Agent platform updates
Feature updates, improvements, bug fixes, and billing changes across Agent modules, ordered from newest to oldest.
Evolution: Multimodal evaluation launched
New Features
- Evaluation — Multimodal evaluation launched. Add evaluation datasets containing multimodal data, select multimodal graders, and launch evaluation tasks against multimodal data
Fixed
- Fixed an issue where trajectory evaluation tasks could not select trajectory data from application output
Evolution: Trajectory evaluation and evaluation interaction improvements
New Features
- Evaluation — Trajectory evaluation launched. Create evaluation datasets containing trajectory data, select trajectory graders, and launch evaluation tasks against trajectories
- Evaluation — Interaction improvements for dataset creation, dataset data addition, and preset grader filtering
Managed Agent: Feature updates
New Features
-
New built-in tool
todo_write: the agent maintains a structured todo list to break down multi-step tasks and track the status of each step. See Agent Tool Configuration. -
New optional models:
deepseek-v4.1-flash,deepseek-v4.1-flash-1017, andglm-5.3. See Define an Agent.
Managed Agent: Feature updates
New Features
- Session-level agent overrides and MCP request headers: when creating a session, use the
agentobject to override the system prompt, tools, MCP servers, and skills, and usemcp_configsto inject custom headers for an MCP server. See Create a Session.
Managed Agent: Feature updates
New Features
-
Environment warm-up: environments with preinstalled packages may need warm-up after creation; a preparing environment cannot be referenced by a session. See Cloud-hosted Environment.
-
Smart modes:
model.idnow supports four routing tiers:auto,performance,balanced, andeconomy. See Define an Agent.
Managed Agent: Feature updates
New Features
- The per-file upload limit is increased to 50 MB. See File Upload and Mounting.
Managed Agent: Feature updates
New Features
- Scheduled tasks (formerly Deployments) are now available: bind an agent and trigger it on a cron schedule or manually; each run creates a session. See Scheduled Tasks.
Managed Agent: New web search and page fetching tools
New Features
- Two new built-in tools:
web_search, which searches the internet and returns the title, link, and content summary of relevant pages (CNY 0.03 per call), andweb_fetch, which reads the main content of a given web page and converts it to text (free of charge for a limited time). See Agent Tool Configuration.
Evolution: Prompt optimization launched
New Features
- Optimization — Prompt optimization launched, supporting version-comparison optimization and debugging-feedback optimization
Security: Security module launched
New Features
- The Security module is now available, providing built-in, out-of-the-box native security capabilities: Agents get default protection as soon as they go live, with no extra integration or rework.
- Provides protection overview, Agent assets, security policy, and risk & audit capabilities, forming an inventory–intercept–audit security loop.
- Advanced protection provides 7 security policies and risk monitoring capabilities that can be enabled on demand, and is currently free for a limited time.
See Security user guide.
Managed Agent: Memory Store launched
New Features
- Memory Store is now available: a persistent mounted resource with a three-layer model of store, memory, and version. Mount it at session creation (read-only or read-write); the agent reads and writes it with file tools, historical versions are recorded automatically, and bodies can be queried and redacted. The console Memory Stores page and the Memory Store API are both available. See Memory Store and the Memory Store API.
Evolution: Alert management launched
New Features
- Alert Management launched
See Alert management.
Managed Agent: Feature updates
New Features
- The session event stream now supports Delta streaming: enable
event_deltas[]formessageandreasoningto stream text chunks as they are generated. See Session Event Stream.
Changes
- The built-in
download_filetool is retired and replaced bymark_artifacts, which registers agent output files. See Agent Tool Configuration.
Managed Agent: Multimodal user messages
New Features
- The Send Event endpoint (
POST /sessions/{session_id}/events) now accepts multimodal content blocks in user messages:image(image_url/file_id),video(video_url/file_id), andfile(file_id+filename). Images and videos enter the model's multimodal vision channel; files are materialized into the session sandbox filesystem and read by the agent on demand. See Send Event.
