> 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/07.md). # July 2026 ## v2026.07.27 *** ### ✨ New Features * **Interactive App Cards in Conversations (MCP Apps)**: Interactive cards (with videos, buttons, and external links) can now be rendered directly in the conversation message flow, with support for full-screen mode and continuing the conversation via "Ask AI". (Previously: AI replies could only be presented as text, so interactive scenarios such as surveys and plan selection could not be completed within the conversation.) * **Doc Quick Entry**: The conversation input bar now lets you browse, search, and preview document templates and apply one with a single click, with fidelity indicators showing how well the layout is preserved. (Previously: applying a document template required leaving the conversation to work elsewhere, and template application could not guarantee layout fidelity.) * **MaiGPT Per-Role Capability Permissions**: Role settings gain a new permissions tab that lets you control models, built-in features, official tools, skills, and connectors on a per-role basis; unauthorized items are simply hidden. (Previously: there was no way to restrict which MaiGPT models and features members could use by role.) * **Knowledge Base Full-Document View**: The file content page adds a toggle between full-document and chunk-card views, reassembling the original text with dual-track markers showing chunk boundaries and overlap regions. (Previously: after a document was chunked, it could only be viewed chunk by chunk, with no way to see where each chunk sat in the original text or how chunks overlapped.) * **Crawl Task Names and List Filters**: Crawl tasks can now be named at creation, and the list adds a search box plus type and status filters; schedule-triggered tasks are automatically named with the schedule name and date. (Previously: crawl tasks had no identifiable names, and the list could not be searched or filtered.) * **Crawl Result Preview**: Before submitting, you can preview the list of URLs that will be crawled and those excluded (completes in seconds), avoiding wasted crawl quota from misconfigured rules. (Previously: after configuring crawl rules, you had to run the full crawl to find out which pages it would cover, costing both time and credits.) * **AI Rule Assistant Prompt Persistence and Recall**: The original description given to the AI Rule Assistant is saved when a task is created, and automatically brought back for adjustment when editing a schedule. (Previously: after generating crawl rules from natural language, the original description was not saved, so there was no way to know where the rules came from when editing a schedule.) * **Self-Service Account Deletion and Admin Account Erasure**: The profile page adds an account deletion feature (with impact preview and confirmation flow); the platform admin side adds a batch account erasure operation. (Previously: personal data deletion requests could not be completed self-service, and the platform side lacked a reliable account erasure tool.) * **Member Credit Quota Alerts**: The Credit usage page adds alert settings that automatically send an in-app announcement and email when a member's usage reaches a threshold or their quota is about to expire. (Previously: there was no advance warning before a member's quota was nearly used up or about to expire β€” they only found out when they were blocked.) * **Team API Integration Tab**: The team edit page adds an API integration tab providing complete API examples consistent with the AI assistant page. (Previously: the team edit page had no API call examples, so developers had to piece together the completions URL themselves.) * **Team Conversation-Level Token Budget Guardrail**: Team safety guardrails add a conversation-level token budget cap that stops execution once the cumulative total is reached, and clarify the naming of the per-message cap. (Previously: the existing token cap only limited a single run and could not govern cumulative consumption across a whole conversation.) * **MaiGPT Focus Mode**: Collapse both side panels with one click so the conversation area fills the full width; hover at the edge to temporarily preview a sidebar. (Previously: the left navigation bar and conversation history panel took up width, limiting the conversation area.) * **Meeting Record Member Access Management**: Meeting details add member access list management, letting you add and remove members who can view the meeting. (Previously: there was no way to configure which members could access a meeting record.) * **Meeting Record Source Labels and Audio Source Billing Details**: The meeting list adds a source column (Desktop/Web), Credit transaction details label transcription charges by microphone versus system audio, and an auditable dual-track settlement mechanism is established. (Previously: there was no way to tell which device created a meeting or which audio source a charge came from.) * **Organization Application Approval Flow**: An organization approval flow can be enabled: applicants fill in application details when creating an organization, and the trial plan is granted only after approval; both pending and declined states show corresponding notices. (Previously: the trial plan was granted automatically at organization creation, which was abused by repeatedly creating organizations for unlimited trials.) * **SynxDB Database Support**: Text to SQL external databases and knowledge base vector databases add a SynxDB option (a PostgreSQL-compatible MPP data warehouse). (Previously: Text to SQL and knowledge base vector stores did not support SynxDB.) * **On-Premises Offline Speech-to-Text (LocalAI)**: Adds LocalAI offline speech-to-text support, usable for both knowledge base audio files and offline meeting record imports, with language and prompt settings. (Previously: on-premises environments could not use cloud speech-to-text services.) * **Voice Agent Adds gpt-realtime-2.1 Model**: The voice agent model list adds gpt-realtime-2.1, with corresponding pricing. (Previously: OpenAI's latest realtime voice model could not be selected.) * **One-Paste Auto-Split of Credentials on the Login Page**: Pasting space-separated credentials into the account field automatically splits them into the fields and submits the login, with support for multiple consecutive spaces as separators. (Previously: a copied "account password" pair had to be pasted manually in two steps before submitting.) * **Organization and Conversation Record Deep Links**: Admin console URLs support organization and record parameters, so opening a link automatically switches to the specified organization and expands the conversation record; users who are not logged in are redirected to login first and then back to the original page. (Previously: external links could not open a specific organization and conversation record directly β€” users had to switch manually.) * **MaiGPT Connector Preference Memory**: Each member's connector enablement preferences are remembered automatically, and new conversations carry over the previous settings. (Previously: frequently used connectors had to be re-enabled at the start of every new conversation.) * **MCP Connectors Support the CIMD Authorization Mode**: Supports the OAuth Client ID Metadata Document mode, so these services can now connect normally. (Previously: some MCP services (such as Tableau hosted MCP) support neither dynamic registration nor pre-registered credentials, making OAuth connection impossible.) ### πŸš€ Core Performance Optimization * **Faster Tool and Skill List Loading**: List queries were refactored so load time no longer grows with platform data volume, and a cross-organization information leak was fixed. (Before: the attached tools and skills menus loaded increasingly slowly as data grew, especially with large skill packs.) * **Faster Batch Q\&A Excel Export**: Duplicate queries were eliminated so the export completes in a single query. (Before: exporting large volumes of Q\&A records was slow due to duplicate queries.) * **MaiGPT Conversation List Loading Optimization**: Per-row duplicate queries were eliminated, making the list load faster. (Before: each row in the conversation list triggered an extra query, slowing loading when there were many conversations.) ### 😊 User Experience & Interface Optimization * **Complete Connector Experience Overhaul**: Adds a connector card selector and dedicated management page, second-column navigation in the tools area, clarified authorization subjects (shared vs. per-contact authorization) with corresponding guidance, and fixes the authorization flow for per-contact-authorized tools. (Previously: connector selection, creation, and management were scattered across nested modals, authorization subjects were unclear, and deleting