> For the complete documentation index, see [llms.txt](https://docs.maiagent.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.maiagent.ai/release-note/untitled/archives/04.md). # April 2026 ## v2026.04.29 *** This update adds duplicate URL handling strategies for knowledge base crawler page imports, letting users flexibly choose to skip, overwrite, or allow duplicates. It also fixes several issues affecting AI assistant stability, including browser tool operation failures and conversation interruptions caused by specific third-party connectors. ### ✨ New Features * **Knowledge base crawler page import conflict strategies**: When batch importing crawled web pages into a knowledge base, you can now choose how to handle duplicate URLs—skip existing ones, overwrite old content, or allow duplicates to be kept. Before importing, you can also pre-check whether a URL has already been crawled, avoiding unnecessary repeated work. Ideal for teams that need to regularly update web content in their knowledge bases. ### 🛠️ Bug Fixes * **Fixed specific third-party connectors causing complete AI assistant conversation failure**: When using Composio connectors (such as Google Sheets) with AWS Bedrock models, connector tool names exceeding the platform limit caused 100% of that AI assistant's conversations to fail. After the fix, the system automatically handles tool name length, and connector functionality works normally again. * **Fixed occasional AI assistant browser tool operation failures**: When the AI assistant used the browser tool to perform click or scroll operations, parameter format issues occasionally caused the operation to fail and interrupt the task. After the fix, browser tool operation stability has returned to normal. * **Fixed the "View Source" button being obscured in knowledge base document preview**: In the text node preview dialog for knowledge base documents, the "View Source" button overlapped with the fullscreen and close buttons in the upper right corner, making it impossible to click correctly. After the fix, the buttons are properly arranged and no longer obscure each other. ## v2026.04.28 *** ### ✨ New Features * **Gemini 3.1 Flash image generation tool**: MaiGPT's AI assistants gain a fast image generation option. Compared to the existing Gemini 3 Pro Image tool, the Flash version generates images faster, making it ideal for conversation scenarios that require instant image creation. Supports low, standard, and high quality levels for the AI to select automatically. ### 😊 User Experience and Interface Improvements * **Simplified AI assistant monitoring metrics**: The monitoring table previously scattered more than 20 quality score and processing time columns; these are now consolidated into two main columns: "Quality Score" and "Processing Time". Click to expand and view the complete sub-metrics with explanations, greatly reducing the reading burden. Rarely used advanced columns are now hidden by default, but users can still enable them from the table settings. * **Improved AI assistant knowledge base utilization**: AI assistants linked to knowledge bases now more proactively use knowledge base content to answer questions, reducing the cases where users had to explicitly request "search the knowledge base" to get relevant answers. ### 🛠️ Bug Fixes * **Fixed being unable to view source code while Canvas is generating**: While Canvas is generating content, clicking the "View Source" button now shows the streaming source code in real time. Previously, switching to source code mode was still covered by the "Compiling" loading animation. * **Fixed layout issues in the contact details panel**: Multiple visual defects in the contact "Details" panel on conversation platforms have been corrected, including gray blocks appearing in the button area, inconsistent component spacing, and insufficient text contrast. * **Fixed the bottom input bar being obscured on iPhone Safari**: When browsing the AI assistant marketplace (MaiGPT) on iPhone Safari, the message input area at the bottom is no longer obscured by the browser toolbar, so messages can be entered and sent normally. * **Fixed the embedded chat widget button being obscured**: For the MaiAgent chat widget embedded in websites, after shrinking the browser window and reopening the widget, the chat dialog no longer covers the launcher button in the lower right corner. The dialog automatically resizes to avoid the button and returns to its original size when the window is enlarged. * **Fixed multiple stability issues in the admin console**: Fixed intermittent errors on several admin console pages: an error when the member invitation list loaded before the table was ready, an exception triggered by leaving the page after uploading knowledge base files, and a crash when clearing the date range in usage statistics. * **Fixed the mistranslation of the "Faithfulness" quality score**: The Faithfulness score label in AI assistant monitoring has been corrected in both Traditional and Simplified Chinese from the mistranslated "honesty score" to the industry-standard term "faithfulness score", consistent with the official RAGAS Chinese documentation. ## v2026.04.27 *** ### ✨ New Features * **AI Copilot agent suggestion system**: Customer service agents can receive AI-generated reply suggestions in real time during conversations, adopting them directly or using them as a reference when composing replies. The Copilot feature can be enabled or disabled in the settings page of each conversation platform (Web Chat, LINE, Messenger, WhatsApp, Slack, Telegram, and more). * **Conversation Credit consumption summary sidebar**: View a conversation's total Credit consumption, category breakdown (vectorization, LLM input/output, tools, etc.), and statistics (turn count, average, median) directly in the conversation details sidebar, with CSV report export. * **GPT Image 2 image generation support**: The image generation tool adds the OpenAI GPT Image 2 model, supporting up to 4K (3840×2160) resolution, transparent backgrounds, and multi-turn reference images. MaiGPT internal assistants come equipped with this tool automatically. * **MaiGPT internal assistant time-related questions**: The MaiGPT internal assistant can now answer time-related questions such as "What's today's date?" and "What time is it now?". Newly created MaiGPT conversation platforms have this enabled by default, and existing ones are automatically updated to enable it. * **Automatic session continuation for embedded conversation pages**: Embedded conversation pages add an automatic Token refresh mechanism, so long sessions are no longer interrupted by login expiration and users do not need to log in again. ### 😊 User Experience and Interface Improvements * **Improved AI assistant schedule configuration interface**: The AI assistant schedule form adds a visual schedule builder that lets you directly select the execution frequency and time, replacing manual format entry. * **Refined