Managed Agent: Feature updates
New Features
- Sessions now support environment variables: pass
environment_variableswhen creating or updating a session and read them by name at runtime. See Create a Session.
Managed Agent: Credential vault egress injection configuration
New Features
- Credential vault now supports Allowed Hosts and Injection Location configuration. Credentials are referenced as
${VARIABLE_NAME}placeholders by the agent; at egress, the gateway replaces the placeholder with the real secret in the Authorization header for matching hosts only—unmatched hosts receive the placeholder. See Credential Vault Authentication.
Managed Agent: Webhook notifications launched
New Features
- Webhook notifications are now available, letting you push status-change events for resources such as sessions, agents, and scheduled tasks to a designated callback URL. For onboarding, see Webhook notifications; for the API contract, see Webhook API.
Sandbox: Isolated cloud sandbox launched
New Features
- Sandbox launched, providing an isolated cloud sandbox environment for code execution, browser operations, and file handling.
- Three base images available:
code-interpreter-v1,browser, andall-in-one. - Create and manage templates in the console, configuring resource specs, file mounts, network policies, environment variables, and lifecycle.
- E2B SDK/API-compatible, supporting instance lifecycle management and data-plane operations.
See Sandbox quick start.
Managed Agent: Feature updates
New Features
- The Bailian CLI now supports Managed Agent: use
bl managed-agentto manage agents, environments, sessions, and related resources. See Use the CLI.
Managed Agent: Feature updates
New Features
-
Multi-agent collaboration is now available: a coordinator agent orchestrates multiple member agents to complete tasks together. See Multi-Agent Collaboration.
-
Tool approval policies are now available: set each tool to always allow or always ask. See Agent Tool Configuration.
Managed Agent: Commercial billing takes effect
- Managed Agents billing takes effect at 2026-08-17 09:00:00 (UTC+8). For billing items, free quota, and an example, see Billing.
Evolution: Documentation launched
New Features
- Evolution documentation launched
RAG: Knowledge base documentation launched
New features
- RAG knowledge base documentation launched
See RAG quick start.
Managed Agent: Documentation launched
New Features
- Managed Agent documentation is now live
See Managed Agent quick start.
Managed Agent: Feature updates
New Features
- Managed Agent is now available
Agent hosting: Managed agent runtime launched
A new managed agent runtime API is now available, with the platform hosting sessions and tool execution, Learn More
Knowledge base RAG: Knowledge retrieval service launched
A new knowledge retrieval service supports joint retrieval across multiple knowledge bases and hybrid ranking, Learn More
Knowledge base RAG: Knowledge Q&A service launched
A new knowledge Q&A service generates answers based on large models combined with retrieval, Learn More
Skills: Skill capability packs launched
New Skill capability packs let agents use official or custom skills, Learn More
Data connection: Data connection module launched
A new data connection module supports data sources such as MySQL/Yuque/OSS, Learn More
Applications - Tongyi Tingwu Agent: Discounted ASR resource packs
The Tongyi Tingwu Agent now offers discounted ASR resource packs, designed for use cases such as automotive sales and customer profiling. Learn More
Spring AI for Alibaba Cloud: New documentation: Calling Model Studio applications
New documentation shows how to use the Spring AI for Alibaba Cloud framework to call agents and workflows in Model Studio. Learn More
Application components - data connection: New ListCategory API for data connection
The new ListCategory API for data connection lets you query the category list of the source application. Learn more
Tingwu Agent: WebSocket protocol support for industrial command transcription
The Tingwu Agent supports the WebSocket protocol for transcribing industrial commands. Documentation for this interaction protocol is now available. Learn more
Application - data connection (formerly application data): New ChangeParseSetting API for data connection
The new ChangeParseSetting API for data connection lets you configure parsing methods by file type. Learn more
Application invocation API: New DashScope API for agent applications
The new DashScope API reference for agent applications is now available. It supports single-turn and multi-turn conversations, streaming responses, file-based Q&A, and visual understanding. Learn more
Application - multimodal interaction development kit: Launched the server-side Java SDK for the multimodal interaction development kit