a connector silently affected tools sharing the same URL.) * **MaiGPT Attachment Menu Optimization**: The menu adds a scrollable fade hint, and enabled connectors are pinned to the top with a badge. (Previously: the attachment menu was scrollable with no hint, and enabled connectors were hard to find mixed into the list.) * **Unified In-Progress Chat Indicator Visuals**: In-progress text is unified with a shimmer effect, the typing indicator dots sit outside the message box, and the bubble appears only once output begins. (Previously: waiting indicators were inconsistent β€” a mix of empty bubbles, spinners, and text.) * **UI Component Upgrades: Model Selector and Skill File Browser**: The model selector is upgraded to a command-palette style (search, keyboard navigation, full accessibility support); the skill file directory becomes a GitHub-style file tree. (Previously: the model selector and skill file directory had dated interaction experiences.) * **MaiGPT Conversation List Right-Edge Cleanup**: Action buttons sit flush against the right edge and the scrollbar fades out on hover, keeping the right edge clean and uncluttered. (Previously: hovering over a conversation row crowded the action buttons against the scrollbar with extra blank space.) * **Model Dropdown Ordering Fix**: Platform models are always listed first in curated order, with organization-created models sorted after them by creation time. (Previously: organization-created models were inserted out of order among the platform models.) * **Logged-In Users Also Fill In the Data Collection Form**: The flow now shows the form after login before entering the conversation, and the login identity is not overwritten by the form. (Previously: logged-in users were never asked to fill in the data collection form.) ### πŸ” Security & Governance Enhancements * **Web Chat Canvas Content Security Hardening**: HTML content is now rendered in an isolated sandbox and the fallback display path is sanitized, so malicious scripts cannot access the user's page. (Previously: AI-generated HTML content carried a cross-site scripting (XSS) risk.) * **Human Verification and Rate-Limiting Protection**: Introduces Cloudflare Turnstile human verification (invisible verification on the login page, escalating to an interactive challenge past the limit), WebSocket message rate limiting, and authentication endpoint throttling, with limits adjustable from the admin side. (Previously: entry points such as login and registration lacked protection against bots and malicious traffic.) * **Web Chat Embed Origin Allowlist and Public Endpoint Throttling**: Each Web Chat can configure an allowlist of domains permitted to embed it; public endpoints gain per-IP rate limits. (Previously: Web Chat could be embedded by unauthorized domains, and public endpoints could be called without limit.) * **Conversation API Deny-by-Default Permissions**: Permissions now default to deny, with only the operations required for anonymous Web Chat explicitly opened, eliminating silently public endpoints. (Previously: conversation API operations not explicitly listed were open to unauthenticated users by default.) * **Anonymous Upload Throttling and Scope Reduction**: Anonymous uploads gain per-IP rate limits and are restricted to chat attachments only; all other uploads require login. (Previously: anonymous upload URLs could be obtained without limit and used to upload arbitrary files.) * **Attachment Storage Billing and Expired Attachment Cleanup**: Adds billing items for attachment storage and derived storage, and periodically cleans up expired attachments based on conversation activity (active conversations are unaffected). (Previously: the storage cost of attachments and derived files was not reflected in Credits, and old attachments accumulated indefinitely.) * **Credit Billing Accuracy Improvements**: Prompt cache reads and writes are billed correctly per provider discounts (for example, OpenAI cache reads get the discounted rate), and team runs and interrupted streaming scenarios are billed correctly, keeping accounts consistent with actual usage. (Previously: multiple billing paths under- or mis-billed: prompt cache usage for team runs and OpenAI models was not recorded, and gateway requests whose streams closed early were not billed.) ### πŸ› οΈ Bug Fixes * **Fixed Normal Agent Team Tool Flows Being Misjudged as Loops and Interrupted**: When a team's entry Agent used two or more tools in one round (such as delegation plus a knowledge base query), the safety mechanism misjudged it as a loop and force-stopped it, truncating the reply. The misfiring consecutive-step check was removed while keeping the token budget and iteration cap guardrails, so normal multi-tool flows are no longer interrupted. * **Fixed Voice Agent Delegated Queries Returning Empty Replies**: During voice conversations, Agent-delegated queries always received an empty reply and could not answer. The voice delegation path is no longer wrongly blocked by the human-handoff check, restoring normal voice Q\&A. * **Fixed Text-to-Image Generation Failing**: After Google retired the preview image generation models, all text-to-image calls failed. Switched to the GA Gemini image generation models, with existing tools and pricing converted automatically, restoring image generation. * **Fixed Team Entry AI Assistant Being Accidentally Deletable**: The entry AI assistant auto-created for a team could be accidentally deleted from the assistant list, breaking the linked conversation platforms. Entry assistants no longer appear in the assistant list and cannot be deleted through assistant management; they are removed only when the team itself is deleted. * **Fixed Images Uploaded via Microsoft Teams Not Being Readable by AI**: Images uploaded through the Teams attachment flow were treated as generic documents, so the AI could not read them and the interface showed a file card instead of an image preview. Files are now correctly identified as images by their actual type, restoring AI image reading and image previews. * **Fixed AI Slides Element Selection and Editing Issues**: Some slide decks could only be edited page-wide, clicking a component showed no selection frame, and the selected shape did not match what actually got modified. Element selection and edit targeting now align with visible content, and unlabeled shapes can be correctly selected and edited. * **Fixed MaiGPT Waiting Indicator Display Glitches**: The waiting indicator could linger on multiple historical messages at once or float in the wrong position. The indicator now appears only on the latest message and disappears once the reply completes. * **Fixed AI Assistant Message Text Color Preview Not Responding**: Adjusting the AI assistant message box text color in appearance settings did not update the live preview on the right. The preview now reflects text color settings in real time. * **Fixed Third-Party Login Failing After Saving Settings**: After saving third-party login from the settings page, all logins were rejected, and the email domain field could not be configured. Saving the settings now automatically enables that login method, and the form adds the email domain field. * **Fixed "Create Organization" Occasionally Not Responding to Clicks**: The "Create Organization" item in the organization dropdown occasionally did nothing when clicked. A click timing conflict was eliminated, so clicking reliably opens the creation dialog. * **Fixed Jagged Button Text in the Onboarding Tour**: Button text in tour popovers showed jagged white edges. Removed the redundant text shadow, making button text crisp. * **Fixed Ask AI Sidebar Covering Host Site Modals**: After the sidebar opened, modals subsequently opened by the host site were hidden behind it and could not be operated. While the sidebar is open, newly appearing full-width modals are automatically detected and shrunk, making them operable again. * **Fixed Knowledge Base Audio Transcription Failures**: Some audio files (such as larger m4a files) were reported as invalid format when going through batch transcription. Audio is now normalized uniformly before transcription, so previously failing files process correctly. * **Fixed Parsing Failures for Certain PDFs**: Some PDFs exported from Word failed to parse entirely due to abnormal internal structures. Abnormal structures are now skipped automatically so processing continues, and the whole document parses correctly. * **Fixed Conversation Platform Unread Count Not Resetting**: After an unread conversation was deleted, the unread badge count lingered while the list showed no unread conversations. Deleting an unread conversation now releases the unread count in sync, and existing inconsistent data across the platform was corrected. * **Fixed Irrelevant Results in Member