actions & tasks block can also be collapsed.
Document Analyst card unified - the separate Doc Analyst and KB Doc Analyst cards have been merged into a single Document Analyst card. In the Configure modal, you can now choose whether the agent analyzes user-attached files, the agent's knowledge base documents, or both. No functionality change - just less to set up. Applies to both legacy and NextGen agents.
Vision processing for image pages in the knowledge base - when Document Analyst targets a knowledge-base page that's an image, the agent now runs a fresh vision pass on the original asset (with the user's question and conversation context) instead of relying on stored OCR. Unlocks deep, in-question analysis of complex schematics, diagrams, and other image-heavy documents stored in the agent's knowledge base.
TODO.md no longer treated as a real file - the planner and replanner now recognize the [**TODO.md**] marker (used by the UI to render a clickable to-do chip) as a UI element, so the agent no longer wastes steps trying to look it up in the knowledge base.
🛠 Improvements
PDF Viewer demo from Citations settings - a new "Play PDF viewer demo" action in Chatbot Settings → Citations opens an in-context preview of the PDF citations experience.
Persona character limit shown live - the persona editor now shows the over-limit error as you type and disables Save until you're back under the limit, instead of waiting for a server validation error on save.
Cleaner Ask page title - the page title now reads Ask • [Agent Name] (centered dot separator) and no longer duplicates the agent name in the heading.
End-user conversation history is smoother - if a teammate starts chatting as a guest and then logs in, those earlier conversations now carry over to their account. The history panel slides in cleanly and stays open when you switch between conversations, and subdomain share-links now keep the history panel accessible when history is enabled.
Hybrid search auto-enables with Advanced Numeric Search - turning on Advanced Numeric Search at agent creation (or via the patch endpoint) now automatically enables hybrid search, since Advanced Numeric Search depends on it.
Claude 3.5 Sonnet has been retired. All agents previously using Claude 3.5 Sonnet have been automatically migrated to Claude 4.6 Sonnet. No action needed.
🔌 API & MCP
API reference rewrite across Tasks, Sources, Pages, and Conversations - cleaner descriptions, consistent NextGen terminology, per-value enum documentation for agent_capability, crawl_status, index_status, data_refresh_frequency, and format, plus accurate behaviour for poll-once-then-404 task results, file size limits on POST /sources, and the labels[] filter syntax on conversation send-message.
is_advanced_numeric_search removed from the public API - this setting is web-only by design. It is no longer accepted on POST /api/v1/projects or exposed in OpenAPI schemas.
🐞 Bug Fixes
Fixed the issue where the Smart Tasks Configure modal save button silently did nothing. Settings now save and the success toast appears.
Fixed the issue where editing an already-enabled tool's configuration (e.g. tweaking a custom action) returned a "max enabled tools" error if you were at your plan's limit. The limit now only applies when actually enabling a new tool.
Fixed the Confluence spaces selector showing a broken-image artifact when an external space icon failed to load. The fallback stack icon now renders cleanly.
Fixed Multi-Agent appearing in the Quick Start section for brand-new users. Multi-Agent only makes sense when you have existing agents, so it's now hidden until you do.
Fixed the post-subscription redirect sending active teams to the demo-agent screen after re-subscribing. Active teams now stay on the normal billing flow; only truly empty dashboards get the onboarding redirect.
Fixed the Conversation Insights agentic-actions filter showing raw internal keys (e.g. expensive_model_claude-4.6-sonnet-thinking) as filter labels. Expensive-model usage now displays as Added Cost: Claude 4.6 Sonnet Reasoning.
Fixed the NextGen file manager auto-opening on mobile as soon as the first file in a conversation was generated. It no longer pops open on small screens.
Fixed an edge case on Claude where the NextGen agent could re-process the user's message instead of progressing through the plan.