The new server-side Java SDK includes documentation covering downloads, installation, key APIs, and code examples. Learn more
Knowledge base RAG: Added logging and monitoring to the knowledge base
All retrieval calls from the knowledge base are now delivered to Log Service, supporting auditing, troubleshooting, statistics, and alerts. Learn more
Application - multimodal interaction development kit: Launched the mobile Android SDK for the multimodal interaction development kit
The new mobile Android SDK includes documentation covering downloads, installation, key APIs, and code examples. Learn more
Multimodal interaction development kit: Added a license mode to the RTOS C SDK
The RTOS C SDK for the multimodal interaction development kit now supports a license mode, covering semi-managed and fully managed access on embedded hardware. Learn more
Application component - prompt engineering: Launched the prompt engineering API for application components
A new prompt engineering module and its API are now available as an application component, allowing you to manage prompt templates, Learn more
Application Publishing and Sharing - UI designer: Launched the UI designer
The Model Studio UI designer is now available. It integrates Mobi low-code capabilities, allowing you to build web applications with a drag-and-drop interface. Learn more
Memory: Memory 2.0 is now available
Introduces long-term memory: automatically extracts memory segments and user profiles from conversations, and supports sharing a single Memory across multiple applications. Learn more
Qwen Web Search Agent: Web Search Agent now available
Available in the Marketplace with a product introduction, API reference, and usage instructions. Learn more
Quanmiao - PPT generation: PPT generation API for Quanmiao now available
The Quanmiao solution now includes a complete API set for PPT generation, with interfaces to manage templates, documents, and generated works, Learn more
Multimodal Interaction Development Kit: New Linux C++ SDK
A new Linux C++ SDK is now available for the Multimodal Interaction Development Kit. The documentation provides download and installation instructions, and key interface examples. Learn more
Knowledge Base: Knowledge bases support subscription-based resource plans
Knowledge bases offer two billing methods: pay-as-you-go and resource plans. You can purchase the RAG Standard Edition Resource Plan or the RAG Ultimate Edition Resource Plan in the console. For more information, see Knowledge Base Billing.
Application Evaluation: New application evaluation
Supports evaluation sets for agents, workflows, and custom applications to build an evaluation system suited to your business needs. For more information, see New Evaluation Sets.
Application - Tongyi Multimodal Translation: New Web Translation JSSDK
The newly available web translation JSSDK lets you make your website multilingual simply by inserting a script. Learn more
Plugins: Image Generation plugin now available on the international site
The official plugin list on the international site now includes text_to_image (Image Generation). Learn more
MCP: New official MCP service
The official MCP service is now live, supporting both in-platform integration and access by third-party services such as Amap Maps. Learn more
Multimodal Interaction Development Kit: New Android/iOS Lite SDKs
The Multimodal Interaction Development Kit now offers new Android and iOS Lite SDKs and a voice list. Learn more
Application - Knowledge base: New ranking model and instruction intervention mode for knowledge base retrieval
The Retrieve interface now includes a ranking model option and an instruction intervention mode, enhancing the platform's retrieval capabilities. Learn more
Long-term Memory: Release of new long-term memory and user persona management APIs
The new long-term memory improves upon the legacy long-term memory in the following ways:
Open APIs let you integrate with any application and support multiple applications sharing the same memory library.
Higher speed and efficiency: lower latency and better recall performance for memory retrieval.
Automatic extraction: automatically extracts key information from conversations and deduplicates it without manual input.
Improved retrieval algorithms: adds semantic search capability, significantly improving retrieval accuracy and response speed.
User persona capabilities: adds full user persona extraction and management.
Knowledge Base: Support for creating audio and video knowledge bases via API
This API supports creating two types of knowledge bases: unstructured knowledge bases based on documents or audio/video files, and structured knowledge bases for data queries or image Q&A. For more information, see CreateIndex – Create a Knowledge Base.