Search**: Searching short strings (such as 00) matched internal IDs not even visible on the page, surfacing irrelevant members. Search now matches only fields visible on the page (name, email, member number). * **Fixed Some MCP Tools Silently Failing**: Some tools (such as Tableau data queries) were quietly removed from the available list due to spec parsing failures, leaving the AI to guess tool names blindly. Spec parsing compatibility was strengthened, restoring the affected tools. * **Fixed Hooks Unable to Save Due to a Deleted Assistant**: After an AI assistant linked to a Hook was deleted, the Hook page still showed that assistant and every save failed. Deleting an assistant now cleans up Hook associations in sync, so Hook settings can be saved again. * **Fixed AI Agent Monitoring Dashboard Data Anomalies**: Trend chart timelines had gaps, error type fields were inconsistent, load failures showed fake empty data, and rapidly switching filters mixed old and new data. Timeline gaps are now filled, error types unified, failures clearly indicated instead of showing empty data, and data always matches the current filters. * **Fixed Copilot Page Getting Stuck After Switching Organizations in Another Tab**: After switching organizations in another tab, the Copilot page's conversation lookup got stuck and could not load. When stuck, it now automatically follows the organization switch and resumes loading; conversations in use are unaffected. * **Fixed LINE Quick Reply "Prefill Input Box" Sending Immediately**: Quick reply buttons configured to prefill the input box sent the message immediately when tapped on LINE, inconsistent with Web Chat behavior. Tapping on LINE now opens the keyboard with the text prefilled, without auto-sending. * **Fixed On-Premises Database Connection String Compatibility**: Database connection strings configured per the installation guide raised errors when creating tables from CSV uploads. Legacy connection string formats are now handled automatically, so configurations following the guide work correctly. * **Fixed On-Premises Single-Cell Spreadsheet Parsing Failures**: Excel/ODS files containing only a single cell failed to parse after upload. Short-content spreadsheets are now correctly treated as data, and this fix has been backported to the on-premises version. ## v2026.07.20 *** ### ✨ New Features * **AI Gateway Custom Models Can Be Assigned Directly as AI Assistant Models**: Bring-your-own-key (BYOK) custom models configured in an organization's AI Gateway can now be selected directly from the model menus in AI assistant settings and MaiGPT conversations, with badges marking custom models in the menu. Tool calling, image understanding, and streaming replies are all supported, and billing uses a service-fee markup model rather than full-price metering. * **AI Gateway Fallback Group Automatic Failover**: With a fallback group configured, when the primary model's upstream service fails (such as overload, timeout, or connection errors), the system automatically retries with the next model in the group, so conversation service is not interrupted by a single model provider's outage. * **Voice Assistant Supports Knowledge Base and Agent Mode**: Voice conversations are no longer limited to the real-time model's general answers β€” substantive questions are delegated to the AI assistant's full question-answering pipeline (using knowledge base Q\&A or Agent mode depending on the assistant's settings). The call and transcript interfaces present the delegation process and sub-agent replies as task cards, and voice transcripts are clearly labeled. * **Inline File Preview Panel**: The output panels in MaiGPT and Web Chat now include embedded file previews β€” files in the knowledge base sidebar, file cards generated in conversations, and files in the Output sidebar can all be previewed directly in the right-hand panel (PDF, Office documents, CSV, Markdown, images, and more), with no need to download files before viewing them. * **In-Place AI Editing for AI Slides**: Slide projects gain four AI editing capabilities β€” selected-page rewrite, in-place text block editing, AI optimization suggestions, and whole-deck style unification. Every edit automatically creates a version snapshot, so you can compare before-and-after differences at any time and restore with one click. * **AI Slides Quick Entry and Template Management**: The MaiGPT home page adds a quick entry for AI Slides, letting you pre-specify template, page count, and other settings before generation and skip the step-by-step Q\&A. Slide templates support a "My Favorites" collection, and you can import your own templates (PPTX, PDF, JSON, or URL). * **Offline Audio Transcription for Meeting Records**: Meeting records support uploading existing audio files for offline transcription β€” no live recording required. Upload a recording after the meeting to generate a transcript, with support for long audio files and a reliable failure-retry mechanism. Upload progress is shown in real time, and the status updates automatically upon completion. * **MCP Apps Interactive Tool Cards**: Web Chat supports MCP Apps (the official MCP UI extension standard) β€” for AI assistants with MCP connectors attached, tool results can be rendered directly in the conversation as interactive app cards (such as a timetable picker). Interaction results are sent back to the conversation in real time for the AI to continue processing, eliminating back-and-forth text descriptions. * **Standard OAuth Connection for MCP Clients**: When MCP clients such as Claude Desktop, Claude Code, and claude.ai connect to MaiAgent, they can now sign in via the standard OAuth 2.1 authorization flow instead of manually copying and pasting an API Key. Organizations can also self-manage OAuth client applications through the API (including key regeneration and revocation). * **Agent Teams Support the Completions API**: AI assistants bound to an Agent Team can now be invoked through the Completions API (both streaming and non-streaming). Agent handoffs are returned in real time, and billing and handoff state persistence match web conversations; API integrators no longer receive "Team mode not supported" errors. * **Completions API Supports Human Handoff**: Conversations integrated through the Completions API can now use the human handoff flow (service-hours checks, queuing, offline notifications). Handoff-related notifications are snapshotted at creation time, so the handoff messages seen by visitors, the admin console, and external channels are fully consistent. * **Cloudflare Access Third-Party Login**: Added Cloudflare Access (zero-trust network access) as a third-party login method β€” organizations can let members sign in to the platform directly after authenticating through Cloudflare Access, with a configurable option for automatic member account provisioning; unauthorized accounts leave an audit record. * **Custom Context Data Injection for Embedded Chat**: Integrators who embed Web Chat into their own websites can inject custom data (such as medical record numbers or identity keys and values) when opening the chat. This data is provided to the AI assistant with every message as conversation context, enabling personalized responses based on the visitor's identity; it is automatically cleared on identity switch or logout to prevent data residue. * **Support for OpenAI gpt-5.6 Series Models**: The platform now supports OpenAI's latest gpt-5.6 family (Sol / Terra / Luna), which works normally with tool calling; organizations can select these models from the model menu. * **Database Q\&A Supports SynxDB**: Database Q\&A (Text-to-SQL) adds SynxDB as a database type option, so organizations using SynxDB data warehouses can connect directly for natural language queries. * **Credit Consumption Columns Added to Chat Record Export**: The Excel export of AI assistant chat records adds 6 credit columns (vectorization, LLM input, LLM output, tools, other, and total). Exported figures match the AI assistant monitoring dashboard, making reconciliation and cost analysis easier. ### πŸš€ Core Performance Optimization * **Parallel Processing for Test Dataset Evaluation**: Dataset evaluation now runs in multiple parallel batches, dramatically shortening overall evaluation time for large datasets β€” for example, 202 test cases previously took about 100 minutes with sequential processing; with parallelization, the time shrinks proportionally to the number of concurrent batches. * **Faster Keyword Search in the Conversation List**: Fixed the "interface request timeout" error when large organizations (over 1.5 million conversations) searched member names by keyword in the conversation list β€” query time dropped from over 27 seconds to about 13 seconds, no longer exceeding the interface wait