Smart Tasks billing now tracks actual agent work - the credit cost of each Smart Tasks run scales with how much the agent actually read and produced, instead of a flat per-tool-call charge. Lightweight runs cost less; heavy runs (large files, long generated code) cost what they should.
Smart Tasks breakdown in the cost popup - the NextGen badge popup now shows Smart Tasks as a collapsible row that opens into three sub-rows: Task setup, File parsing, and Code generation. Users can see exactly which stage drove the cost on any Smart Tasks run.
Cleaner dashboard rendering - switched the dashboard chart library to a lighter, canvas-based renderer. Charts no longer clip inside the chat bubble, legends sit cleanly below the plot, and resize behavior is more predictable. Dark-background regressions are gone; charts default to a clean white background.
Tighter agent/UI naming for dashboards - the agent now refers to dashboards by their title in chat, matching what shows up in the file manager.
First-time NextGen enable preserves NextGen defaults - turning NextGen on for the first time on an existing agent no longer reverts max_actions_per_query to 1 or carries over stale legacy settings. The optimal NextGen preset is applied cleanly.
Fewer duplicated steps and empty-prompt errors on Anthropic models - hotfix for NextGen agents using Claude where the agent could repeat execution messages or send an empty prompt mid-task.
Deep document analysis fan-out fix - resolved a race condition where parallel deep-dive analysis on multiple files could miss one of the files. Deep dive now consistently fans out to every requested document.
🐞 Bug Fixes
Fixed the issue where some columns in uploaded CSV files (e.g. url, post_date) were silently ignored on newer agents. All column data is now preserved end-to-end during ingestion.
Multi-Agents are now available - combine several specialized agents under one chat, and let end-users pick which one they talk to.
New Multi-Agent card on the Create page.
Users can build a Multi-Agent by attaching child agents to it.
Child agents can be added or removed at any time on the Multi-Agent's build page.
Rule of thumb: Multi-Agent controls UI, child controls AI.
When interacting with a Multi-Agent for the first time, the end-user is asked which child they want to chat with.
Users can switch the child they're talking to at any time, even mid-conversation.
Available on all deployments except SGE.
Available on Premium and Enterprise.
🤖 NextGen Updates
Generate Dashboard action - NextGen agents can now produce interactive dashboards (dynamic charts, filters, formulas) from whatever data is in the conversation: uploaded files, knowledge-base results, or numbers computed by the agent. The planner invokes this automatically when data is best presented visually, or on direct user request. Controlled by a new permission under Smart Tasks.
Side-panel artifact viewer - clicking a generated file in the file manager now opens it in a preview panel alongside chat. Supports markdown, plain text, images, and live dashboards. No download needed.
Full cost transparency on every NextGen message - the NextGen badge in the bottom-right of any response opens a popup with a full breakdown of the query steps and the exact credit cost. Same breakdown visible on Customer Intelligence pages.
Extended file type support for Smart Tasks uploads - when Smart Tasks Read Files is enabled, end-users can upload xlsx, xls, csv, tsv, json, md, and pptx alongside the usual document formats, so the agent can read and analyze them directly.
Conversation memory - NextGen conversations automatically retain earlier context: decisions, files, numbers, preferences - even 20+ messages later. Scoped per conversation; new conversations start fresh.
Smarter, more reliable agent behavior:
Smart Tasks carries memory of previous code-execution steps across the conversation and always returns a final summary.
Deeper document analysis works reliably again after a round of fixes.
The agent no longer hallucinates file names or entity references when re-planning mid-task.
Re-planning is more efficient, so complex NextGen messages use fewer credits.
NextGen polish:
The NextGen badge now has a proper hover state.
NextGen agents with no knowledge base can now hold a conversation, and past conversations are visible.
Deep document analysis is skipped automatically when the agent has no indexed pages.
Toggling NextGen on or off now saves cleanly, with no stale settings left behind.
The API error returned when NextGen is disabled now names the exact setting to enable.
When NextGen is off, Knowledge Base appears in the locked tools section alongside the other NextGen-only tools.