Workflow Application: Add multimodal generation node
The multimodal generation node calls Alibaba Cloud multimodal models to generate images, videos, or audio content based on configured prompts and parameters.
Application - Multimodal Interaction Development Kit: You can now configure the Multimodal Interaction Development Kit
The application configuration page now lets you configure four modules: voice interaction, understanding and generation, skills, and agents. Learn More
Knowledge Base: Add knowledge base monitoring API
Query monitoring data for a specified knowledge base over a specific time range. For more information, see GetIndexMonitor – Get Knowledge Base Monitoring Data.
Knowledge Base: Add update knowledge base API
Update part of the configuration for a specified knowledge base. For more information, see UpdateIndex – Update a Knowledge Base.
Workflow Application: Add asynchronous execution mode
Add asynchronous execution mode for text generation.
In asynchronous execution mode, workflows run in the background. The system returns a task ID immediately. Use the task ID to query the execution result. View historical asynchronous tasks in the Task Hub.
Knowledge Base: Support RAM users to activate knowledge bases and cost allocation management
To allocate costs across departments or projects, use the tag feature to label workspaces for cost allocation. For more information, see Cost Allocation Management.
Knowledge Base: Support adjusting initial retrieval parameters for knowledge bases
Lowering the values of Initial Vector Retrieval TopK and Initial Keyword Retrieval TopK reduces the number of tokens sent to the reranking model, significantly lowering costs. For more information, see Model Invocation Costs.
Knowledge Base: Knowledge base commercial billing
Alibaba Cloud Model Studio knowledge base service starts billing on January 4, 2026. Total charges consist of instance fees and model invocation fees. For more information, see Knowledge Base Billing.
Agent Application: Release of new agent application (Agent 2.0)
The new agent application unifies knowledge bases and MCP as tools. Agents autonomously plan when and in what order to call them. It fully displays the model's reasoning process and tool calling steps.
Knowledge Base: Support for audio and video knowledge bases
Upload audio and video files to create knowledge bases for:
Intelligent search and Q&A based on audio/video content, such as live stream replay Q&A, course tutoring, and customer service quality inspection.
Secondary creation of audio/video assets, such as generating scripts and captions or providing editing suggestions.
For more information, see Create and Use a Knowledge Base.
Workflow Application: Knowledge base nodes support different invocation methods
Knowledge base nodes support three invocation methods: mandatory, intelligent, and legacy. They also support parsing document, image, and table-type knowledge bases.
Workflow Application: Add multimodal data nodes. Add file-type parameter to start node
Add four nodes: document parsing, image parsing, video parsing, and audio parsing. These extract structured parameters from corresponding input files.
Add file-type parameter to the start node to upload files.
Application Invocation: Support invoking Alibaba Cloud Model Studio applications via Responses API
Synchronous invocation API: ideal for real-time interactive scenarios requiring immediate results. Easily reuse existing OpenAI codebases or quickly integrate tools from the OpenAI ecosystem. For more information, see Synchronous Invocation API Reference.
Asynchronous invocation API: for long-running tasks, set background to true in the request. The API returns a task ID immediately for later querying and management. For more information, see Asynchronous Invocation API Reference.
High-code Application: New application type
Deploy AI backend services based on Python project structure. Includes built-in automated O&M, observability, and Simple Log Service. For instructions, see High-code Application.
Workflow Application: Add one-click import for Dify workflows
Import Dify workflows with one click. Nodes and parameters map automatically after import. For instructions, see Workflow Import and Export.
Agent Application: Upgrade file Q&A
Support full-text citation, chunk retrieval, and custom processing modes to deeply analyze and automate tasks for documents, images, audio, and video files. For instructions, see File Q&A.
Agent Application: Upgrade reply functionality
Automatically switch to a better-performing model when the primary model is limited in specific scenarios. For more information, see Agent Application.
Knowledge Base: Simplify data import workflow
Import files or data directly when creating a knowledge base. No need to pre-import them on the application data page.