limit. * **Conversation Platform Tab Count Query Optimization**: A dedicated index was created for the conversation count queries behind each conversation platform tab (pending, in progress, etc.). Counts are now completed directly from the index without reading rows one by one, eliminating platform-wide database pressure spikes during peak hours. * **Document Parsing Stability Improvement**: The knowledge base's deep document parsing (Docling) adds a cluster-wide concurrency cap, preventing multiple large documents parsing simultaneously from exhausting memory and interrupting the parsing service, making document processing more stable and reliable. ### 😊 User Experience & Interface Optimization * **MaiGPT Input Bar Overhaul**: The model selector moves inside the input bar (matching the inline menu style of mainstream AI chat tools). When too many skills and tools are selected, they automatically collapse into a count chip that expands on click. Also fixed the issue where the "+" menu's submenu got stuck open when hovering over "Upload File" or "Deep Research." * **MaiGPT Conversation Sidebar and Search Interface Optimization**: The conversation history sidebar unifies group headers, loading skeletons, and list layout; the search panel's row spacing and heights are standardized, with loading skeletons aligned to actual data; conversation titles are no longer truncated early by action buttons; and reply streaming renders more smoothly without text flickering. * **New Visual Style for Meeting Transcripts**: Plain-text meeting transcripts in meeting records adopt a new visual style with clear paragraph separation for easier reading, plus paragraph focus support; blank paragraphs are automatically filtered from display. * **Shimmer Effect for In-Progress States**: All "in progress" text hints in conversations (tool executing, analyzing results, thinking, skill in use) now use a unified shimmer sweep animation, consistent with the thinking block's existing animation, making in-progress states recognizable at a glance. * **Web Chat Loading Indicator and Agent Switch Card Sync**: Web Chat's message loading indicator and multi-agent switch cards now align with the admin console β€” a consistent loading animation shows while waiting for AI replies, and a switch card appears when a multi-agent assistant switches agents, so visitors can clearly see which agent is currently serving them. * **Credit Quota Display Improvement**: The credit quota list shows the full expiration date and time (no longer just the date), and usage is displayed on two lines with thousands separators. The quota card format on the profile page is aligned to match, and the "N days remaining" expiration reminder displays correctly again. * **Deep Links in System Notification Emails**: The buttons in two types of system emails β€” low credit balance alerts and data retention warnings β€” are now deep links: clicking them automatically switches to the corresponding organization and opens the relevant settings dialog directly, with no need to find the organization and feature entry yourself. * **Image Search Toggle and Query Scope Limits**: AI assistant settings add an "Image Search" toggle to disable the automatically enabled image search tool. Image search now also respects the knowledge base and file scopes specified by the API, no longer searching knowledge base content outside the specified scope. ### πŸ” Security & Governance Enhancements * **Outbound Request Security Protection (SSRF)**: Outbound requests from API tools and scheduled event notification (Webhook) deliveries now have comprehensive destination address protection β€” requests targeting internal networks, cloud metadata, and other private addresses are blocked, with re-validation on every connection (including redirects), preventing probing of internal systems through tool configuration. * **Strengthened Read-Only Protection for Database Q\&A**: Database Q\&A (Text-to-SQL) adds two layers of write protection β€” syntax checking now validates statement by statement (blocking bypasses that smuggle writes via multi-statement queries), and before query execution the database connection is forced into read-only mode so the database itself rejects any write. * **API Pagination Size Cap**: Platform API pagination now has a uniform cap of 100 items per page, preventing abnormally large single queries from causing full-table database scans and safeguarding overall service stability; normal interface operations are unaffected. * **Keycloak SSO Certificate Verification Enabled by Default**: Server certificate verification for Keycloak single sign-on is now enabled by default, eliminating man-in-the-middle attack risk. Customers with special certificate chain situations can be individually exempted by the platform; all other organizations (including future ones) use the secure default. * **Tamper-Proof Protection for Audit Logs**: Platform audit event records (23 categories of operation trails) now have database-level protection against modification and deletion (append-only), and audit records are periodically archived to write-once immutable storage β€” meeting ISO 27001 requirements for tamper-proof audit logs and strengthening enterprise compliance evidence. * **Private Network Attachment Uploads for Enterprises**: Enterprise customers connecting to the platform via site-to-site VPN can route attachment uploads through the customer's private network path instead of a public internet endpoint β€” API traffic and file uploads stay on the private line end to end. * **Hook System Included in the Enterprise Plan**: The Hook system (event hook automation) is now officially part of Enterprise plan entitlements β€” Enterprise plan organizations automatically gain access, which is automatically revoked when the plan expires or is downgraded; the platform retains the flexibility of manual activation for individual organizations. ### πŸ› οΈ Bug Fixes * **Fixed AI-Generated Videos Not Visible in Web Chat**: After an AI assistant generated a video, Web Chat visitors previously could not see it (only the admin console could). Generation progress and video playback are now fully displayed, and an intermittent record-creation error that interrupted follow-up replies after video generation completed has also been fixed. * **Fixed Batch Image Generation Showing Only the First Image**: When batch-generating multiple images at once in MaiGPT, Web Chat previously showed only the first image; all generated images are now fully displayed. * **Fixed Unable to Create Conversations After Google Login in Web Chat**: After visitors signed in to Web Chat with a Google account, creating a conversation previously produced an "authentication credentials were not provided" error (username/password login worked fine). Visitors signing in with Google are now automatically added to the organization the Web Chat belongs to, and conversation creation works again. * **Fixed Incorrect Credit Billing Records for Image Processing**: Two image-related credit billing errors have been corrected β€” conversation attachments parsed with AI vision are now billed by actual processing volume (previously some usage was unbilled or billed under the wrong category), and knowledge base image vectorization volume is now recorded correctly. Transaction details also add a vision-parse consumption hint, clearly distinguishing it from regular conversation consumption. * **Fixed Agent Team Conversation Interruptions on External Channels**: When AI assistants bound to an Agent Team served LINE and other external channels, each message previously opened a new conversation (agents completely lost context), and the Team edit page also failed to open. Conversations now continue correctly, and the edit page works again. * **Fixed Errors in Knowledge Base Batch Re-Parsing**: Running batch re-parse on files without a designated parser (auto-detect state) previously returned a system error; re-parsing now completes normally. * **Fixed Text-to-Image Failures with CMYK Reference Images**: Generating images with a print-color-mode (CMYK) image as the reference previously failed outright; the system now automatically converts the color mode before submitting, and generation works again. * **Fixed Text Extraction Errors in Certain PDFs**: Some PDFs exported from Word contain faulty embedded accessibility tag information, causing text extracted after upload to a knowledge base to differ from the actual page content (for example, numbers becoming dashes). Parsing now skips the problematic tag information and uses the actual page text. * **Fixed File Processing Failures from Knowledge Base Custom Field Type Conflicts**: Custom fields on knowledge base files (such as version numbers or URLs) could previously cause batch processing failures from mixing numeric