🛠 Improvements
Image citation limit per response - new setting in Personalize > Citations controls how many image or OpenGraph citations are rendered inline per response. Options: 1, 2, 3, or All. Extras beyond the limit show as regular numbered citations - nothing is lost, just less clutter on image-heavy answers.
Model selector upsell - users without model-selection access can now open the dropdown and see every model listed with a lock icon. Clicking any locked model opens a Contact Sales upsell explaining that Enterprise unlocks frontier models from OpenAI, Anthropic, and Google.
MCP / custom actions limit configurable per Enterprise plan - the per-agent cap on MCP and custom actions is now adjustable per Enterprise plan on request. Default of 3 stays on all other plans.
Instant Viewer marked as legacy - the Instant Viewer integration card now shows a Legacy badge with a tooltip pointing to the newer PDF Citations feature in Agent Settings.
OneNote files display as files in the SharePoint and OneDrive file pickers, not folders.
🔌 API & MCP
Page search - GET /api/v1/projects/{projectId}/pages now accepts an optional search query parameter that filters pages by URL or filename (case-insensitive substring match). OpenAPI spec updated.
Clearer NextGen API error - when use_nextgen_agentic_flow is disabled, POST /projects/{id}/tasks and POST /projects/{id}/tasks/{taskId}/messages now return a message naming the exact setting to enable.
🐞 Bug Fixes
Fixed the issue where the Ask Me Anything conversation would unexpectedly scroll to the bottom when image citations shuffled in an older message.
Fixed the issue where HTML entities (e.g. ', &) appeared as literal text in the conversation list sidebar and on Analyze and Ask widgets.
Fixed the issue where Instant Viewer opened for non-PDF inline citations (it should only open for confirmed PDFs; other citations now open as external links).
Fixed the issue where the PDF preview panel could carry stale state across conversations, showing the wrong document or highlight when switching chats.
Fixed the issue where the Instant Viewer legacy badge tooltip incorrectly showed a "Chat to learn more" button.
Fixed the issue where the CustomGPT Copilot send button stayed active on empty or whitespace-only input, and the input's resting and active states were inverted.
Fixed the Labels editor: the Delete All Labels button no longer throws a route error, and the add-label form is no longer artificially capped in width.
Fixed the Copied! tooltip in copy-to-clipboard controls so the confirmation appears in the right place.
End-user Conversation History (Premium) - personal conversation history is now available on all agent deployments (except SGE).
Team members: conversations sync across devices and browsers, persist indefinitely unless the user deletes them.
Guests: conversations save to browser storage - lost if the cache is cleared or the device changes.
Vimeo integration is now live as a native agent source.
Ingest Vimeo channels, user profiles, showcases, and individual videos.
Citations include timestamps that deep-link into the exact moment in the video.
Auto-sync on Premium plans.
Video must have transcription available for sync to work.
🤖 NextGen Updates
API support for NextGen agents - full API parity with UI.
Tasks: GET /projects/{id}/tasks (list), POST /projects/{id}/tasks (create), GET /projects/{id}/tasks/{taskId} (poll), PUT /projects/{id}/tasks/{taskId} (rename), DELETE /projects/{id}/tasks/{taskId} (delete), GET /projects/{id}/tasks/{taskId}/export (CSV/PDF export).
Messages: POST /projects/{id}/tasks/{taskId}/messages (send follow-up), GET /projects/{id}/tasks/{taskId}/messages (list), GET /projects/{id}/tasks/{taskId}/messages/{promptHistoryId} (fetch single).
Files: GET /projects/{id}/tasks/{taskId}/messages/{promptHistoryId}/files (list), GET …/files/{fileId}/download (download artifact), DELETE …/files/{fileId} (delete artifact).
NextGen toggle also available via PUT /projects/{id}/settings using use_nextgen_agentic_flow.