conversation platform list filter tab styling**: The "Mine / Unassigned / All" filter tabs on the conversation list have changed from an underline style to a pill-shaped segmented control, with tab widths adapting to content so text is no longer truncated. * **Answers preserved when switching between survey questions**: When filling out multi-question surveys, custom input values entered previously are no longer lost after returning to a prior question. The selection and focus visual styles of options have also been unified, making it easier to distinguish the currently focused state from the selected state. * **Conversation platform sidebar search bar alignment and scrollbar refinement**: Unified the search bar spacing and alignment of the conversation platform and conversation lists, and changed the scrollbar to a lightweight style that only appears on mouse hover, no longer occupying fixed space. * **Credit admin console protection against accidental changes**: Credit transaction records and top-up records are now read-only, preventing business staff from accidentally modifying or deleting them and causing accounting inconsistencies. The quick-edit feature for billing parameters has been removed; edits must be made from the detail page to avoid accidental changes. ### 🛠️ Bug Fixes * **Fixed being unable to view source code during Canvas generation**: While Canvas is streaming generated content, clicking the "View Source" button now correctly shows the live-streamed source code, no longer covered by the "Compiling" loading overlay. * **Fixed the input area being obscured by the browser toolbar on mobile devices**: On mobile browsers such as iOS Safari, the bottom input bar is no longer obscured by the browser's dynamic toolbar, so users can enter and send messages normally. * **Fixed contact tag edits not syncing to all conversations**: After editing a contact's tags in the conversation details panel, all conversation list items for the same contact now update their tags in sync, instead of only the current conversation. * **Fixed sorting jumps on the role permission management page**: On the role permission management page, after modifying an individual permission and reloading the list, items no longer jump to a different page. All permission-related resources are uniformly sorted by creation time in descending order. * **Fixed conversation platform settings pages failing to open**: The settings pages for the LINE, Telegram, TeamBot (Azure Bot), and Slack conversation platforms work normally again, no longer showing a blank screen when opened. * **Fixed permission errors on Agent marketplace user conversation pages**: When Agent marketplace and MaiGPT users enter conversation pages, they no longer fail silently due to insufficient permissions. The system automatically sends requests with the correct permission scope. * **Fixed handling of corrupted OCR document recognition models**: A corrupted OCR text recognition model file no longer makes the feature completely unusable; the system provides a clear error message. Model files are now pre-downloaded, avoiding failures from downloading externally at runtime. ## v2026.04.23 *** ### ✨ New Features * **Edit contact tags directly in the conversation details panel**: Customer service agents can add or remove contact tags directly in the right-hand conversation details panel, without opening the "Transfer/Assign" dialog. The tag selector supports search, and the empty state provides a "Manage Tags" quick navigation shortcut. * **AI assistant in-conversation message search**: AI assistants gain the ability to search conversation messages, retrieving historical messages within a conversation along with their surrounding context. Ideal for scenarios where the AI assistant needs to reference past conversation content to answer questions. * **Browser tool**: AI assistants gain web browsing capability, directly accessing web page content to obtain real-time information. Users can enable this feature for AI assistants in the tool settings. ### 😊 User Experience and Interface Improvements * **Visual improvements for conversation list filter tabs**: The "Mine/Unassigned/All" filter tabs on the conversation platform's "All Conversations" list have changed from an underline style to a pill-shaped segmented control, eliminating clashes with the page's color scheme. Tab widths adapt to content so longer labels like "Unassigned" are no longer truncated, and unread count badges are now integrated into the tab buttons. ### 🛠️ Bug Fixes * **Fixed system errors when conversation platforms were configured with invalid roles**: When an administrator specifies a role that does not exist or does not belong to the organization in conversation platform settings, the system no longer returns an incomprehensible error message; instead, it clearly indicates which roles are invalid. * **Fixed missing notifications when knowledge base document parsing fails**: When uploading documents to a knowledge base, if an error occurs during parsing (such as an invalid API key or unsupported file format), the system no longer silently shows success without creating any index; it now clearly displays the reason for the failure. * **Fixed structured output failures with Bedrock models**: AI assistants using AWS Bedrock models no longer encounter the "the last message must be a user message" error when structured output is enabled. After the fix, structured output works normally across all model providers, including Bedrock, OpenAI, and Azure. * **Improved knowledge base embedding service stability**: The knowledge base vector embedding service now automatically retries when it encounters transient network errors (such as 502, 503, and 504), reducing knowledge base indexing failures caused by brief network fluctuations. ## v2026.04.22 *** ### ✨ New Features * **Q\&A interaction keyboard navigation and ranking question type**: The Web Chat Q\&A interaction area adds full keyboard support, including arrow keys to switch options, Enter to select and submit, Space to toggle multi-select, and Esc to skip. A new ranking question type supports mouse dragging and Shift+arrow key keyboard sorting. Single-choice questions now select directly with Enter, no longer requiring an extra click on the submit button. \[maiagent-django] **\[⚠️ Manual review recommended: frontend integration pending confirmation]** * **Automatic WhatsApp Token expiration notifications**: The system automatically detects whether the WhatsApp connection Token has expired, and when it does, creates an organization announcement to notify administrators to handle it promptly, preventing the WhatsApp channel from failing to send and receive messages due to Token expiration. ### 😊 User Experience and Interface Improvements * **AgentOps evaluation record processing timing analysis**: AgentOps evaluation records add a "Pre-processing Time" metric, showing the time spent on system initialization, request validation, and Agent