and text types (this once caused processing errors for hundreds of files). All custom field values are now normalized to text format when written to the search index, so type conflicts no longer occur. * **Fixed Reply Failures from Conflicting Thinking Settings**: If an AI assistant's thinking budget exceeded its maximum output length (for example, when settings were not synced after switching models), certain models rejected all requests. The system now automatically adjusts the thinking budget into a safe range, and replies are no longer interrupted. * **Fixed Reply Interruptions When AI Assistants Process Many Files**: When Agent mode processed a large number of files at once (such as dozens of spreadsheets), accumulated content could previously exceed the model's limit, causing the entire reply to fail and all completed work to be lost. The system now automatically keeps accumulated content within the model's processing range and, when the limit is genuinely exceeded, shows an understandable error explanation (rather than a technical error message). * **Fixed Intermittent Download Failures for Code Interpreter Output**: When viewing files produced by the code interpreter, a system error previously occurred if the workspace snapshot was updated during the download; the system now automatically retries with the latest version, and downloads work again. * **Fixed Not Redirecting to the Original Target Page After Login**: When opening a shared link (such as an Agent Marketplace conversation share) while logged out, users previously landed on the home page after login instead of the page they originally intended to visit; both password and Google login now correctly return to the original target page. * **Fixed Pinned Organizations and Remembered Organization Disappearing**: The pinned organization list and the last selected organization previously vanished entirely after 7 days of inactivity (a cache expiration mechanism was applied by mistake); they are now retained permanently, and existing pins that have not yet expired are automatically rescued. * **Fixed Duplicate "Typing" Indicators in Web Chat**: During AI replies (especially when the connection briefly dropped and reconnected), Web Chat could previously show two "typing" bubbles at the same time; only one is now shown at a time, positioned consistently below the latest message. * **Fixed Raw Translation Codes Showing in the Interface**: The "Knowledge Base β†’ Tag Management" edit dialog in the English and Simplified Chinese interfaces, and the crawl parameter preview item in credit usage, previously displayed raw translation codes instead of normal text; all missing translations have been filled in, and the five locales now display consistently. * **Fixed Ask AI Opening a Non-Functional Window Before Joining an Organization**: When newly registered users who had not yet created or joined any organization clicked Ask AI, a non-functional chat window previously opened; a clear prompt now appears, guiding them to create or join an organization first. * **Fixed AWS Marketplace Subscription Message Reception Failures**: For customers subscribing to the platform through AWS Marketplace, subscription status notifications were previously rejected by the system, delaying subscription processing; the reception chain has been repaired, and trial credit issuance is isolated to prevent double-stacking with official subscription credits. ## v2026.07.13 *** ### ✨ New Features * **One-Click Copy and Multi-Format Export for Meeting Summaries**: The meeting summary tab in Meeting Records adds an export toolbar: copy the full summary with one click, export it as Markdown, PDF, Word (docx), or plain text, or send the summary content directly by Email β€” no more manual select-copy-paste. * **Web Chat Supports Google Account Login**: The Web Chat native login page adds a "Sign in with Google" button, so end users can sign in to conversations directly with their Google account without registering separate credentials. Takes effect after enabling the Google login toggle in the admin console's login settings. * **Automatic Parser Detection on Knowledge Base Upload**: When uploading PDF, Word (docx), or PowerPoint (pptx) documents to a knowledge base, the system automatically determines whether the document is a scan or image-heavy content and switches to the most suitable parser (such as OCR or vision parsing), so users no longer have to judge for themselves. The parser dropdown's default item in the upload dialog shows "Auto-detect (default)," and the parser ultimately used can be confirmed and changed in the file list. * **Per-Call LLM Usage Details Added to Quality Control Detail**: The token statistics section of the AgentOps quality control detail adds an expandable "LLM Call Details" block: every model call in an Agent mode reply is listed with its sequence number, duration, input/output tokens, and cache write/read tokens, making it easy to analyze where token usage grows across multi-round Agent iterations. * **Live Progress and Stall Alerts for Automated Evaluation**: While an AgentOps automated evaluation runs, the completed/total counts on the detail page now tick up in real time; if an evaluation goes more than 10 minutes without any progress update, the page shows a persistent "may have been interrupted" warning, so users can spot stuck evaluation tasks early instead of refreshing manually over and over. * **Billing Page Separation for Partner-Managed Organizations**: For organizations whose plans are managed by a partner (reseller/agency), the billing management page now shows a "Your plan and Credits are managed by your partner" notice card and hides all purchase and upgrade entry points, preventing accidental direct purchases from the platform; the Credits usage page remains visible. The billing page also gains proper loading and load-failure states, with a retry button on failure. * **Credit Billing Introduced for Attachment Parsing**: Attachment parsing by AI assistants in conversations now consumes credits: document parsing is billed by text node count and audio transcription by seconds, with billing aligned with knowledge base parsing, and usage shown on the conversation dashboard. Audio transcription also moves to background worker processes: shorter audio files get their transcript in the same turn, while longer ones automatically continue the reply once transcription completes β€” resolving the earlier issue where audio files could not be parsed in some environments. ### πŸš€ Core Performance Optimization * **AI Reply Cache Hit Rate Optimization**: The prompt caching mechanism for AI replies has been optimized: knowledge base retrieval content and time/location information that changes each turn are moved out of the model's cache scope, so multi-turn conversations hit the cache consistently. Measured over three conversation turns, model input usage cost dropped by about 53%, with the biggest credit savings in high-traffic, long-conversation scenarios. * **Gemini Model Connection Stability Improvement**: Connection management when using Gemini (Vertex AI) models now uses a shared connection pool, fixing the earlier problem of idle connections accumulating on every request and improving service stability under sustained high traffic. ### 😊 User Experience & Interface Optimization * **Web Chat Message Interface Upgrade (Tool Cards, Research Plans, and Inline Citations)**: The Web Chat message interface syncs the admin console's new design with four improvements: 1. When an AI reply pauses for more than 0.8 seconds, a contextual hint appears ("Analyzing results…" after a tool finishes, "Thinking…" while thinking), so long waits are no longer blank; 2. Knowledge base retrieval and web search results are presented as cards that expand to show source files, relevance, and search links; 3. When the AI plans a task, a "research plan" list appears with per-item progress; 4. Citation numbers in answers become interactive inline citation markers β€” hover to preview the source file name, original passage, page number, and relevance, and click to open the source file. * **MaiGPT Interface Redesign (Sidebar, Quick Search, and All Chats Page)**: The MaiGPT interface has been fully redesigned: 1. The left sidebar adopts a new layout with clear sections, a narrower default width, and drag-to-resize (the width is remembered); 2. A new cmd+K (Ctrl+K on Windows) quick search panel searches conversation titles and message content in real time; 3. A new "All Chats" page centralizes browsing and searching all conversation history, with keyword hit highlighting and infinite scroll. * **FAQ Bulk Import Moved to Background Processing with Import History**: Excel bulk import of knowledge base FAQs now runs as background batch processing: you can leave the page after