Polling model: GET /tasks/{taskId} returns processing + chunk events while the session runs, then completed + remaining events on first read after the session ends. Rate-limited via a separate tasks.poll bucket so polling doesn't consume the general API quota.
NextGen badge with queries breakdown - NextGen messages now show a unified bolt + "NextGen" badge. Hovering the badge reveals a popup with per-tool task counts and the actual query cost. Also visible on Customer Intelligence pages.
Chat with a NextGen agent that has no knowledge base - NextGen agents that rely only on tools (web search, custom actions, MCP) with no indexed documents can now chat. Previously these agents were blocked by a "add knowledge to your agent" gate.
Actions page polish on first NextGen activation - five sensible default actions (Knowledge Base, Smart Tasks, Doc Analyst, KB Doc Analyst, Web Search) are auto-enabled the first time NextGen is turned on. Incompatible actions render fully greyed out below the NextGen banner so what's available is obvious at a glance.
Knowledge Base Doc Analyst: multi-page search - Doc Analyst can now search and analyze multiple pages in one query instead of one at a time.
Agent accuracy and reliability improvements:
Cleaner response formatting - agents no longer occasionally leak stray code or markup into chat answers.
More reliable coding tasks when using Anthropic models (no more hangs).
Smarter tool selection - the agent is more aware of which tools are enabled and passes parameters through them more reliably.
More accurate answer classification feeding into Customer Intelligence metrics.
NextGen bug fixes:
Knowledge Base Doc Analyst and hybrid search no longer fail intermittently.
GPT-4.1 removed as a coding-task model (it was causing failures).
Max tasks per query is no longer reset to the plan default when NextGen is re-toggled - the default is set only on the very first activation.
🛠 Improvements
Customer Intelligence: Agentic Action filter - filter the Analytics Overview by prompts that used a specific agentic action.
Agent Initiative: don't prompt on mobile (Premium) - new Live Chat setting prevents agent-initiative starter-question prompting on mobile devices.
Azure provider flexibility for GPT models - several GPT models (GPT-4.1, GPT-4.1-mini, GPT-4o-mini, GPT-5.1, GPT-5.2, o4-mini variants) no longer hardcode OpenAI as the provider, so Azure routing works when selected at the project or plan level.
Chat input placeholder fade - the Ask Me Anything input now fades long placeholder text at the right edge instead of awkwardly clipping or causing horizontal scroll.
🔌 API & MCP
Multi-file upload - POST /projects/{id}/sources (plus POST /projects and PUT /projects/{id}) now accepts a files[] array of up to 50 files per request, matching the UI. The existing singular file input remains as a legacy alias for backward compatibility.
Source Labeling: AND/OR label grouping - nested AND/OR logic on source labels is now supported via API, letting callers express complex access rules.
🐞 Bug Fixes
Fixed the issue where Source Labeling with grouped labels returned "I'm sorry" responses on pages the user had legitimate access to.
Fixed the issue where users with zero source labels could incorrectly see all pages when Source Labeling was enabled, instead of the expected empty-set enforcement.
Fixed the issue where SharePoint and OneDrive folder changes failed with Undefined array key "cTag" errors during sync.
Fixed the issue where creating a new conversation did not appear in the sidebar until page refresh when the project's User Conversation History setting was Hidden or Team-Members-only.
Fixed the issue where YouTube citations lost their timestamp deep link when the video was referenced via a starter question.
Fixed the issue where large photo-heavy PDFs hung during upload.
Fixed the issue where Azure-hosted reasoning and GPT-5 deployments returned Unsupported parameter: 'max_tokens' errors.
Knowledge Base Doc Analyst - a new agentic action (formerly known as Deep Dive) is now live.
When a user asks for more details about a specific source, the agent automatically finds that document and runs the Doc Analyst analysis flow on it.
Works on all source types and document formats.
Available settings: Improve answer relevance (Cohere reranker) and allowed document length.
