> 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/05.md).
# May 2026
## v2026.05.29
***
### π Security and Governance Enhancements
* **Strengthened MaiAgent Role Tool and Skill Permissions**: Strengthened the role-based tool and skill permission controls to ensure members can only use tools and skills authorized for their roles during conversations. Organization administrators can manage member access with greater confidence through role settings.
### π οΈ Bug Fixes
* **Fixed an issue with the WhatsApp phone number registration flow**: Registering a phone number through the WhatsApp channel failed with an error, preventing registration from completing. After the fix, the WhatsApp phone number registration flow works normally again.
* **Fixed a data-passing issue in Canvas workflow transfer nodes**: When using the transfer feature in Canvas workflow design, transfer nodes did not pass data correctly, preventing the workflow from running properly. After the fix, transfer nodes correctly pass all required data and the workflow transfer feature works normally again.
## v2026.05.28
***
### β¨ New Features
* **Balance alerts and grant information on the Credits overview page**: The Credits overview page now integrates a new information card that shows the account balance, balance alert configuration status, and the time and amount of the last and next Credits grants at a glance. You can track the full Credits status without visiting other pages, with color-coded warnings when the balance is low.
* **Automatic retry for interrupted AI assistant streaming**: Added an AI assistant-level setting for automatically retrying interrupted streams. When an AI response is interrupted mid-stream, the system automatically re-fetches the response; on the user's side, already-produced tool call results are kept and the text portion is automatically replaced with the retried result, with no need to resend the message manually. Enable or disable this in the AI assistant's basic settings.
### π Performance
* **Automatic recovery for interrupted test dataset evaluations**: When a system update is deployed, in-progress test dataset evaluations no longer get permanently stuck in the "Running" state. The system automatically finalizes completed test case results and marks the evaluation as "Partially completed," so users can view the finished results. An hourly cleanup mechanism was also added to prevent any lingering stuck evaluations.
### π User Experience
* **Automatic adaptation of the AI assistant's custom response length limit**: When configuring an AI assistant's custom response length limit, switching to custom mode now automatically initializes the value to the selected model's maximum instead of showing a blank field. When switching models, the value automatically adjusts to the new model's valid range, preventing errors caused by settings that exceed model limits.
### π οΈ Bug Fixes
* **Fixed an issue where Flex message cards did not display in Web Chat conversations**: When using Web Chat, Flex message cards in AI assistant replies (such as button menus and image-text combinations) were discarded during streaming, so the message bubble could not render the Flex content. After the fix, Flex messages are correctly preserved during streaming and displayed in full.
* **Fixed an issue where the conversation area showed abnormal patterns in Web Chat dark mode**: With dark mode enabled in Web Chat, the conversation area background displayed a repeating pattern of downward arrows. After the fix, the conversation area displays normally in dark mode.
* **Fixed an issue where the message input box background disappeared on the conversation page**: Due to a styling issue, the message input box on the conversation page lost its background color and border and became transparent. After the fix, the input box's normal white background and gray border are restored.
* **Fixed abnormal button spacing in Agent UI Flex message previews**: In Agent UI Flex message previews, the first button in the button area sat flush against the image or text above it with no spacing. After the fix, buttons maintain the standard 12px spacing from the content above.
* **Fixed a system error when entering an unsupported connection type for MCP tools**: When adding an MCP tool, entering an unsupported connection type (such as a local command) previously caused an internal system error. After the fix, a clear error message appears on save, explaining that only HTTP/HTTPS connections are supported.
* **Fixed tool call record detection failures when AI assistants use Anthropic models**: When an AI assistant made tool calls using Anthropic models, some tool call records could not be detected correctly, which could cause tool call results to go missing from the conversation history. After upgrading the SDK and correcting the detection logic, tool call records are fully preserved.
* **Fixed request errors in structured output containing nullable object fields**: When an AI assistant used the structured output feature, a custom JSON Schema containing nullable object fields produced an incompatible format that caused the OpenAI API to reject the request (400 error). After the fix, the format is automatically corrected to ensure all field definitions meet API requirements.
## v2026.05.27
***
### β¨ New Features
* **Agent UI officially released**: Agent UI has moved from beta to an official feature. All users with access can directly use the Agent UI menu, the FlexMessage editor, and the Agent UI tab in AI assistant settings without additional activation.
* **Bedrock streaming automatic retry setting**: When using AWS Bedrock models, if the stream is interrupted during transmission, the system can automatically retry to ensure a complete response. Administrators can enable this feature in each chat platform's settings; the chat interface labels retried messages, and the credit details also flag additional charges incurred by retries.
* **Multilingual link fields for the announcement system**: System announcements now support multilingual link fields, letting you set a corresponding link for each language version of an announcement.