workflow preparation, letting administrators analyze the sources of response latency more precisely. ### 🛠️ Bug Fixes * **Fixed document OCR parsing failures with self-hosted AI models**: When using self-hosted AI models such as vLLM or ollama for document OCR parsing, the API URL was incorrectly assembled with a duplicated path, causing 404 errors and parsing failures. After the fix, URL assembly is correct for all AI model providers. * **Fixed timeout errors when updating knowledge base tags**: When updating knowledge base tags, a timeout handling error occurred in certain situations. The fix also adds a synchronous update method, improving the stability of tag updates. * **Fixed long conversation analysis tasks failing due to timeouts**: Increased the execution timeout limit for conversation analysis tasks, fixing the issue where longer or more complex conversations failed during analysis due to timeouts. * **Fixed PDF translation failing to run in specific environments**: Fixed a compatibility issue between the PDF translation feature's internal processing component and the background task framework, ensuring PDF translation works normally in all deployment environments. * **Fixed incorrect display of IQ Service search results**: Corrected the knowledge title (now extracted from the document's actual title), directory name (now the knowledge base name), and numbering fields of IQ Service integrated search results, ensuring search results display correctly. * **Fixed server errors on the Line channel for malformed requests**: When the Line Webhook received malformed requests missing the signature header, it previously returned a 500 server error and triggered monitoring alerts. After the fix, it returns the correct 400 error code. ## v2026.04.20 *** ### ✨ New Features * **AI assistant browser tool**: AI assistants gain web browsing capability, querying web page content in real time to help answer questions. Ideal for use cases where the AI assistant needs to obtain the latest information from the web. * **Web Chat conversation event banners**: The Web Chat interface adds real-time conversation event notifications; users can see conversation state changes such as transfers, queuing, and assignments in the chat timeline, no longer having to guess what stage the conversation is currently in. * **Improved Web Chat voice call experience**: Voice calls add dial tone audio feedback and a quick entry point from the admin console. A standard dial tone plays while connecting, so users clearly know the call is being placed. The Web Chat public URL settings area in the admin console adds a "Voice Mode" button; with the voice assistant enabled, the voice conversation page can be opened with one click. ### 🚀 Core Performance Improvements * **Faster knowledge base image processing**: When uploading documents containing large images to a knowledge base, the system automatically resizes the images before building the index, reducing processing time and resource consumption. ### 😊 User Experience and Interface Improvements * **Web Chat output file panel moved to a side tab**: When the AI assistant generates files through the code interpreter, the output panel opens automatically on the side. The file count updates in real time on the tab label, and you can freely switch between the canvas and output tabs to manage all output files in one place. * **Comprehensive improvements to the knowledge base "AI Search" experience**: The knowledge base search page has been renamed "AI Search", with search results now shown as a collapsible list displaying an 80-character preview when collapsed, making it easy to quickly scan multiple results. Each result adds a "Preview" button that opens the original document preview directly. All citation source areas add a unified "View Source" entry that opens the document preview window with one click. * **Keyboard support for interactive Q\&A**: The interactive Q\&A panel adds full keyboard operation: arrow keys to switch options, Enter to confirm a selection, Space to toggle checkboxes in multi-select questions, and Esc to close the panel. Focused options automatically scroll into view, and mouse and keyboard operation can be mixed freely. ### 🛠️ Bug Fixes * **Fixed list pagination settings not being preserved**: After selecting the items-per-page count on list pages such as MaiAgent roles, skills, and contacts, navigating to a detail page and back no longer resets it to the default. The crawler result import dialog also adds support for custom items-per-page counts. * **Fixed abnormal code block display in MaiGPT chat**: In chat messages, array indices inside code blocks and inline code (such as `arr[0]`) were misidentified as citation links, corrupting the code display. After the fix, code segments are protected before citation conversion, and citation sources are checked for existence, ensuring code displays correctly. \[0]\`) is no longer mistakenly converted into a citation hyperlink. Also fixed the issue where conversion was triggered even when no citation sources existed. * **Fixed inconsistent button corner radii in the Admin interface**: Button groups in the admin console display correct corner radii again, and the quick-insert prompt area on the AI assistant settings page also fixed overflow and font size issues. * **Fixed buttons being pushed off screen when MaiGPT titles are too long**: Overly long conversation titles on the MaiGPT page are automatically truncated, with the full title shown on hover. The knowledge base selector, share, and collapse buttons on the right are no longer pushed off screen and remain visible and clickable at all times. * **Fixed abnormal tool calls in the deep research feature**: When using the deep research feature, request handling for tool calls works normally again. * **Fixed a compatibility issue between Gemini models and JSON format output**: For AI assistants using Gemini models with JSON format output configured, sending a message containing only an image (no text) no longer triggers an error. Behavior for other models (Claude, OpenAI) is unaffected. ## v2026.04.17 *** ### ✨ New Features * **Custom items-per-page for table pagination**: The paginator on all admin console tables now supports manually entering the number of items per page (up to 500), instead of only choosing from fixed options (10/50/80/100). Custom values are saved automatically and applied on the next visit, and saved custom options can be deleted from the dropdown at any time. ### 🔐 Security and Governance Enhancements * **Automatic authentication token revocation on logout**: On logout, the system automatically sends the authentication token back to the backend for revocation, ensuring the token becomes invalid immediately after logout. This prevents logged-out credentials from being reused, improving account security. ### 🛠️ Bug Fixes * **Fixed the MCP connector settings form being autofilled by browsers**: When adding a private MCP server, the "Server