Agentic cost matches regular Doc Analyst. On Enterprise plans, the agentic cost can be configured independently.
Known limitations: one document at a time, no schematic support yet, translation requests may be refused. Improvements coming soon.
NextGen Agents (Closed Beta) - new tools and self-service configuration on the Actions page:
New Knowledge Base (RAG) tool with an accuracy algorithm setting.
New Smart Tasks (Code Execution) tool with three configurable permissions: Read Files, Create Files, and Make External Connections.
Actions page now splits tools into NextGen-compatible and NextGen-incompatible groups with a clear visual divider and banner.
Max tasks per query increased from 10 to 15.
Getting started guides published for customers in the beta.
🛠 Improvements
Microsoft Entra ID on SSO page - added the Entra logo to the SSO configuration page with a link to the new setup guide.
PDF processing reliability - the AI vision models GPT-5 and GPT-5-mini have been retired and migrated to GPT-5.4 mini, resolving timeout issues on PDF + vision workflows. GPT-5.4 mini is ~30% faster than the previous default.
Integration file picker redesign - updated the file and space picker UI for SharePoint Docs, SharePoint Sites, OneDrive, and Confluence with an improved checkbox + icon + label layout.
Customer Intelligence: clickable rows - prompt summaries and missing content items on the Analytics Overview page are now clickable, navigating directly to the full conversation.
Default prompt display - the AMA chat input now shows the full configured default prompt instead of truncating it. A non-blocking warning appears in settings when the prompt exceeds the recommended display length.
Minor copy and link updates - refreshed default prompts, contact sales link, documentation links, and Knowledge Base Doc Analyst and NextGen labels across the app.
🐞 Bug Fixes
Fixed the issue where Standard plan customers with query addons were seeing inflated query limits (e.g., 6,000-13,000 instead of 3,500) due to addons from previous billing cycles incorrectly entering the proration formula.
Fixed the issue where updating page metadata from the UI marked all fields as user-edited even when values were unchanged, blocking subsequent auto-sync updates from integrations and webhooks.
Fixed the issue where SharePoint and OneDrive files that were moved kept their old URL and title in citation links, sending users to the wrong location.
Fixed the issue where API citations were not validated within the correct project context, potentially exposing citations from other projects in streaming and non-streaming responses.
Fixed sidebar icons being partially clipped when the sidebar was collapsed on project pages.
Fixed "Connected agents" label capitalization to "Connected Agents".
Fixed the issue where Gemini model responses lost whitespace in streamed chunks.
Fixed the agent's fallback ("sorry") message not appearing when using tool-augmented mode (File Analyst or Web Search) and no relevant sources were found.
Fixed large PDF uploads crashing workers due to excessive memory usage during text extraction.
Fixed remote tabular files (.xlsx, .ods, .tsv) added via URL failing to crawl.
NextGen Agents (Closed Beta) - self-service configuration is now available for the advanced agentic orchestration engine introduced last week.
New toggle on the Intelligence tab to enable NextGen for any agent.
NextGen (BETA) badge is visible across all pages when the feature is enabled on an agent.
New Max tasks per query setting controls how many tasks the agent can perform in a single query. Starting value is 10.
If the agent hits the task limit, users can continue in the next query.
🛠 Improvements
Source Labeling editor (Closed Beta) - a new Labels tab is now available on the Build page for accounts with Source Labeling enabled:
Builders can see which labels are assigned to which documents, and add or remove labels directly.
New Manage Labels modal allows creating and deleting labels at the agent level.
Simplified citation settings - removed the separate "Display Citations / Show Citation UI" toggle. Citation visibility is now controlled only through citation type toggles, eliminating a confusing extra step.
Updated CustomGPT.ai logo - the new brand identity has been applied across the entire application, including all logo variants and favicons.
Master Agents child picker redesign - the child agent picker in Master Agents has been updated to match the new agent UI design, with improved color theming and styling.