### π Performance
* **Knowledge base web crawler upgraded to a new engine**: The underlying web crawler for knowledge bases has been upgraded, improving the accuracy and completeness of web content extraction.
### π User Experience
* **Clear message when conversation permission is missing**: When a user without conversation permission (for example, Agent Marketplace not enabled) tries to use Web Chat, the system no longer shows the generic "Failed to create conversation" message. It now clearly states "No conversation permission. Please contact an administrator to enable it," helping users understand the problem and seek assistance.
* **Improved AI tool prompts when credits are insufficient**: Unified the response behavior of all AI tools (including SQL query, code execution, MCP connectors, API tools, image generation, video generation, and web search) when credits are insufficient. Previously, some tools would abruptly interrupt the AI assistant's response flow; now the AI assistant proactively informs the user that credits are insufficient and suggests topping up, so the conversation is not interrupted unexpectedly.
* **Improved video generation status tracking**: The video generation feature adds structured status tracking, giving users a clearer view of video generation progress.
### π οΈ Bug Fixes
* **Fixed batch approval failures in announcement management**: In the admin console, batch-approving auto-translated announcements no longer fails due to database query conflicts. This issue caused every batch approval attempt by administrators to error out.
* **Fixed the Web Chat location-aware response feature not working correctly**: Due to an incorrect configuration read path, the location-aware response feature in Web Chat did not work even when enabled in the chat platform settings. After the fix, location-aware responses correctly provide localized replies based on the user's location.
## v2026.05.26
***
### β¨ New Features
* **WhatsApp Business coexistence mode**: WhatsApp Business Account (WABA) settings add coexistence mode support, allowing multiple WABA accounts to be connected simultaneously with no need to repeat setup for each account. Ideal for businesses serving different customer segments through multiple WhatsApp numbers.
### π Performance
* **Improved AI model streaming response stability**: When using an AI assistant, if the model's streaming response has not started sending any content within 45 seconds, the system automatically detects this and resends the request, preventing users from waiting a long time without receiving a reply. Administrators can enable this feature gradually for individual chat platforms.
### π User Experience
* **Code Interpreter tool now shown in credit details**: In credit transaction records and the pricing page, the Code Interpreter tool now displays the correct Chinese name and billing unit instead of the internal system code. Supports five languages: Traditional Chinese, Simplified Chinese, English, Japanese, and Korean.
* **Agent Marketplace permission description added to role settings**: On the role settings page, explanatory text has been added under the "Agent Marketplace" permission, clearly stating that this permission controls whether members can start conversations with AI assistants in the Agent Marketplace, helping administrators configure role permissions more accurately.
* **Credits overview page adjustment**: Removed the "Purchase Credits" button from the balance card on the Credits overview, so users are no longer directed to the currently hidden billing page.
### π οΈ Bug Fixes
* **Fixed batch upload failures for knowledge base files**: When different files in a knowledge base used different data formats for custom attributes (for example, writing a number to a field first and text later), the entire batch of files would fail to index. After the fix, the system automatically normalizes all custom attributes to text format, preventing upload failures caused by format conflicts.
* **Fixed some MCP tools being unusable**: When connecting certain third-party MCP tools, tool definition formats that did not fully conform to the standard specification prevented the system from correctly parsing the parameter structure, making the tools uncallable. After the fix, the system automatically handles common non-standard formats, improving compatibility with a wide range of MCP tools.
* **Fixed a background cleanup error in the browser tool**: After an AI assistant used the browser tool, the background cleanup process occasionally hit execution errors that could leave browser resources unreleased. After the fix, the cleanup process runs stably and browser resources are reclaimed normally.
## v2026.05.25
***
### β¨ New Features
* **MCP connector OAuth auto-discovery**: When connecting to a third-party MCP service, the system automatically discovers the server's OAuth authentication information (following the RFC 9728 standard). Users only need to provide the MCP server URL to complete the OAuth authentication setup automatically, with no need for administrators to fill in technical parameters manually.
* **Tag management and AI analysis added to the Copilot sidebar**: The sidebar of the embedded Copilot page adds conversation tag and contact tag management plus an AI analysis button. Support agents can manage tags and trigger AI analysis directly in Copilot without switching to the main chat platform page. Also fixed an issue where data did not update in real time when entering directly via a link.
* **WhatsApp Business account coexistence settings**: The WhatsApp Business Account (WABA) settings page adds coexistence settings, making it easier for users to manage the coexistence of multiple WhatsApp accounts.
### π User Experience
* **Clickable tags on list pages**: Tags on list pages (AI assistants, knowledge bases, databases, tools, chat platforms, and more) can now be clicked to jump directly to the corresponding detail page, with no need to manually copy the name and search. Tags that overflow the display area appear in a popup titled "Associated AI Assistants" for easy identification, and also support click-through navigation. Also fixed an issue where the list did not reload after pressing the clear button in the search box on the knowledge base, test dataset, and automated testing pages.