URL" field was misidentified by browsers as a login form and autofilled with a saved email address, causing URL format validation to fail. After the fix, form fields no longer trigger browser autofill. * **Fixed AI assistant tool search loading failures**: When AI assistants searched for and loaded external tools, loading failed in some situations. After the fix, tool search and loading work normally again. ## v2026.04.16 *** ### ✨ New Features * **Knowledge base search AI summary**: The knowledge base search test page adds an AI summary feature. After a search, a summary is automatically generated above the results, similar to Google AI Overview. Search results appear immediately while the AI summary streams in progressively, without affecting search speed. Supports responses in multiple languages including Traditional Chinese, Simplified Chinese, English, Japanese, and Korean. * **Web Chat FlexMessage message action support**: Web Chat now supports FlexMessage message action buttons. In LINE, FlexMessage buttons can send messages directly to the AI assistant, but in Web Chat they previously only showed a browser alert. After the update, clicking a FlexMessage message button sends the message directly to the AI assistant and gets a reply, consistent with the LINE experience. * **Conversation platform dispatch system**: Conversation platforms add a dispatch system that supports AI automatically assigning conversations to appropriate team members. When the AI assistant determines human intervention is needed, it can automatically transfer the conversation to the designated team member, improving customer service efficiency and response speed. ### 😊 User Experience and Interface Improvements * **Redesigned announcement dialog visuals**: The system announcement dialog has been visually redesigned. The new version displays different colors and icons based on severity, aligns the title and date to the left, and fixes plain-text line breaks not displaying correctly. * **Customizable number of Web Chat conversation starter questions**: Web Chat previously always displayed 2 random conversation starter question examples; administrators can now customize the display count, ranging from 1 to 10. The default remains 2, compatible with existing behavior. ### 🛠️ Bug Fixes * **Fixed processing failures for knowledge base documents containing special characters**: Some PDF and Word documents contained special control characters after parsing, causing processing to fail when uploading to a knowledge base. After the fix, the system automatically cleans invalid control characters from documents, ensuring all document types can be processed normally. * **Fixed Web Chat attachment sending failing due to missing filenames**: When sending attachment messages through Web Chat, a missing filename field in the attachment data caused sending to fail. After the fix, the system automatically extracts the filename from the file path, ensuring attachment messages are sent normally. * **Fixed Telegram message sending failing on transient errors**: When the Telegram channel encountered a transient network error while sending a message, the message failed outright. An automatic retry mechanism has been added that retries with increasing intervals on transient errors, improving message delivery rates. * **Fixed LINE messages failing to be delivered due to expired reply tokens**: When a LINE channel reply token expired, the message could not be delivered to the user. After the fix, the system automatically detects expired reply tokens and switches to push messaging to deliver the message, ensuring delivery. ## v2026.04.15 *** ### ✨ New Features * **Fullscreen voice call page**: Voice calls were previously presented only as a bottom popup, limiting the interaction experience. A new standalone fullscreen call page includes an immersive background, large avatar display, language switching, and call control buttons, providing a more intuitive voice interaction experience. Ideal for use cases requiring extended voice conversations. * **Conversation event timeline and tag system**: The conversation details page adds a complete event timeline that visualizes the history of transfers, assignments, status changes, and other events. A custom tag system also replaces the previous priority mechanism, letting each conversation platform independently configure its own tags to meet diverse classification needs. * **Knowledge base FAQ batch import and export**: When managing large numbers of FAQs, adding them one at a time is inefficient. New batch upload and export features let you process multiple FAQ entries at once via file, ideal for teams that need to manage Q\&A content at scale. * **Conversation Canvas map display**: AI assistant conversations add a Google Maps map type for Canvas, rendering interactive map markers, routes, and info popups. Ideal for travel planning, logistics tracking, store navigation, and other scenarios requiring geographic information display. * **Conversation analysis**: Conversation platforms add conversation analysis capability, automatically analyzing conversation content and generating insight reports to help administrators understand conversation quality and service trends. ### 🚀 Core Performance Improvements * **Faster Web Chat page rebuilds**: When the number of Web Chats exceeded tens of thousands, the page rebuild task processed them one by one and timed out. It now processes them in parallel batches (100 per batch), ensuring page updates for large numbers of Web Chats complete successfully without timing out. ### 😊 User Experience and Interface Improvements * **Automatic sync back to source files after editing knowledge base documents**: After editing knowledge base content from plain-text files (txt, md, csv, etc.), the system automatically merges the edits and syncs them back to the original file. Non-plain-text files (PDF, DOCX, etc.) are unaffected. If the sync fails, only a reminder is shown; the knowledge base content is still saved. * **AI assistant role instruction Token estimate display**: Below the "Role Instructions" input area on the AI assistant settings page, an estimated Token count is displayed in real time (in the format "about XX tokens"), helping users understand the Token consumption of their instructions. * **Random display hints for Web Chat greetings and conversation starter questions**: Hint text has been added next to the "Greeting" and "Conversation Starter Questions" headings on the Web Chat settings page, clearly explaining the random display rules: 1 greeting is randomly shown each time, and 2 conversation starter questions are randomly shown each time. * **Relocated the reset conversations button**: The "Reset All Conversations" button has moved from next to the action icons on the conversation platform list page into the "Enable Reset Command" section of each conversation platform's edit page. As a low-frequency, irreversible operation, adding a layer of friction