uploading, and the system completes the import in batches, greatly improving stability for imports at the tens-of-thousands scale. The bulk upload dialog also adds an "Import History" tab showing each import's processing progress and per-row success/failure results. * **Cited Documents Show Retrieval Markers for Knowledge Base Sources**: In a conversation's "Cited Documents" block, each cited document's custom fields are now marked with a πŸ” retrieval marker, letting users see at a glance which fields were sent to the AI model to participate in answering. The markers reflect the knowledge base's current field settings. * **Unified Resource Permission Display Rules**: Permission presentation is now fully unified across seven resource types β€” AI assistants, knowledge bases, conversation platforms, databases, tools, Agent UI tools, and skills: action buttons the user lacks permission for are always disabled with a hover explanation of why (distinguishing "role lacks permission" from "quota exhausted"), so filling out a form only to hit an error no longer happens; without edit permission, edit and settings pages are truly read-only (including the knowledge base's document upload, tags, custom fields, FAQ, and other sub-pages), making "view" mean what it says. * **Web Crawl Request List and Detail Page Information Completed**: The web crawl request list adds an ID column (one-click copy) and a type column (one-time / scheduled). The detail page adds a progress percentage and per-page completion times, and provides a step-by-step read-only recap of "what settings this crawl request was created with" (URL, crawl settings, execution mode, schedule, knowledge base import settings). * **Unified List Status Tag Styles**: Status and type tags across the admin console's list pages (crawl import status, evaluation results, error types, tool execution status, export requests, schedules, connectors, API records, organization member status, and more β€” over 17 locations) are unified into a consistent pill style. Color semantics are unchanged, and visual consistency across pages is improved. * **Left Sidebar Menu Reordering**: The admin console's left sidebar menu is reordered according to product planning: the 12 top-level items and each group's sub-items are arranged in a new order, and "Agent Marketplace" moves up from 9th to 3rd, making it easier to find. Feature names, paths, and permissions are all unchanged. * **More Varied Typing Wait Animations**: The waiting animation before an AI reply adds two new styles, "bubble" and "orbit," which rotate randomly with the original breathing blink for three variants total (the same one never appears twice in a row). Animation pacing and trajectories have been fine-tuned based on feedback, and the system respects the user's reduced-motion preference. * **Animated Entry Splash**: The loading screen when entering the admin console is now animated: the logo is painted in with a brush stroke, then the brand wordmark appears letter by letter in a handwriting style. The system loads normally in the background, and the splash fades out only after the animation finishes and the home page has rendered, avoiding a glimpse of an unfinished blank screen. On very slow networks it automatically falls back to the static logo without affecting entry. * **Wording Adjustment in Human Handoff Notification Settings**: In the "Notifications" sub-tab of human handoff settings, fields like "CC Members / CC Groups" are reworded to "Send Copy Notification (Members) / (Groups)," with related help text adjusted accordingly, lowering the comprehension barrier for customer service staff without technical backgrounds. * **Conversation Titles Fully Preserved Without Truncation**: Titles of newly created conversations now fully preserve the first message content β€” no longer fixed-truncated at 20 characters, and no longer forcibly prefixed with a person's name. Display length is now adjusted automatically by the interface based on column width, and with the resizable sidebar you can see more complete titles. Existing conversation titles are unchanged. * **Embedded Web Chat SDK Mobile Experience Improvements**: Three mobile improvements for the embedded Web Chat SDK: 1. On iPhone, tapping the input box no longer auto-zooms the page (effective by default; partners need no configuration changes); 2. The embedded window allows the geolocation permission, so AI assistants with GPS-aware replies can obtain coordinates normally; 3. When switching locale via the SDK, the opening message of a conversation that has not yet started is automatically regenerated in the new language. ### πŸ” Security & Governance Enhancements * **Strengthened Path Protection for Download Links**: The download feature for files produced by AI tools now uses stricter path validation: whatever form the input takes (including multi-encoded variants), the file path actually read is physically confined to the output directory, eliminating the path traversal attack surface. This is proactive hardening following a security scan event; no actual data exposure was found. * **Fixed Web Chat Login Authentication Setting Being Accidentally Disabled**: Fixed an issue where a Web Chat conversation platform's login authentication setting was accidentally turned off when saving other settings (such as enabling the data collection form). After the fix, only an explicit action to disable login authentication removes the setting; ordinary settings saves no longer affect the login requirement. ### πŸ› οΈ Bug Fixes * **Fixed SSO Users Unable to Enter Conversations After Logging In to Web Chat**: Fixed an issue where Google SSO, Keycloak, and SAML users, after logging in to a login-required Web Chat, were incorrectly denied entry to conversations (shown as lacking chat permission). After the fix, members auto-joined via SSO can enter conversations normally and receive real-time replies. * **Fixed New Members Being Bombarded with Historical Announcements**: Fixed an issue where newly registered users or members newly joining an organization were bombarded with historical announcements upon entering the platform: announcements published before they joined are now automatically marked as read, no longer popping up banners, dialogs, or unread badges; historical announcements can still be browsed in the announcement list. * **Fixed Linked Tool Step Expansion and Mispaired Results**: Fixed an issue in internal Q\&A tool progress cards where clicking to expand a single tool step expanded all steps in tandem, and execution results could appear under the wrong step or duplicate steps appeared (more common on certain models such as Gemini). After the fix, each step expands independently and results map correctly. * **Fixed "Select All" Under Knowledge Base Search Filters Selecting the Entire Base**: Fixed an issue in the conversation knowledge base selection panel where pressing "Select All" after filtering by search actually selected the entire knowledge base. After the fix, "Select All" in a search context selects only the documents matching the search, the selection count matches what is displayed, and submitted conversations retrieve only the selected documents. * **Fixed Knowledge Base Upload Queue Not Disappearing**: Fixed an issue where, when bulk-uploading many files to a knowledge base (especially near the 25-file limit), the upload queue panel lingered at the top of the table even after everything showed complete. When some files in a batch fail, the remaining successful files now finish uploading normally instead of the whole batch getting stuck. * **Fixed Renamed Pinned Conversations Not Updating Immediately**: Fixed an issue in MaiGPT where, after renaming a "pinned" conversation, the pinned section of the sidebar still showed the old name until the page was manually refreshed. After the fix, renaming behaves consistently for pinned and unpinned conversations and takes effect immediately. * **Fixed Overlapping Wait Indicators During Streaming Replies**: Fixed two display issues during streaming replies in internal Q\&A / test conversations: the typing-dots bubble lingered on screen even after the AI had started producing output, and a "Thinking…" hint appeared at the bottom while content was still streaming. After the fix, only one progress indicator shows at a time, and the messaging is no longer contradictory. * **Fixed Custom API Tools Mislabeled as Web Search**: Fixed an issue where custom API tools in conversations (those with search / fetch in their names) were mislabeled as "Searched the web" or "Read a web page." After the fix, only the platform's built-in web search/read tools show semantic titles; custom tools always show "Used {tool name}." * **Fixed the Browser Operation Panel Disappearing After a Page Refresh**: Fixed an issue where the AI