### π οΈ Bug Fixes
* **Fixed Web Chat welcome messages not rendering Markdown formatting**: Web Chat welcome messages configured with Markdown syntax such as headings, bold text, and lists now render formatted content correctly, consistent with how other messages are displayed.
* **Fixed MCP tools failing to execute correctly with non-text parameters**: When using MCP tools (such as the system operations assistant) to modify settings containing boolean or numeric values, parameters are no longer incorrectly converted to text, and the operation completes normally. This issue previously caused tools to retry repeatedly until hitting the limit and still fail.
* **Fixed AI assistant interactive replies being misjudged as errors**: When an AI assistant replies through interactive UI elements (such as option buttons or images), the system no longer misjudges the reply as failed. Interactive replies now display correctly, and message content is no longer replaced by an error message.
## v2026.05.22
***
### π οΈ Bug Fixes
* **Fixed AI assistants going off topic in long conversations**: During long conversations with an AI assistant that has memory enabled, when early conversation records were archived by the system, the AI assistant could misinterpret the current user input, producing responses unrelated to the user's actual question. After the fix, the AI assistant correctly recognizes the conversation history structure, accurately understands the current user input, and gives relevant responses.
* **Fixed knowledge base attachment insertion failures**: An error occurred when adding content to a knowledge base via attachments, preventing data from being imported. After the fix, users can import content into knowledge bases via attachments normally.
## v2026.05.21
***
### π οΈ Bug Fixes
* **Fixed real-time message send failures after the admin console page sat idle**: When using the admin console Web Chat, sending a message after the page had been idle for a while failed because the real-time messaging connection's authentication had expired. After the fix, reconnection automatically uses the latest authentication information, so resuming after idle no longer interrupts operations.
* **Fixed Excel/ODS files with minimal data being rejected on knowledge base upload**: When uploading Excel, XLS, XLSX, or ODS files with only 1-2 cells of data to a knowledge base, the system mistook the only data for column headers and rejected the file as empty. After the fix, the system correctly recognizes and processes spreadsheet files with minimal data.
* **Fixed AI conversations failing after uploading images on iOS phones**: When taking a photo or pasting an image via iOS Safari for an AI conversation, the image format label reported by the browser could mismatch the actual format (for example, a JPEG image labeled as PNG), causing the AI model to refuse to process it. After the fix, the system detects the actual format directly from the image file contents instead of relying on the format information provided by the browser.
* **Fixed occasional interruptions in Gemini model streaming responses**: AI assistants using Gemini models occasionally encountered error interruptions during streaming responses when hitting safety filtering or length limits. After upgrading the integration component, streaming handling is more robust and the related interruptions are resolved.
## v2026.05.20
***
### β¨ New Features
* **API request tracing for conversation messages in the admin console**: In the admin console's conversation records, you can now directly view all downstream API requests triggered by each message (language model calls, vector search, reranking, and more), with color-coded labels distinguishing request types (blue for language models, green for vector search, orange for reranking), making it easy for administrators to quickly pinpoint issues.
### π Performance
* **Faster startup when the database query feature connects to large external databases**: When using the database query feature to connect to external databases containing thousands of tables, the system no longer needs to pre-scan all table structures. Startup time drops from growing with the number of tables (thousands of tables could take tens of seconds) to completing instantly.
### π οΈ Bug Fixes
* **Fixed the AI assistant file query tool showing duplicate results**: When using the AI assistant's file query tool to search spreadsheet files (xlsx, csv, and more), the same file no longer appears in multiple duplicate results. Previously the system automatically converted uploaded spreadsheets into internal processing formats, causing the same file to appear up to three times in search results. After the fix, only the original file actually uploaded by the user is shown.
## v2026.05.19
***
### β¨ New Features
* **Forced triggering of user-selected skills and tools**: Skills or tools selected by users in Web Chat conversations are now reliably triggered. Previously, the AI might not use the specified tool after selection; now the system ensures the AI prioritizes the user-specified skill or tool in the first round of its response.
* **New real-time voice conversation models**: The voice assistant adds two real-time voice model options: OpenAI gpt-realtime-1.5 and Google Gemini Live 2.5 Flash native audio. Also updated the OpenAI voice option list, removing unsupported voices and adding two new voices, verse and cedar.
### π User Experience
* **Improved plan information tab in the admin console**: The plan information tab on the organization management page adds "Add Plan" and "Add Top-up" quick action buttons, so operations staff no longer need to work across multiple pages. The plan history table adds a billing mode column, the table layout is clearer, and translation support for all five languages has been completed.
* **Improved MaiAgent role database permission settings**: The database permission selector in role settings has been improved. Selected items are now correctly preserved when switching pages, no longer losing selection state due to page changes.