reduces the risk of accidental clicks. ### 🛠️ Bug Fixes * **Fixed AI assistant citation numbers not matching their sources**: Fixed the issue where citation numbers in AI assistant replies (e.g., \[1], \[2]) opened the wrong document. Citation numbers now correctly correspond to the knowledge base documents the AI actually referenced, and the same document no longer appears multiple times in the citation list. \[1], \[2]) could point to the wrong citation source document. After the fix, citation numbers correctly correspond to their citation sources, and the same citation source no longer appears more than once. * **Fixed the knowledge base document viewer notice banner being hidden by scrolling**: In the knowledge base document viewer, the notice banner previously scrolled away with the page. After the fix, the banner is pinned to the top of the viewer, unaffected by page scrolling and zooming, and remains visible at all times. * **Fixed task retry failures when creating SQL knowledge base tables**: When uploading a SQL file to create tables, if the task needed to be retried (such as when re-dispatched), it could fail because the placeholder record could not be found. After the fix, if the table has already been created successfully, the result is returned directly, and inconsistent states now produce a clear error message. * **Fixed a single AI evaluation metric parsing failure causing the entire evaluation to fail**: When running AI evaluations, if the language model returned an unexpected format for one evaluation metric, the entire evaluation task was interrupted. After the fix, a parsing failure in a single metric no longer affects the evaluation of other metrics, making evaluation results more complete and reliable. ## v2026.04.13 *** ### ✨ New Features * **Contact tag management**: Previously, contacts could not be categorized, making it hard to quickly filter specific groups among large numbers of customers. A complete tag management feature has been added, supporting custom tag creation, batch adding or removing tags in the contact list, filtering contacts by tag, and viewing tags directly in the customer info panel on the conversation page. * **Voice assistant interruption control settings**: When using the voice assistant, it was previously not possible to configure whether users could interrupt the AI's voice response. New interruption control settings let you enable or disable "Allow Interruption" and configure the minimum speech duration and minimum word count thresholds required to trigger an interruption, ideal for teams that need fine-grained tuning of the voice interaction experience. * **Code Interpreter output file panel and environment upgrade**: When using the Code Interpreter feature, files generated by the AI previously could not be viewed or downloaded from the frontend. Both the admin console and Web Chat now add an output file panel for browsing and downloading AI-generated files directly. The execution environment also comes pre-installed with common data analysis and document processing packages plus Chinese fonts, and supports reading attachments uploaded in the conversation. * **Usage statistics Credits metric and billing categories**: Organizations on Credit-based billing plans previously could not see the detailed breakdown of credit consumption in usage statistics. A new Credits metric option provides 7 standard charts and three-level billing category drill-down (category pie chart → item ranking bar chart → member detail expansion), making it easy to understand the structure of credit consumption. * **AI assistant proactive questioning**: AI assistants could previously only answer questions passively and could not proactively ask when unsure of user intent. A new proactive questioning feature lets the AI assistant automatically generate structured questions with options during conversations, helping guide users to clarify their needs and improving conversation accuracy. * **Japanese and Korean language support**: The platform previously supported only Traditional Chinese, Simplified Chinese, and English. Complete Japanese (日本語) and Korean (한국어) language support has been added, covering all 5,000+ translation entries, with automatic browser language detection and switching, so Japanese and Korean users can operate the admin console in their native language. ### 😊 User Experience and Interface Improvements * **Unified Canvas compilation status indicators**: During streaming generation, non-React Canvas types (Mermaid, Markdown, HTML) previously showed raw source code without a consistent loading indicator. All Canvas types now uniformly display a spinner and "Compiling" text during compilation, automatically switching to the rendered result on completion. * **Improved API key management interface**: The API key list page removes redundant table gridlines for a cleaner layout, the revoke action uses a more intuitive delete icon, and revoked or expired keys now display a disabled status button instead of being hidden, making key status easier to identify. ### 🔐 Security and Governance Enhancements * **Sensitive information protection in system logs**: Strengthened the security of system logging by updating logging practices to avoid accidentally recording sensitive credential information, ensuring users' authentication data is properly protected in system logs. ### 🛠️ Bug Fixes * **Fixed screen flicker when sending the first message in Web Chat**: When a user sent their first message in Web Chat, the welcome message animation overlapped with the quick question disappearance animation, causing visible flickering and jumping. After the fix, the screen transitions smoothly when the conversation starts, with no visual interference. * **Fixed the meeting notes page not scrolling on small-screen devices**: When viewing the meeting notes detail page on small-screen devices such as a MacBook Air, the page content was cut off and could not be scrolled down. After the fix, the page scrolls normally, ensuring meeting notes can be read in full on all devices. * **Fixed error messages for SQL database connection tests**: When entering a malformed connection URL (such as one containing placeholder text) in SQL database connection settings, the system returned an unclear server error. After the fix, a clear validation error message is displayed, making it easy for users to correct their connection settings. * **Fixed occasional AI assistant scheduled task execution failures**: When using the AI assistant scheduling feature, tasks occasionally failed due to expired database connections. After the fix, the system ensures a valid database connection before each scheduled execution, making scheduled tasks run more reliably. ## v2026.04.10 *** ### ✨ New Features * **Conversation platform tag classification and event tracking system**: Conversation platforms previously offered only a four-level numeric priority (0-3), with no flexible classification. Each conversation