browser operation tool's panel no longer appeared after refreshing the page: when loading a conversation, if the backend still has an active browser session, the panel is now automatically restored. Also fixed the panel showing a leftover frame from the previous conversation when switching conversations. * **Fixed Inconsistent Streaming Retry Defaults for New AI Assistants**: Fixed an issue where the "streaming auto-retry" setting was overridden to off by the form default when creating a new AI assistant. After the fix, this option defaults to on for newly created AI assistants, automatically retrying when reply streaming is interrupted; existing assistants' explicit settings are unaffected. * **Fixed Share Links Not Opening the Conversation**: Fixed an issue where Web Chat conversation share links, when the AI assistant did not have "show previous conversations" enabled, opened only a brand-new blank conversation. After the fix, as long as sharing is enabled, share links always correctly load the shared conversation content. * **Fixed Audio Files Failing to Parse**: Fixed an issue where audio files always failed to parse in cloud storage environments (reporting the audio file could not be found); knowledge base audio import and conversation attachment audio parsing both work again. * **Fixed Complete Failure of the Online Parsing Service**: Fixed an issue where all files immediately failed to parse when knowledge bases used the LlamaParse / MaiAgent Parser (Online) parser (caused by the upstream service retiring an old parsing mode). The service has been switched to a still-supported parsing mode with comparable table and image extraction quality, and file parsing works again. * **Fixed a Single Blocked Page Failing an Entire Document in Vision Parsing**: Fixed an issue with the vision parser where a single page blocked by the model-side content filtering mechanism caused the entire document's parsing to fail. After the fix, blocked pages are marked as skipped while the remaining pages parse normally, and skipped pages are logged for traceability. * **Fixed Indexing Failures for Large Image-Rich Documents**: Fixed an issue where large PDF documents with many images (roughly 20MB and above) failed during knowledge base indexing. Image indexing requests are now automatically batched by size, so large image-rich documents index successfully. * **Fixed Special Image Formats Breaking an Entire Conversation**: Fixed an issue where pasting an image format the model does not support (such as a website favicon ICO file mis-saved as PNG) into a conversation left every subsequent turn of that conversation unable to reply. After the fix, unsupported images are safely skipped before entering the conversation, and previously affected conversations are automatically repaired on load, restoring normal replies. * **Fixed One Faulty Tool Making an Entire MCP Connector's Tools Disappear**: Fixed an issue with external MCP connectors (such as HubSpot's official MCP service) where a single tool with an unparseable parameter format made all of the connector's tools silently disappear. After the fix, only that tool is skipped, and the remaining tools are provided to the AI assistant normally. * **Fixed GitHub-Authorized MCP Connectors Unable to Complete Verification**: Fixed an issue where MCP connectors using GitHub as the authorization source (such as GitHub's official remote MCP service) could not complete OAuth authorization. After the fix, the GitHub remote MCP service can be added as a connector and authorized normally; authorization behavior for other existing services is unaffected. * **Fixed the Code Execution Tool Becoming Unusable from Resource Hoarding**: Fixed an issue where the code execution tool did not release compute resources when execution was interrupted (such as by a user aborting a reply or a connection drop), and the accumulation eventually caused "resources full" errors in other conversations. The system now reliably releases resources in any interruption scenario and automatically reclaims leftover resources older than 30 minutes. * **Fixed IQ Knowledge Reply Failures with Attachments**: Fixed an issue where the IQ Knowledge built-in workflow failed the entire reply when audio or document attachments were included. After the fix, attachments are always read only after background parsing completes, so replies no longer fail because an attachment is still being parsed. * **Fixed Inconsistent Error Rate Statistics on the Dashboard**: Fixed an issue on the monitoring dashboard where conversations actively intercepted by Hook rules were miscounted as system errors: such records are now classified as a separate "Hook Intercepted" type and excluded from error rate calculation (consistent with the handling of user-initiated aborts). The error total and per-type sums are consistent again, and historical records have been corrected as well. ## v2026.07.06 *** ### ✨ New Features * **MaiGPT Thinking Effort Selection and Model Capability Indicators**: When chatting with MaiGPT, you can now choose the AI's thinking effort (low, medium, or high) to adjust reasoning depth based on task complexity. When switching models, the interface instantly shows whether the model supports thinking mode, making it easy to choose the most suitable model. * **MaiGPT Knowledge Base Panel Search Improvement**: The MaiGPT knowledge base side panel now uses a two-tier search design of "unified document search + knowledge base scope multi-select filter." The first row is a multi-select filter for knowledge base scope, allowing multiple knowledge bases to be checked at once; the second row is a unified document search box. This replaces the previous tab-based search target design and fixes the checkbox state contamination issue when switching across knowledge bases. * **Hook Management Interface Optimization and Web Chat Message Hook Indicators**: The Hook management console list page has been fully optimized, adding ID and creator columns, displaying descriptions independently, and turning associated AI assistants into clickable links. The Hook creation flow now starts with template selection before proceeding to the settings page to complete the full configuration, preventing incomplete settings from being submitted. Web Chat messages now include Hook processing indicators β€” a shield icon means the message was modified, a red prohibition icon means the message was blocked, and administrators can customize the indicator text. ### 😊 User Experience & Interface Optimization * **Knowledge Base Web Crawl Source Information Enhancement**: The knowledge base now displays system metadata fields in the UI, allowing administrators to see more complete source information and crawl status when viewing crawled sources. ### πŸ” Security & Governance Enhancements * **SIP Sync Error Sensitive Information Protection**: Raw sync error messages that could contain sensitive information have been removed from the SIP settings and SIP Trunk list, replaced with a sync status badge showing the "Sync Failed" state, preventing leakage of sensitive information such as tokens and hostnames. ### πŸ› οΈ Bug Fixes * **Fixed Conversation Request Failures in Specific Scenarios**: When a system log write task was scheduled to a terminated temporary process, all subsequent conversation requests would continuously fail until the worker restarted. After the fix, log flushing only handles tasks belonging to the current process, and conversations are no longer interrupted by cross-process task conflicts. * **Fixed Browser Tool Session Interruption**: When using the AI assistant's browser tool, leaving the page and returning would lose the browser session state, requiring operations to start over. After the fix, browser session state is preserved across processes, and returning to the page allows continuing previous browser operations. * **Fixed LDAP OU User Counts and Group Mapping**: When using LDAP/AD integration, the user count for all OUs (organizational units) on the scope settings page always displayed as 0, and group mappings identified in DN format could not be matched correctly. After the fix, both OU user counts and group mappings display correctly. * **Fixed Deep Research Progress Bar Display Anomaly**: When using Deep Research in the admin console chat, the progress panel always displayed 0% with no phase description or step details. After the fix, the progress percentage, current phase, and step information all display correctly. * **Fixed Chat Expand/Collapse Animation Jumping**: In Web Chat, expanding or collapsing sections such as thinking processes, tool cards, and research progress caused the screen to jump momentarily and push messages above. After the fix, expand/collapse animations run smoothly without screen jumping or text jitter. * **Fixed Unresponsive