### π οΈ Bug Fixes
* **Fixed long text messages overflowing the layout in Web Chat**: When using Web Chat on mobile devices, sending very long text with no natural break points (for example, a continuous string without separators) caused the message bubble to exceed the screen width. After the fix, messages wrap correctly and no longer overflow the layout.
* **Fixed errors when uploading the same attachment across conversations**: Uploading the same file in different conversations triggered a "This file has already been uploaded" error. After the fix, the system correctly handles identical attachments across conversations without errors.
## v2026.05.18
***
### β¨ New Features
* **Text reply toggle for voice AI assistants**: Voice AI assistants add a "Voice Reply" toggle. With voice synthesis turned off, the AI assistant's replies appear as text messages in the conversation interface, suitable for scenarios that do not require voice output.
* **Credit consumption records link to AgentOps details**: In the organization settings' credit consumption records, the ID of conversation-type consumption records is now a clickable link that opens the AgentOps conversation detail window directly, making it easy to trace the credit usage details of individual conversations. Also removed the confusing "Description" column.
### π Performance
* **Faster chat platform tab count loading**: Loading of chat platform tab counts (Mine / Unassigned / All) is dramatically faster. Medium-sized organizations (about 9,000 conversations) dropped from 304 ms to 8.7 ms (a 35x improvement), and large organizations (about 198,000 conversations) dropped from over 600 ms to 153 ms (about a 4x improvement).
### π User Experience
* **Q\&A messages presented as standalone cards**: Q\&A content submitted by users via forms is now presented as standalone cards in the conversation interface. Questions appear in bold with answers arranged as subtitles, visually distinct from regular conversation messages, making form responses easy to identify at a glance.
* **Pre-Chat form custom titles and multilingual switching**: The Web Chat Pre-Chat form supports administrator-customized titles and subtitles (including multiple languages), and adds a language switcher so visitors can change the interface language before filling it out. The form card visual style has also been simplified, and dropdown menus now appear gray when nothing is selected, preventing visitors from mistakenly thinking a selection has been made.
* **Improved Code Interpreter output panel reliability**: The Code Interpreter output panel is now driven by structured signals. Whether the AI assistant writes files with absolute or relative paths, the output panel automatically detects and displays the produced files. Previously produced files also display correctly when historical conversations are reloaded.
### π οΈ Bug Fixes
* **Fixed the Webhook URL copy button layout issue on channel settings pages**: Fixed a layout issue on the LINE, Telegram, Slack, and Teams channel settings pages where the webhook URL copy button was pushed below the input box. The copy button is back on the same line as the input box.
* **Fixed feature menu hover triggering large numbers of duplicate requests**: Quickly sweeping the mouse across the feature menu no longer triggers large numbers of duplicate API requests. With debouncing added, operations are smoother and unnecessary server load is reduced.
* **Fixed usage statistics displaying too low for API-accessed AI assistants**: For AI assistants accessed via the Completions API, input token data in the usage statistics panel was not recorded correctly; actual usage was about 11 times the displayed value. After the fix, statistics correctly reflect actual usage. This issue did not affect billing amounts, only the display in the usage statistics panel.
* **Fixed slow updates to the expired member credit quota list**: After a member's credit quota expired, the "Expired" list on the quota allocation page could take up to 1 hour to update. After the fix, the update frequency is increased to a check every minute, and expired members move to the "Expired" list correctly within 1 minute.
* **Fixed a system error when voice calls ended**: Fixed an issue where residual audio data still attempted to transmit after the connection was closed when a voice call ended, producing a system error. This fix eliminates about 50 related error records per week and improves voice call stability.
* **Fixed a system error after meeting recordings were deleted**: When a user deletes a recording record before the meeting recording completes, the system no longer errors out because the record cannot be found. Deleted or nonexistent recording records are handled correctly without affecting other features.
* **Fixed knowledge base data cleanup failures when deactivating an organization**: Fixed an issue where deactivating an organization caused the cleanup job to fail midway because knowledge base-related data such as citation sources and tags were cleaned up in the wrong order. All related data is now fully removed in the correct order.
## v2026.05.15
***
### π Performance
* **Improved service connection stability**: Corrected inconsistent connection timeout settings between servers and the load balancer, eliminating the occasional 502 connection errors users encountered and improving overall service availability.
### π User Experience
* **Fixed selected count display in the member assignment interface**: In the MaiAgent role "Assign Members" interface, the "N selected" counters on both sides now work correctly again instead of always showing 0.
*
* **Automatic refresh of the code interpreter output panel**: After an AI assistant generates files with the code interpreter, the output panel on the right automatically displays the newly generated files, with no need to click refresh manually.
### π οΈ Bug Fixes
* **Fixed coordinate parsing errors in the AI assistant browser tool**: When an AI assistant used the browser tool to click page elements, operations occasionally broke off due to incorrectly formatted coordinates. After the fix, the system automatically corrects the format issue so the AI assistant can continue and complete the operation.