platform can now define its own custom tags to freely label and filter conversations. Complete conversation event tracking has also been added, recording the history of each assignment, transfer, queuing, and human takeover, with event notifications pushed via real-time messaging. When the AI assistant transfers a conversation, it also automatically classifies it using the available tags. * **MCP connector internal network access settings**: Enterprise users with on-premise MCP services can configure allowed internal network ranges (CIDR), letting MCP connectors securely access internal services while maintaining security protections against external networks. ### 🚀 Core Performance Improvements * **Improved knowledge base search connection stability**: Improved the search service's connection management mechanism to automatically detect stale connections invalidated by thread switching and re-establish them immediately. Under high-load conditions, the stability of the knowledge base search service is significantly improved. ### 😊 User Experience and Interface Improvements * **Improved schedule execution record retention**: Schedule execution audit records now directly store each AI assistant reply. Even if conversation messages are processed by subsequent workflows, historical replies remain fully accessible from the schedule execution records, making it easy to trace the results of scheduled tasks. ### 🛠️ Bug Fixes * **Fixed AI assistant conversation failures caused by tool name conflicts**: When the code interpreter and custom tools (or MCP tools) were enabled at the same time, duplicate tool names (such as identically named command-line tools) caused the AI model to return an error and made conversation impossible. After the fix, the system automatically resolves tool name conflicts, ensuring conversations work normally. * **Fixed security validation failures when uploading knowledge base files**: When uploading files to a knowledge base, some invocation methods sent file data in special formats that the security validation service could not process correctly, returning a server error. After the fix, security validation correctly identifies and processes these files. * **Fixed schedule-dedicated conversations being deleted by mistake**: When the system automatically cleaned up empty conversations, it did not check for associations with schedule-dedicated conversations, potentially deleting conversations still in use by schedules. After the fix, schedule-dedicated conversations are no longer affected by the cleanup process. * **Fixed abnormal AI thinking mode settings**: The AI assistant's thinking mode settings could cause service issues in certain situations due to incorrect formatting. After the fix, format validation has been added at multiple service layers, ensuring thinking mode operates reliably. ## v2026.04.09 *** ### ✨ New Features * **Organization API key management**: Organization administrators can now centrally manage API keys from the settings page, including creating new keys (with member and expiration date assignment), viewing the status of all keys, and revoking keys that are no longer needed. The full key value is shown for copying at creation time; afterwards, only the prefix is shown for security. * **Credit usage statistics and analysis**: A complete Credit usage statistics dashboard has been added. Organization administrators can view Credit spending trends over time and analyze usage across multiple dimensions, including AI assistant, member, MaiAgent role, conversation platform, contact, and voice assistant. Billing categories support three-level drill-down, from overview to detail at a glance. * **Filter conversations by contact tags**: The conversation list adds filtering by contact tags, with support for selecting multiple tags at once. Customer service agents can quickly find conversations in specific categories (such as VIP or a specific product line), improving conversation triage and handling efficiency. ### 😊 User Experience and Interface Improvements * **Improved invitation management page tabs**: Invitation management on the profile page has been split from a single long table into three separate tabs: "Invitation Codes", "Invitation Links", and "Invitation Emails". Each tab has its own search field and filters; invitation links support one-click copying, and invitation emails support status filtering and resending. * **Increased knowledge base batch import limit**: The batch import limit for web crawl results has been raised from 50 to 200 per operation. Users who need to import large amounts of web data into knowledge bases can reduce the number of operations by about 75%. ### 🛠️ Bug Fixes * **Fixed some language models failing to respond**: Some language models that do not support streaming responses (such as DeepSeek R1 provided through AWS Bedrock) returned no response or errors during conversations. The system now automatically detects a model's streaming support and switches to full-response mode for models that do not support it, ensuring conversations proceed normally. * **Fixed indefinite waiting when AI conversation streams were interrupted**: When conversing with AWS Bedrock models, if the streaming response was interrupted midway, the system waited indefinitely for the next response chunk. After the fix, a 90-second timeout mechanism reports the error within a reasonable time when a stream is interrupted, preventing users from waiting endlessly with no response. Errors occurring mid-stream that were not correctly captured have also been fixed, ensuring errors are properly recorded and classified. ## v2026.04.08 *** ### ✨ New Features * **AI assistant voice conversation interruption control**: When using the voice assistant, you can now configure whether users are allowed to interrupt mid-reply, and adjust the minimum duration and minimum word count thresholds that trigger an interruption, making voice conversation interaction more natural and fluid. * **Reset all conversations for third-party conversation platforms**: Administrators managing third-party platform (LINE, Messenger, Instagram, Telegram, Teams, WhatsApp, Email, Slack) conversation channels can now reset all historical conversations in a channel with one click. A second confirmation is required before the operation, and the number of reset conversations is displayed afterward. * **MaiGPT language model selection memory**: Previously, after switching language models in MaiGPT, new conversations always reverted to the default model. The system now automatically remembers each user's preferred language model and applies the last selection when creating new conversations, with no manual switching each time. * **Organization activation records and event export**: The admin console adds an organization activation records page where administrators can view organization activation history and export operational event records, making it easy to track member activity and conduct