PDF Viewer settings guidance - added inline help tooltips to PDF Citations settings explaining how the built-in PDF viewer and citation highlighting work.
🔌 API & MCP
OpenAPI schema improvements - a series of updates to the API reference:
Added human-readable enum value descriptions for enable_citations, response_source, citations_view_type, chatbot_model, and agent_capability.
Added step-by-step authentication guide to BearerTokenSecurityScheme.
Added request body examples for Create Agent, Send Message, and Update Settings endpoints.
Improved is_chat_active description with polling guidance for new agent readiness.
Updated parameter descriptions across all major endpoints (Agents, Pages, Conversations, Messages, Labels, Sources, Reports, Settings).
🐞 Bug Fixes
Fixed PDF Viewer not loading on Share Link deployments when using a custom subdomain.
Fixed URLs being stripped from messages sent through the Slack integration.
Fixed citations not appearing in responses when the Drive Conversions agentic action was active.
Fixed the Deep Dive credit multiplier being applied to all queries when the feature was enabled, even if Deep Dive was not actually triggered for that query. Usage stats now correctly show the number of documents analyzed.
Advanced Agentic Orchestration (Closed Beta) - a major upgrade to how agents handle complex, multi-step tasks. Available on request for selected customers.
Agent can break down complex requests into steps, plan an approach, execute actions, and deliver a final answer - all in one conversation.
New File Manager panel lets users upload files and view AI-generated documents directly in the chat.
Cross-session memory - agent retains context from previous conversations, so users don't need to repeat themselves.
Real-time streaming responses during multi-step tasks, with live progress updates as the agent works through each step.
Context-Rich Starter Questions Management - builders can now manually manage the context-rich starter questions generated by Webpage Awareness:
For every page saved by Webpage Awareness, builders can edit both the display message (what visitors see) and the agent message (what gets sent to the agent when clicked).
Manually edited questions are not removed after 30 days, unlike auto-generated ones.
URLs in the management modal are now cleaned up (no protocol/www prefix) and truncated for readability.
Available at Settings > Starter Questions > Manage (per saved webpage).
CLI Authentication - the CustomGPT CLI tool can now authenticate via a browser-based OAuth flow. Running the CLI opens a browser window for authorization, and a token is automatically issued back.
🛠 Improvements
PDF Viewer & Citations overhaul - a major update across citations and PDF viewing:
PDF Viewer is now available on all deployments: Share Link, Live Chat, Embed, and Website Copilot. Previously limited to Instant Viewer deployment only. Configure at Personalize > Citations. Enterprise-only.
Citations page restructured with a new master Enable Citations toggle and clear settings hierarchy - sub-settings only appear when their parent is enabled.
OpenGraph citations redesigned - raw URL overlays replaced with structured, fully-clickable resource cards showing document title, source domain, and favicon.
PDF Viewer styling improved - updated navigation, layout, and mobile responsiveness.
WCAG accessibility compliance improved for classic citations.
Instant Viewer deployment remains as a legacy option but is strongly discouraged. Existing agents with Instant Viewer configured have their settings transferred automatically.
Product Lookup role is now available on Standard plans. Previously required a Premium subscription.
Two-factor authentication UX upgrade - the TOTP verification page now uses a modern 6-field OTP input with automatic submission when the code is fully entered or pasted. Recovery code login remains fully supported.
SSO logout improvements - SSO-authenticated users can now properly log out, including guests from embedded agents. Fixed redirect URLs, tenant identification, and handling for cases where no logout URL is configured.
"Keep PDFs for viewing" now enabled on existing SharePoint and OneDrive sources. The setting was previously only available for newly connected sources.
SharePoint and OneDrive folder names now sync automatically when renamed in the source.
On mobile, the browser status bar now matches the chat header color when the chatbot is opened in fullscreen mode.
🐞 Bug Fixes
Fixed an issue where the in-app support copilot was not appearing on the dashboard.