*
* **Fixed SSL connection compatibility issues in AI assistant tools**: For third-party services using legacy SSL certificates (such as TWCA Root 1998), administrators can enable a compatibility mode in the tool settings so AI assistants can connect to these endpoints successfully. Other security checks (certificate chain, hostname, validity period) remain enabled.
*
* **Fixed a data storage issue in conversation history records**: After a database driver upgrade, conversation history record data was double-encoded, resulting in an incorrect storage format in the database. After the fix, new conversation records are stored in the correct format, and existing affected data is restored via a repair program.
## v2026.05.14
***
### β¨ New Features
* **SCIM identity sync management interface**: Enterprise administrators can now choose SCIM as the identity sync method in organization settings (in addition to the existing LDAP). Supports SCIM endpoint connection settings, token management (including 180-day expiration warnings and one-time display), group mapping (with CSV import/export and automatic role creation), and sync log review.
* **Web Chat pre-conversation data collection form**: Administrators can configure a pre-conversation form in the Web Chat platform settings, supporting 4 field types (text, email, phone, dropdown) with up to 10 fields. The form supports multilingual labels and hint text in 13 languages, drag-and-drop reordering, and automatic mapping of visitor-submitted data to contact fields.
* **Genesys Cloud Open Messaging integration**: Administrators can configure the Genesys Cloud connection in the chat platform Webhook settings, with the system automatically managing OAuth token renewal. Supports "Test Connection" credential validation, a banner alert when credentials expire, and credential sharing across multiple Webhooks on the same chat platform.
* **"AI processing" indicator messages for Slack and Telegram**: Extended the "AI processing" indicator message feature from Teams to the Slack and Telegram platforms. Administrators can enable this feature in the chat platform settings, customize the indicator text, and set the initial delay and repeat interval; the placeholder message clears automatically after the AI replies.
* **Agent smart attachment interpretation toggle**: Administrators can enable "Agent smart attachment interpretation" in the Web Chat platform's AI awareness settings. When enabled, the AI assistant decides on its own whether to read attachment contents based on the conversation context, instead of always processing them into the knowledge base. Disabled by default; must be enabled manually.
* **Copilot AI suggestion version switching**: After support agents regenerate an AI suggestion, they can browse all historical versions using left/right toggle buttons and choose the most suitable version to send or edit. While the AI is generating a new version, the switcher automatically locks to the latest version. Also fixed an issue where the regenerate button could be clicked repeatedly at certain moments.
### π Performance
* **Deduplicated knowledge base retrieval content**: Optimized the knowledge base multi-path retrieval merging mechanism to automatically remove duplicate document fragments, keeping only the most relevant results. This effectively reduces duplicate content processed by the AI, lowers token consumption, and improves answer quality.
### π User Experience
* **Improved batch credit quota distribution flow**: When batch-distributing credit quotas to roles, added role search and infinite scroll loading, switched to a single batch-distribution submission (instead of one request at a time), and now display distribution result statistics in real time (successful / skipped / failed counts).
* **Improved chat platform field descriptions in the API documentation**: When integrating via the API or MCP, developers can now see complete descriptions of each channel's configuration fields, such as the Web Chat welcome message and greeting, directly in the API documentation, with no need to consult the source code.
### π οΈ Bug Fixes
* **Fixed MaiGPT skill selection being cleared after sending a message**: When using MaiGPT, selecting a skill or tool and then sending a message no longer clears the selection state. Skill and tool selections persist within the same conversation and are cleared automatically only when switching to a different conversation.
* **Fixed misaligned columns in the external database list**: The name column of the external database tab list now correctly displays the database name and type label, with no more layout misalignment.
* **Fixed the Web Chat ID copy button wrapping on the Webhook settings page**: In the chat platform Webhook settings, the Web Chat ID copy button now displays correctly side by side with the input box instead of wrapping to the line below.
## v2026.05.13
***
### β¨ New Features
* **Web Chat pre-conversation form**: When enabled, website visitors must fill in administrator-configured fields (name, email, phone, or custom dropdowns) before starting a conversation. The system automatically creates or matches contact records, improving visitor management efficiency. Supports interfaces in 13 languages, and visitors who have already filled in the form skip it automatically on their next visit.
* **MaiGPT skill menu and quick actions**: The MaiGPT conversation page adds a unified "+" menu that integrates five features: file upload, skills, tools, connectors, and Deep Research. Users can select skills from the menu (divided into system built-in and organization custom); selected skills appear as blue labels next to the input box and clear automatically after the message is sent (single-use effect).
* **AI assistant voice reply toggle**: Administrators can control in AI assistant settings whether the voice assistant replies by voice. When turned off, the voice assistant can still receive voice input but responds in text, suitable for scenarios that need voice input but prefer text replies.