audits. Supports Traditional Chinese, Simplified Chinese, and English interfaces. * **Code execution panel support in the embedded web version**: Users of the embedded web version (Web Chat) can now correctly see the output panel for code execution results. Previously, because the API did not provide the relevant configuration information, the frontend could not correctly determine whether to display this panel. ### 😊 User Experience and Interface Improvements * **Refined stop generation button visuals**: The stop generation button during conversations has changed from a white square on black to a white background with a dark gray outline, for a lighter, softer look. Corresponding color adjustments have also been made in dark mode, for a more consistent overall style. * **Improved scheduled task conversation platform selection and interval settings**: In scheduled task settings, "Inbox" has been renamed "Conversation Platform", with icon labels showing each platform type (such as LINE, Web, Admin). The schedule interval setting has changed from a single seconds input to four simultaneous fields for days, hours, minutes, and seconds, making it more intuitive to use. * **Paginated external database table descriptions**: The table description list on the external database settings page adds pagination. Previously, with more than 20 tables, the list could not display all the data; all table descriptions can now be browsed through pagination. * **Completed permission name translations in organization overview**: On the organization overview page, some permission names (meeting notes permission, meeting list permission, real-time transcription permission, AI assistant marketplace permission) previously displayed as English key values; Traditional Chinese, Simplified Chinese, and English translations have now been completed. * **Message sending disabled while attachments are uploading**: When uploading an attachment in a conversation, the send button is automatically disabled until the attachment finishes uploading, preventing users from sending incomplete messages before attachments are ready. ### 🔐 Security and Governance Enhancements * **Third-party package remote code execution vulnerability patched**: Patched a remote code execution (RCE) security vulnerability in the jsonpickle package by upgrading to the fixed version, ensuring the security of the platform's data processing. * **Refined security scanning process and scope**: Streamlined the security scan detection scope to ensure only code actually deployed in production is scanned. Removed container configuration files no longer in use (eliminating 3 high-risk scan findings) and corrected false-positive flags from static analysis tools, improving the accuracy and credibility of security scan results. ### 🛠️ Bug Fixes * **Fixed two issues in the scheduled task edit form**: When switching schedule types (such as from interval to one-time), previously configured delivery options are no longer cleared; the execution time format for one-time schedules has been corrected to a format the backend accepts. * **Fixed incorrect error messages after automatic retries were exhausted**: After the AI assistant's automatic retry mechanism exhausted all attempts, the returned error message was incorrect. After the fix, the original error cause is displayed correctly, helping users understand the problem. * **Fixed pagination errors when browsing message history**: When browsing message history, if the pagination cursor was corrupted or contained an invalid time format, the system no longer produces a 500 internal error and instead gracefully guides users to reload the conversation. * **Fixed boolean value formatting issues in contact custom attributes**: When using boolean values (such as "yes/no" flags) in contact custom attributes, inconsistent letter casing caused errors during data updates. The system now automatically normalizes boolean value formats, ensuring consistent formats for data storage and search. * **Fixed attachment file reading**: When the AI assistant processes user-uploaded attachments, the file reading logic has been corrected, ensuring attachment content is fully read and processed. * **Fixed handling issues with special characters in messages**: Special characters in conversation messages could cause message processing to fail. After the fix, the system uniformly sanitizes special characters, ensuring messages are sent and displayed normally. * **Fixed disabled AI models still appearing in the menu**: The Flash image model discontinued by Google has been removed from the selectable list, preventing users from selecting a model that no longer works. ## v2026.04.07 *** ### ✨ New Features * **Agent scheduling**: AI assistant tasks previously had to be triggered manually; you can now set up schedules for AI assistants to run automatically. Supports three schedule modes: timed execution (Cron), interval execution, and one-time schedules, with a choice of in-conversation or standalone execution, plus designated inboxes and message delivery channels (conversation or Webhook). * **Added GPT-5.4 mini and GPT-5.4 nano model support**: The platform adds the GPT-5.4 mini and GPT-5.4 nano models, which users can select directly. ### 🚀 Core Performance Improvements * **AI model compatibility upgrade**: Upgraded the AI framework packages to support the latest Bedrock models, ensuring the platform can use the latest AI model capabilities. * **AI model performance monitoring tools**: Added an AI model performance metrics analysis tool, supporting P95/P99 percentile statistics for time-to-first-token and throughput per second, making it easy to monitor model response quality. ### 😊 User Experience and Interface Improvements * **Completely redesigned tool and skill selectors**: Redesigned the AI assistant's tool and skill selection interfaces. They now use a search-first two-column grid layout, add selected-item tags for quick overview and removal, grouped collapsible display with selected counts, and full-description Popover previews. The skill selector adds search functionality (previously unavailable), greatly improving the efficiency of selecting many items. * **Display names for disabled knowledge base file parsers**: When the parser used by a knowledge base file has been disabled, the dropdown in the file list now displays the parser name marked as disabled, replacing the previously hard-to-recognize UUID code, so users can understand the file's parsing status. ### 🔐 Security and Governance Enhancements * **Platform security package updates**: Upgraded multiple Python and system packages to patch known security vulnerabilities, including high-risk issues such as path traversal, buffer overflow, and denial of service, strengthening the platform's overall security posture. ### 🛠️ Bug Fixes * **Fixed transcript speaker label save errors**: After modifying multiple speaker labels in the transcript viewer, saving occasionally failed with an error. All