Fixed the custom "I don't know" message configuration being hidden after the citations page update. The input field is now restored in Chatbot Settings.
Fixed source label filtering not being applied correctly, which could cause over-permissive access or incorrect "sorry" responses for agents using source labels (closed beta).
Fixed an issue where tasks were not displayed when previewing a shared conversation link.
Fixed a SharePoint picker UI glitch where folder contents would disappear when navigating back into a folder.
Fixed an issue where image/vision analysis was returning empty responses approximately 80% of the time.
Available on Optimal, Accuracy, and Understanding modes.
Available to users with access to the model dropdown.
Gemini 3.1 Pro is now available as an AI model option:
Spends one additional query per message.
Available on Optimal, Accuracy, and Understanding modes.
Available to users with access to the model dropdown.
Improved Numeric Search - a major upgrade to how agents handle part numbers, SKUs, and alphanumeric codes:
Retrieval quality for numeric and alphanumeric queries has been significantly improved.
Response times dropped from ~40 seconds to under 4 seconds on large datasets.
To enable: click the Settings button on the Build page and turn on Improved Numeric Search, then re-index (or at least re-index the files where numeric search matters).
New agents can enable the setting during creation. Existing agents that don't enable it continue using the previous numeric search.
Enabling this adds up to 30 seconds of processing time per document during ingestion, which is why it's opt-in rather than automatic.
🛠 Improvements
Notification call-to-action links (e.g., model announcements) now open in a new tab instead of navigating away from the current page.
🔌 API & MCP
Updated OpenAPI schema:
Added realistic example values across all 17 CustomerIntelligence fields.
Added 429 Resource Exhausted error responses to 69 endpoints.
Added streaming documentation with SSE event format and JavaScript client example for POST /chat/completions.
The is_ocr_enabled parameter now accepts boolean values (true/false) in addition to integers (1/0), restoring backward compatibility across all form requests that use this field.
🐞 Bug Fixes
Fixed an issue where the unsaved changes warning in Chatbot Settings was triggered by background polling refreshes, causing false popups even when no settings had been modified.
Fixed an issue where subscribed users who hadn't chatted with any agent yet were locked out of both Crisp live chat and AppWidget support.
Fixed an issue where inline citations could disappear after a conversation loaded, along with autoscroll jumps during streaming and redundant citation loading after streaming completed.
Fixed an issue where streaming failed when Web Search was enabled on an agent.
When using File Upload source with non-searchable PDFs, users can now choose the new AI Vision processing mode alongside traditional OCR.
One page of PDF processed with AI Vision counts as one Vision Credit.
AI Vision understands images within PDFs but does not extract them as separate files.
OCR has been flagged as "not recommended." AI Vision is available on Standard plans and above, while OCR remains Premium-only.
AI Vision cannot currently be used at the same time as regular image vision processing on the same agent. Users who need both should process PDFs and images separately.
Agentic Actions are now available on Standard plans:
Standard users can enable one Agentic Action at a time.
All Actions have higher query costs on Standard compared to Premium.
Some Actions have feature-specific limits on Standard (e.g., lower Document Analyst word limits, fewer Web Search calls per query).
The exact cost for each Action is now displayed on the Limits page based on the user's plan.
🛠 Improvements
Document Analyst quality upgrade - new "Improve answer relevance" toggle:
When enabled, the most relevant parts of uploaded files are prioritized before being sent to the model, leading to better answers.
The agent still uses its full Document Analyst word limit, but content is ordered by relevance rather than position.
If the uploaded file(s) are below the agent's limit, they are read in their entirety regardless.
ON by default for new agents, OFF for existing agents to avoid disrupting current behavior.
The Document Analyst configuration modal has also been refreshed.
New "Copy Button" visibility setting in chatbot settings. When disabled, the button to copy the agent's entire message is hidden from the chat interface.
The "Share" button in the app header has been replaced with a "Publish" button. Functionality is unchanged, but the action is now immediately visible with a label instead of an icon-only button.