* **Smart attachment handling decisions**: Chat platforms add a smart attachment handling mechanism where the system automatically decides based on attachment type and size whether indexing into the knowledge base is needed. Images, videos, and stickers are automatically skipped; text files are judged by size. The AI assistant can proactively read or index skipped attachments during the conversation. Enabled by default; can be turned off in the chat platform settings.
### π User Experience
* **One-click actions for AI suggestions**: Copilot's AI-suggested replies add three quick action buttons: send directly, edit then send, and regenerate suggestion. Support agents no longer need to manually copy and paste suggested content β they can adopt it with one click or edit before sending, and regenerate with one click if suggestion generation fails.
### π οΈ Bug Fixes
* **Fixed cross-page selection issues in group permission assignment**: When assigning resources such as databases, chat platforms, and knowledge bases to MaiAgent roles, accumulating selections across multiple pages and then pressing the move button previously preserved only the current page's selections. After the fix, selections from all pages are moved correctly.
## v2026.05.12
***
### β¨ New Features
* **Custom Agent workflow timeout message**: For AI assistants using Agent workflows, administrators can now customize the timeout message shown to users (up to 500 characters) when workflow execution exceeds the configured time, replacing the fixed system default text. The timeout minimum has also been lowered from 60 seconds to 30 seconds, letting AI assistants with short workflows report timeout status faster.
* **Dynamic CORS allowlist management**: Previously, adding a customer's Web Chat embed domain required code changes and deployment (taking 1-2 business days). Now administrators can add and manage allowed origin domains directly in the admin console, with changes taking effect immediately. Supports full URLs or regular expressions, dramatically shortening customer Web Chat go-live time.
### π οΈ Bug Fixes
* **Fixed knowledge base AI search summaries not including Metadata**: When using the knowledge base AI search summary feature, AI summaries did not reference annotated data even when Metadata was set to "Enable search." After the fix, AI summaries correctly read all search-enabled Metadata, making search results more complete and precise.
* **Fixed the chat platform tool "Compact mode" display having no effect**: When the tool call display mode was set to "Compact" in the chat platform settings, tool call results still displayed in the full panel, indistinguishable from "Full mode." After the fix, compact mode correctly renders in a single-line format (tool icon + name + status), consistent with Web Chat behavior.
* **Fixed misaligned columns in the external database list**: The "Database name" column in the external database list failed to render, causing all subsequent columns in that row (selected tables, associated AI assistants, creator, and more) to shift one cell to the left. After the fix, all columns align correctly and the database name column fully displays the icon, name, and type.
* **Fixed the output panel failing to refresh**: The output panel's refresh function did not work and data would not update. After the fix, the panel data refreshes normally.
## v2026.05.11
***
### β¨ New Features
* **Tool call API request detail view**: The tool call record detail page adds a "Request Information" section showing the actual HTTP request URL, method, headers, and body sent by the AI assistant. Administrators can directly confirm whether template variables (such as contact IDs) were replaced correctly, without asking the backend team to check logs. Sensitive header values (such as API keys and authorization information) are automatically masked.
* **Metadata displayed in knowledge base search test results**: Search results on the knowledge base search test page add metadata label display. Each search result shows the knowledge chunk's custom metadata in key:value format; when there are more than 6 labels, the full list can be expanded, making it easy for administrators to quickly identify the source and attributes of search results.
* **Mistral 3.5 reasoning mode support**: AI assistants add support for Mistral 3.5 reasoning (thinking) mode. When users select a Mistral 3.5 model, the AI assistant can use its reasoning ability to think step by step before answering, improving answer quality on complex questions.
### π οΈ Bug Fixes
* **Fixed Bedrock models forgetting context in multi-turn conversations**: AI assistants using AWS Bedrock models would forget their previous replies in multi-turn conversations, leading to repeated answers or an inability to continue the conversation. After the fix, the AI assistant correctly remembers the full conversation history and the multi-turn conversation experience is back to normal.
* **Fixed incorrect reasoning mode with AWS Bedrock Claude Opus 4.7**: When using the AWS Bedrock Claude Opus 4.7 model, the reasoning (thinking) feature did not work due to an incompatible API format. After the fix, the platform automatically detects the model version and uses the corresponding reasoning API format, ensuring reasoning summaries are generated correctly.
* **Fixed the screen stalling after deleting the current conversation**: After deleting an in-progress conversation in the Agent Marketplace and MaiGPT, the screen still showed the deleted conversation's old messages and required a manual click to create a new conversation. After the fix, deleting the current conversation automatically creates and switches to a new conversation.
* **Fixed AI assistant skill selections disappearing after cross-page selection**: When configuring AI assistant skills, browsing across pages or using search in the skill selector caused previously checked skills to be unchecked. After the fix, selected skills persist across pagination and search operations.