Checkboxes in Skill and Agent UI Selectors**: In the "Select Skills" and "Agent UI" selection popups, clicking the checkbox icon itself did not trigger selection β€” only clicking the card border worked. After the fix, clicking the checkbox correctly triggers selection. * **Fixed Missing Agent Message Bubble Background Styling**: In the admin console chat, the background color and frame of Agent message bubbles disappeared, making it visually impossible to distinguish Agent replies from other messages. After the fix, Agent message bubbles have their normal gradient background restored. * **Fixed Admin Console Form Field Selection Anomaly**: In admin console edit forms, selecting a dropdown field would trigger a page reload, causing unsaved edits to be lost. After the fix, filter events only apply to the filter panel and no longer affect field operations in edit forms. * **Fixed Agent Historical Conversation Image Handling Anomaly**: The handling logic for exceeding the image count limit in Agent historical conversations was faulty. After the fix, the image count limit works correctly, and excess images are handled properly. ## v2026.07.03 *** ### ✨ New Features * **Conversation Navigator**: Quickly locate specific messages in long conversations. A new thumbnail navigation bar on the side marks each message with a colored dot β€” click to jump to it. The expanded panel previews message content, ideal for deep conversation scenarios that require reviewing context. * **SIP Call Line Management**: Enterprises can now manage SIP Trunk and phone number settings themselves, with support for organization-level call record queries. The new architecture replaces the legacy call center and provides advanced call management features such as number routing and inbound call dispatching. * **Meeting Summary Voice Playback**: Meeting summaries now support voice playback, automatically detecting the summary language and selecting the corresponding voice. Ideal for scenarios where you need to listen to meeting highlights, or international meetings requiring multilingual playback. * **Offline Meeting Recording Import**: Supports uploading external recording files, automatically performing speech recognition and speaker identification to generate transcripts and meeting summaries. Ideal for non-real-time meeting scenarios, such as importing phone recordings or external meeting recordings. * **Hook Knowledge Base Search Capability**: Hook flows now include knowledge base search capability, allowing custom flows to query knowledge base content directly for message enhancement or conditional logic. * **Web Chat Cross-Device Conversation History**: Web Chat users with a contact identity can now see their past conversation records across different devices or browsers, no longer losing conversation history when switching devices. * **Document Processing Skill Expansion**: AI assistants now include Word, Excel, and PDF document processing skills, enabling document generation and editing directly in conversations. Ideal for enterprise scenarios requiring automated production of reports, contracts, or presentations. * **Claude Sonnet 5 Model Support**: The platform now supports the Claude Sonnet 5 model on AWS Bedrock, allowing users to select the latest Claude model in AI assistant settings. ### 😊 User Experience & Interface Optimization * **Meeting Record Detail Page Full Redesign**: The meeting record detail page UI has been fully redesigned, featuring a slim identity bar, a recording information card on the left, and a three-tab panel on the right. The tab order has been adjusted to "Transcript β†’ Meeting Summary β†’ AI Q\&A," with the transcript as the default tab. When no transcript has been generated, the summary tab clearly indicates that a transcript must be generated first. * **One-Click Meeting Summary Copy**: The meeting summary tab now includes a "Copy Summary" button that copies the raw content of the current summary version with one click (preserving Markdown formatting), eliminating manual drag-selection. * **AI Reasoning Process and Tool Call Interface Redesign**: The AI reply's reasoning process now uses an expandable/collapsible panel design with a more refined visual style. Tool call processes are displayed more clearly, the screen no longer shakes when toggling expand/collapse, and timestamps are consistently aligned. * **Web Chat Loading Animation Upgrade**: The loading screens in Web Chat when entering the page, starting a new conversation, or switching conversations have been changed from a spinner to a message bubble shimmer skeleton animation, delivering a smoother and more unified visual experience. * **Account Settings Page Redesign**: The account settings page is now divided into "Organization" and "Personal" levels. The organization level includes total credit usage and an API key management entry point, while the personal level includes personal information and preference settings. Redundant dropdown menus have been removed for a more intuitive information architecture. * **Real-Time Streaming of AI Interactive Content**: Interactive cards and quick reply options generated by AI assistants are now streamed in real time, so users no longer have to wait for the AI's full reply before seeing interactive elements. * **Crawler Import Record Lookup**: The crawler detail page now includes an "Import Records" tab where you can view the import history and results of each crawl operation. ### πŸ› οΈ Bug Fixes * **Fixed Team Canvas Thumbnail Positioning Anomaly**: The team canvas MiniMap thumbnail occasionally jumped to the top or became misplaced, requiring navigating back and re-entering to recover. After the fix, a viewport stabilization mechanism ensures stable initial loading, and the cross-page style leakage issue has been resolved. * **Fixed Embedded Web Chat Flashing Home Page on Launch**: When embedding Web Chat via iframe or SDK, the full home page would briefly flash before entering the conversation interface. After the fix, home page content is no longer shown in embedded mode. * **Fixed Web Chat Input Box Layout Issue**: When attachment upload, voice input, and voice conversation were all disabled in Web Chat, the input box text alignment was broken. After the fix, the input box aligns correctly under any feature combination. * **Fixed Attachment-Only Messages Not Displaying Bubbles**: When sending only attachments without text, the message bubble did not display at all. After the fix, attachment messages correctly display their bubble frame. * **Fixed Conversation List Scrolling and Recent Assistant Icons**: Fixed abnormal scrolling behavior during AI streaming replies (users being force-scrolled while reading), older conversations not auto-loading on large screens, and recently used AI assistant icons not displaying correctly. * **Fixed Web Chat Language Setting Not Persisting**: After users manually switched the language in Web Chat, opening a new tab or reloading the page reset the language to the default. After the fix, the user's chosen language persists across tabs and page reloads. * **Fixed MCP Connector Salesforce Connection Issue**: The MCP OAuth authentication flow had incorrectly assembled authorization URLs and could not correctly retrieve resource metadata when connecting to services such as Salesforce. After the fix, Salesforce MCP services connect normally. * **Fixed Bedrock Model Tool Call Validation Error**: When using AWS Bedrock models, streaming disconnection retries occasionally triggered an "unpaired tool call block" validation error (affecting approximately 0.04% of messages). After the fix, the retry mechanism correctly handles the pairing of tool call blocks. * **Fixed Vertex AI Model Long Conversation Error**: Long conversations with Google Vertex AI models produced calculation errors because the context window parameter was not passed correctly. After the fix, long conversations proceed normally. * **Fixed New User First-Day Notification Not Sent**: The first-day welcome notification after new user registration was not triggered. After the fix, first-day notifications are sent normally. * **Fixed Incomplete Resource Permission Response**: The user permissions API did not return the complete set of resource creation permissions (such as create tool, create Skill), causing inaccurate permission checks for some features. After the fix, all 7 resource permission fields are returned correctly. * **Fixed AI Gateway Integration Example Page Not Scrolling**: When the window height was insufficient, the AI Gateway integration example page content was clipped and could not be scrolled down. After the fix, the page scrolls normally to browse the full content. --- # 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/07.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.