Unsaved changes warning: if you try to leave a settings page or close the tab with unsaved changes, the browser will now warn you before navigating away.
Latest Prompts in the Analyze section are now clickable. Clicking a prompt opens a conversation modal showing the full conversation context.
Chat input layout improvements on mobile, including better spacing when the file upload tool is disabled and a more responsive starting screen layout.
Master Agents have been renamed to Multi-Agents across the platform.
🔌 API & MCP
Fixed 500 errors on the /pages/{pageId}/metadata endpoint when a non-numeric pageId was provided. The endpoint now returns a proper validation error.
🐞 Bug Fixes
Fixed an issue where child agents in Multi-Agent setups could answer questions outside their designated scope by referencing conversation history from unrelated queries.
Fixed an issue where the "Verify Response" (Explainability) button did not appear in the chat interface when message copy actions were hidden.
Fixed an issue where the login and registration pages had insufficient minimum height on mobile devices.
Claude 4.6 Sonnet and Claude 4.6 Sonnet Reasoning are now available in the AI model selector.
Claude 4.6 Sonnet Reasoning uses extended thinking to deliver higher-quality responses for complex queries, at a cost of 2 queries per message.
Master Agents are now available on demand:
Create a master agent that connects multiple specialized child agents under one interface.
Each child agent has its own knowledge base and persona, while the master agent controls the shared look and feel.
End-users can switch between child agents using a dropdown, without leaving the conversation.
To enable Master Agents for your account, contact the sales team.
🛠 Improvements
Smarter document-focused search:
When you ask a question about a specific document, the agent now detects this and focuses its search exclusively on that file.
This means more relevant content from the requested document ends up in the agent's context, leading to better answers.
You don't need to provide the exact file name - the agent can figure it out from your question and conversation history.
The feature is always active and can be disabled via Intelligence > General knowledge base awareness > Disabled.
Image processing for website-sourced agents has been upgraded:
When your agent crawls a website, images are now processed with context from the host page (first 500 words).
This results in richer image descriptions - for example, a product image now includes size, price, and properties from the listing page, not just visual features.
New "Selected Folders Overview" button on the SharePoint source page, letting you see at a glance which folders and files you've selected for your agent.
GPT-5 has been retired. All agents using GPT-5 have been automatically migrated to GPT-5.1 Optimal.
Opus 4.5 is now marked as legacy in the model selector. If your agent uses Opus 4.5, consider switching to the newer Opus 4.6 model.
YouTube URL handling improved - the agent now correctly processes YouTube URLs in various formats when provided via sitemap input.
Bulk update and delete is now available for website contexts in Starter Questions management.
SharePoint and OneDrive sync notifications are now handled more reliably.
Minor copy improvements on the SharePoint and related source pages.
🔌 API & MCP
Comprehensive filtering has been added to the Customer Intelligence analytics MCP tool:
15+ filter parameters including user emotions, intents, content sources, stakeholder status, risk metrics, leads, country, language, and date ranges.
AI assistants connected via MCP can now answer complex analytical questions like "Compare frustrated vs positive users in December" or "Which countries have the most unanswered questions?"
Fixed the issue with MCP server sessions expiring after a couple of hours. Sessions now automatically reconnect.
Fixed the integer schema for the leads filter parameter in the OpenAPI spec.
Added enum constraint for the request_source filter in the Customer Intelligence OpenAPI parameters.
🐞 Bug Fixes
Fixed the issue where batch updating documents on the Build > Documents page was not working.
Fixed the issue where document metadata was displaying incorrect URLs in OpenGraph tags.
Fixed the issue where the data source selection could become stuck after dismissing the model configuration dialog.
Fixed the issue where affiliate tracking IDs were not being saved correctly.
Fixed the issue where Customer Intelligence was showing inflated query counts due to incomplete chat entries being included in the results.