* **Fixed knowledge base metadata key validation issues**: Creating a metadata key with a duplicate name in a knowledge base returned a 500 server error instead of a clear message. After the fix, a clear duplicate-name error message is shown, and system-managed keys are protected from accidental modification.
* **Fixed the typing indicator flickering after streaming replies end**: When using Web Chat, the "typing" indicator flashed once more after the AI assistant's streaming reply completed. After the fix, the typing indicator behaves normally for streaming replies and the conversation experience is smoother.
* **Fixed parsing failures when uploading Word documents to knowledge bases**: In certain deployment environments, uploading Word documents (.docx/.doc) to a knowledge base failed to parse due to temporary file write permission issues. After the fix, Word documents are processed normally across all deployment environments.
* **Fixed Mermaid diagrams not fully rendered when exporting to PDF**: When exporting a Canvas containing Mermaid diagrams to PDF, the export started before the diagrams finished rendering, resulting in missing or incomplete diagrams in the PDF. After the fix, the export waits for all diagrams to finish rendering before proceeding.
## v2026.05.07
***
### β¨ New Features
* **Member credit quota management**: Administrators can now set credit usage quotas for individual members within the organization. In the "Quota Allocation" tab on the Credits page, they can distribute to a single member, batch-distribute by role, and edit quotas and validity periods, with statistics cards providing a real-time view of allocation status. Members can view their own quota usage progress on their personal Profile page. When credits run short, the system distinguishes between "personal quota exhausted" and "insufficient organization balance" with different messages.
* **Separated contact and conversation tags on the conversation detail page**: The tag system on the conversation detail panel has been upgraded to a dual-block design β "Contact tags" follow the contact and are shared across conversations, while "Conversation tags" can be applied automatically by AI or edited manually. A new conversation tag selector supports search and batch operations, and each chat platform settings page adds a "Manage Tags" shortcut that jumps directly to the conversation tag management tab.
### π Performance
* **Improved PDF translation service stability**: The fonts and model resources required for PDF translation are now preloaded at system deployment, preventing the translation service from failing to start due to runtime environment restrictions. After the fix, the PDF translation feature works normally in all deployment environments.
### π User Experience
* **Automatic brand text replacement in white-label versions**: The white-label admin console now automatically replaces brand names in the interface with the customer's custom name. The replacement covers major interfaces including AI assistant monitoring, the conversation list, the database selector, the role permission tree, and third-party login, ensuring a consistent brand experience.
* **Improved wording for the LINE chat platform "AI assistant thinking animation"**: The "Enable loading animation" option in the LINE chat platform settings has been renamed to "AI assistant thinking animation," more clearly describing this feature as visual feedback while the AI processes a reply. All language versions have been updated in sync.
* **Centered logo in the admin console sidebar**: The logo at the top of the admin console sidebar changed from left-aligned to horizontally centered, staying centered whether the sidebar is expanded or collapsed for a more balanced look.
* **Optimized PDF translation layout warning monitoring**: The handling of layout warnings during PDF translation has been optimized; multiple warnings are now merged into a single notification, greatly reducing monitoring system noise. Translation records add a warnings field, making it easy to view the layout issues encountered by individual translations in the admin console.
### π οΈ Bug Fixes
* **Fixed AI assistants producing an extra response after using the "Ask the user a question" tool**: After an AI assistant used the interactive Q\&A tool to ask the user a question, it produced an additional unrelated response message that was written into the conversation record. After the fix, the AI assistant correctly waits for the user's answer after asking a question and no longer generates extra responses automatically.
* **Fixed Deep Research showing an unclear error when the step limit was exceeded**: When Deep Research exceeded the research step limit, it showed a generic error message that left users unable to understand what happened, and the research state could not be reset. After the fix, a friendly message is displayed and the research state is automatically cleared so users can start over.
* **Fixed action buttons in the contact list accidentally toggling checkboxes**: Clicking the edit or delete button in the contact list also checked that row's checkbox and triggered the batch action bar. After the fix, action buttons only perform their corresponding action without affecting row selection state.
* **Fixed being unable to start a new conversation after deleting one in the Agent Marketplace**: After deleting the current conversation in the Agent Marketplace, clicking the "New Q\&A" button did nothing. After the fix, a new conversation can be created and the page switches normally after deleting a conversation.
## v2026.05.06
***
This update adds a waiting indicator message feature to the Microsoft Teams chat platform, giving users real-time feedback while the AI is processing. It also fixes several important issues affecting billing attribution and system stability, and refines tool category naming to more accurately reflect the organization management structure.
### β¨ New Features
* **Teams chat platform waiting indicator messages**: When messaging an AI assistant via Teams, if AI processing takes a while, the system automatically shows a "Processing" indicator message and periodically updates the elapsed time; the indicator clears automatically after the AI replies. Administrators can customize the indicator message text, initial delay, and update interval in the "Feature Settings" tab of the Teams chat platform settings page.