label changes are now submitted in a single batch, eliminating problems caused by many concurrent requests, and the save function is stable again. * **Fixed the code interpreter output panel not displaying**: When using the code interpreter, the output panel did not display if no knowledge base was configured. After the fix, the code interpreter's output panel displays normally whether or not a knowledge base is configured. ## v2026.04.02 *** ### ✨ New Features * **Interrupt AI streaming replies**: While conversing with an AI assistant, you can click the stop button at any time to interrupt the reply being generated. If interrupted during the thinking phase, the message bubble is automatically removed; if interrupted during the reply phase, the partially generated content is kept and labeled "Reply interrupted", making it easy to ask again. ### 🚀 Core Performance Improvements * **Improved background task reliability and connection stability**: Improved the background task processing mechanism by enabling delayed acknowledgment, ensuring tasks are marked as processed only after completion, so tasks are automatically re-dispatched rather than lost when a worker node terminates abnormally. Expired database connections are also proactively cleaned up before message processing, avoiding processing failures caused by invalid connections. ### 😊 User Experience and Interface Improvements * **Improved duplicate tool name notices for AI assistants**: When configuring an AI assistant, if tools with identical names are accidentally added, the error message now clearly lists the duplicated tool names, making it easy to find and fix the problem without checking each one individually. ### 🛠️ Bug Fixes * **Fixed display issues with the embedded Web Chat conversation button**: When Web Chat is embedded in a website, the conversation button no longer flashes the default icon during loading. If a custom button icon is configured, it loads correctly from the API and displays. Before the fix, the button briefly showed the default appearance before switching to the custom style. * **Fixed knowledge base document speaker labels being uneditable**: For speech-to-text documents uploaded through a knowledge base, editing speaker labels no longer produces a permission error. Before the fix, only speaker labels for documents attached directly to an AI assistant were editable; documents under a knowledge base path returned a 403 error. * **Fixed image and sticker attachments in conversations being undeletable**: Deleting image or sticker attachments on conversation platforms no longer produces a "resource not found" error. Before the fix, the deletion returned a 404 error because attachments were associated with conversations incorrectly. * **Fixed batch-imported members not automatically getting API keys**: Members added through batch import now automatically get API keys created. Before the fix, batch-imported members lacked API keys and could not use features requiring key authentication. * **Fixed tag creation failures when uploading multiple knowledge base files simultaneously**: When uploading multiple knowledge base files at the same time, tag creation no longer fails due to key conflicts from parallel processing. Before the fix, processing multiple files concurrently could cause tag creation errors. ## v2026.04.01 *** ### ✨ New Features * **Organization API key management**: Organization administrators can now centrally manage API keys from the settings page, including creating new keys (with member and expiration date assignment), viewing the status of all keys, and revoking keys that are no longer needed. The full key value is shown for copying at creation time; afterwards, only the prefix is shown for security. * **Batch speech-to-text processing**: When uploading long audio files to a knowledge base, the system now supports batch speech-to-text processing, overcoming the previous 60-second single-pass limit. Ideal for teams that need to transcribe long audio such as meeting recordings and customer service calls. * **Custom system email sender account**: Administrators can configure a custom SMTP email account in the admin console (with password encryption), so system notification emails are sent from the enterprise's own mailbox instead of the default shared account. * **Login record tracking and activity log export**: The system adds member login record tracking; administrators can view members' login history in the activity log, with support for asynchronously exporting audit records, ideal for enterprises requiring security compliance tracking. ### 🚀 Core Performance Improvements * **PDF translation 10x faster**: The PDF translation service's throughput has increased from 1 to 10 per second, dramatically reducing wait times when translating large documents. ### 😊 User Experience and Interface Improvements * **Interrupt AI responses in real time**: During a conversation, users can now stop the AI's in-progress reasoning and response generation at any time, without waiting for the full reply to finish, improving real-time control over the conversation. * **Context variables in tool Headers**: When configuring an API tool's Headers, you can now insert conversation context variables directly (such as user ID and session ID), making tool integrations more flexible. * **Custom icons for the embedded chat button**: The chat button embedded in websites now prioritizes the custom icon configured in the admin console, keeping the embedded chat widget's appearance consistent with your brand identity. ### 🛠️ Bug Fixes * **Fixed duplicate Flex Message card display**: When replying with Flex Message cards, the same card was occasionally rendered twice during streaming. After the fix, cards are correctly deduplicated and moved outside the message bubble to display independently, for a cleaner layout. * **Fixed conversation titles wrapping in the sidebar**: In the conversation list sidebar, longer conversation titles failed to truncate correctly and wrapped, breaking the layout. After the fix, titles display on a single line and are automatically truncated with an ellipsis when too long. * **Fixed the broken live preview on the Web Chat settings page**: The live preview area on the Web Chat settings page failed to display, and the action column on conversation platform settings pages was not correctly pinned to the right. After the fix, the preview works normally and the action column is pinned in the correct position. * **Fixed abnormal keyword highlighting in knowledge base search tests**: On the knowledge base search test page, switching between "Raw Text/Highlighted Text" caused conflicts between Markdown rendering and keyword highlighting, corrupting the format. After the fix, keyword highlighting displays correctly in all modes. --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter: ``` GET https://docs.maiagent.ai/release-note/untitled/archives/04.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.