Knowledge Base: Data sources support DMS
Sync data directly from a specified DMS table when creating a knowledge base. For more information, see Knowledge Base.
Knowledge Base: Optimize creation workflow
Classify knowledge bases into three types—Document, Data, and Image—based on use cases to simplify creation. For more information, see Knowledge Base.
Knowledge Base: Add debugging panel
When editing an agent application, adjust knowledge base parameters online and verify retrieval recall performance in real time. For instructions, see Knowledge Base.
MCP: Add external invocation capability
Integrate into third-party applications: configure automatically with one click for quick external invocation.
Integrate into personal projects: invoke using the MCP SDK for flexible coding and deep customization.
Workflow Application: UI upgrade
Retire agent orchestration applications.
Upgrade the workflow application UI.
Knowledge Base: Knowledge base data sources support self-managed MySQL
Structured knowledge bases support syncing data from a specified self-managed MySQL table. For configuration details, see Integrate MySQL Data into a Knowledge Base.
Knowledge Base: Knowledge bases support text-embedding-v4 model
The text-embedding-v4 model upgrades language support, code snippet vectorization, and vector dimension options compared to v3. It suits most use cases.
Knowledge retrieval and Q&A: Knowledge retrieval and Q&A API launched
New knowledge retrieval and Q&A API, with three stages returned in SSE streaming, Learn more
Knowledge Base: Knowledge bases support text-embedding-v3 model
The text-embedding-v3 model upgrades language support, input length, and customizable vector dimensions compared to v2. It suits most use cases.
Knowledge Base: Choose Qwen VL for image parsing
Select qwen-vl-max or qwen-vl-plus models. Pass prompts to guide recognition of layout, elements, and content. Ideal for complex images or charts. For more information, see Create and Use a Knowledge Base.
Knowledge Base: Unstructured knowledge bases support importing offline HTML files
Build unstructured knowledge bases from offline HTML files. For more information, see Create and Use a Knowledge Base.
Knowledge Base: Add weight setting capability
- When an agent application links to multiple knowledge bases, assign weights by source importance. The system prioritizes retrieving relevant information from higher-weight knowledge bases. For instructions, see Create and Use a Knowledge Base.
MCP: Add MCP marketplace and MCP management
- Enable Alibaba Cloud Model Studio-provided MCP services or deploy custom MCP services. Reference them in agent and workflow applications to enhance capabilities. For instructions, see Official MCP Services and Custom MCP Services.
Agent Application: Agent applications support QwQ series models
Excludes plugins, workflows, and audio/video interaction capabilities.
- The QwQ model has strong reasoning ability. It outputs its reasoning process before the final answer. Its math and coding performance (AIME 24/25, LiveCodeBench) and general metrics (IFEval) match DeepSeek-R1 full version and significantly exceed DeepSeek-R1-Distill.
Agent Application: Agent applications support QwQ series models
Excludes plugins, workflows, and audio/video interaction capabilities.
- The QwQ model has strong reasoning ability. It outputs its reasoning process before the final answer. Its math and coding performance (AIME 24/25, LiveCodeBench) and general metrics (IFEval) match DeepSeek-R1 full version and significantly exceed DeepSeek-R1-Distill.
Workflow Application: Workflow applications support qwq-plus and qwq-32b models in large language model nodes
The QwQ model has strong reasoning ability. It outputs its reasoning process before the final answer. Its math and coding performance (AIME 24/25, LiveCodeBench) and general metrics (IFEval) match DeepSeek-R1 full version and significantly exceed DeepSeek-R1-Distill.
Agent Application: Agent applications add multimodal reply enhancement toggle in retrieval settings
Enable multimodal recognition to let agent applications parse chart and image content in knowledge bases and provide precise answers combining visual information.
Agent Application: Agent applications support qwen-vl-plus-latest and qwen-vl-plus-2025-01-25
qwen-vl-plus-latest always points to the latest snapshot version.
qwen-vl-plus-2025-01-25, also known as qwen-vl-plus-0125, belongs to the Qwen2.5-VL series. It extends context to 128K and significantly enhances image and video understanding.