### π User Experience
* **Updated AI assistant tool category names**: In the tool selector on the AI assistant edit page, "Built-in Tools" has been renamed to "System Built-in Tools" and "User Custom Tools" to "Organization Custom Tools," more accurately reflecting how tools are actually managed β provided by the platform or created by organization administrators.
### π οΈ Bug Fixes
* **Fixed access errors on chat platform settings pages**: When using chat platforms such as LINE, Teams, and Messenger, some users encountered server errors when accessing chat platform settings pages, preventing the pages from loading. After the fix, all chat platform-related pages are accessible normally.
* **Fixed an error in AI assistant template reply mode**: When an AI assistant was set to template reply mode without an attached knowledge base, sending a message caused a server error and no reply was received. After the fix, the system handles this situation normally without errors.
## v2026.05.05
***
### β¨ New Features
* **Burmese support added to PDF translation**: The knowledge base's PDF translation feature now supports Burmese (ααΌααΊαα¬), allowing Burmese PDFs to be translated into other languages and vice versa. The language menu ordering has also been adjusted, placing Asian languages first for quicker selection.
### π Performance
* **Faster Text-to-SQL initialization for large databases**: When connecting to large databases with thousands of tables for Text-to-SQL queries, initialization time drops from several minutes to instant. The system now loads table structures on demand instead of scanning all tables at connection time, dramatically reducing wait time.
### π οΈ Bug Fixes
* **Fixed meeting record permissions displaying in Chinese on non-Chinese interfaces**: When using the English, Japanese, or Korean interface, the names and descriptions of meeting record-related permissions no longer display in Chinese. Also fixed organizations with the meeting record feature enabled so they can correctly see and assign meeting record permissions on the role settings page.
* **Fixed scheduled AI assistants occasionally failing to respond after long-running operation**: Scheduled AI assistants no longer have their responses interrupted by conversation history format issues after accumulating many conversation turns. The system automatically corrects format anomalies in the conversation history, ensuring scheduled tasks keep running stably.
* **Fixed occasional deep research web search failures in live conversations**: When using the deep research feature in live conversations, web searches no longer break off unexpectedly, and search results are returned normally and integrated into the research report.
* **Fixed the response when voice assistant API settings had an incorrect format**: When configuring a voice assistant via the API, if the submitted settings format is incorrect, the system returns a clear error message instead of a server error.
## v2026.05.04
***
### β¨ New Features
* **Contact API credential management**: Previously, API tools could only set unified authentication information at the tool level, with no way to differentiate access for different contacts. Now, in the "API Credentials" tab on the contact detail page, you can independently configure header authentication information for different API tools per contact, suitable for scenarios requiring individualized multi-tenant access control.
* **Q\&A card visual presentation**: After users answered a questionnaire, the Q\&A content used to display as plain text in a chat bubble, indistinguishable from regular conversation. Now, Q\&A results are automatically presented as standalone cards (white background, blue border) with each Q\&A pair clearly separated, making answer records easy to read at a glance.
* **Multi-brand white-label deployment support**: Supports configuring the brand name, support email, and the Ask AI assistant via environment variables, with all interface text automatically applying the brand settings. White-label deployment customers can display the correct brand information without modifying the source code.
### π Performance
* **LDAP sync performance optimization**: For enterprise customers using LDAP identity sync, database queries during large-organization syncs are dramatically reduced. Member permission data is pre-computed to avoid one-by-one queries, significantly speeding up synchronization.
### π User Experience
* **Skeleton loading animation while the AI assistant asks a question**: When an AI assistant uses the "Ask the user" tool, a shimmering skeleton loading animation (including question row and option placeholders) is shown while the question is being prepared, replacing the previous blank waiting state. Once the question is ready, it is automatically replaced by the actual question card, so users clearly know the AI is preparing a question. Supported in both Web Chat and the admin console.
* **Improved deep research task tracking**: The deep research feature's AI research assistant now uses a to-do list to plan research tasks and tracks the completion status of each step in real time. The research process is more organized, reducing duplicated or missed research steps.
* **Organization switching on the no-permission page**: When a user's default organization loses access, the no-permission page no longer offers only "Sign out." The page shows the current organization's name and lists the user's other organizations with access; clicking one switches directly, with no need to sign out and back in.
### π οΈ Bug Fixes
* **Fixed the output panel not opening automatically when Code Interpreter generated files via bash**: When the AI generated files using bash commands (such as creating Excel or PDF files with Python libraries), the output panel did not expand automatically and users could not see the download interface. After the fix, the system detects file write records in bash execution results and automatically opens the output panel. Fixed in both Web Chat and the admin console.
---
# 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/05.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.