Workflow Application: Workflow applications support batch nodes
Batch nodes are functional nodes in workflows for handling batch tasks. For more information, see Batch Processing.
Workflow Application: Workflow applications support DeepSeek series models
You can use DeepSeek series models in workflow applications to build task-oriented and conversational workflows.
Agent Application: Agent applications support DeepSeek series models
Integrate DeepSeek series models into agent applications with knowledge bases, long-term memory, and prompt templates to build private knowledge Q&A applications.
Workflow Application: Workflow canvas enhancements
Intent classification nodes support intent modes:
Single-select mode: the large language model selects the best-matching intent from existing configurations.
Multi-select mode: the large language model selects all matching intents from existing configurations.
Text transformation nodes support JSON output format.
For more information, see Workflow Application.
Agent Orchestration Application: Agent orchestration applications support conditional judgment nodes
Agent orchestration applications support conditional judgment nodes. For more information, see Agent Orchestration Application.
Workflow Application: Add audio/video real-time interaction
Easily convert text-and-image chat applications into audio/video real-time interaction applications.
Provides H5/APP debugging windows. Publish to your web/iOS/Android apps using the audio/video SDK.
Agent Application: Add audio/video real-time interaction
Easily convert text-and-image chat applications into audio/video real-time interaction applications.
Provides H5/APP debugging windows. Publish to your web/iOS/Android apps using the audio/video SDK.
Application Components: Add Prompt example optimization documentation
Extensive practice shows that few-shot methods significantly improve large language model reasoning and performance. As Alibaba Cloud Model Studio's few-shot capability, the Prompt example library lets you input user queries (Query) and expected model responses (Answer). During model invocation, it retrieves relevant examples based on the user input to guide output and improve answer accuracy. Ideal for customer service and Q&A scenarios. For more information, see Prompt Example Library.
Application Sharing: Add WeChat and DingTalk sharing channels
Create DingTalk AI bots and WeChat Official Account AI bots. After creation, share the bot webhook address or QR code with target users as needed. For more information, see Application Sharing.
Knowledge Base: Unstructured knowledge bases support importing Excel documents
In addition to structured knowledge bases, unstructured knowledge bases also support importing Excel documents for:
Documents that may include non-Excel formats (e.g., PDF, DOC).
Import Excel documents with different table schemas.
For more information, see Knowledge Base.
Application Observability: Add application observability
The new Application Observability supports end-to-end viewing of Alibaba Cloud Model Studio application processing flows. For more information, see Usage Monitoring and App Performance Analytics.
SFM service-linked role: Add AliyunServiceRoleForSFMTelemetry role
This service-linked role enables Alibaba Cloud Model Studio application observability. Application observability uses this role to access your OpenTelemetry instances. Click SFM Service-linked Role to view.
SFM service-linked role: Add AliyunServiceRoleForSFMAccessingMNS role
This service-linked role enables automatic synchronization of OSS changes. Alibaba Cloud Model Studio data centers use this role to access OSS change messages in your MNS queues. Click Service-linked Role to view.
Knowledge Base: Structured knowledge base data sources support cloud databases
Structured knowledge bases support ApsaraDB RDS as data sources. For details, see Create and Use a Knowledge Base.
Knowledge Base: Unstructured knowledge bases support custom metadata
Add metadata to documents in unstructured knowledge bases to improve retrieval efficiency and precision. For details, see Create and Use a Knowledge Base.
Knowledge Base: Support image and text retrieval
Knowledge bases parse images in documents. Images in structured documents become vectors. Images in unstructured documents get text extracted first, then become vectors. For details, see Create and Use a Knowledge Base.
If a structured knowledge base linked to an agent application contains image indexes, upload images during Q&A. Related data records appear with the question for the large language model. For details, see Agent Application.
The test knowledge base page also supports image and text input to help evaluate image and text retrieval capability. For details, see Create and Use a Knowledge Base.
Data Management: Unstructured data files support tag categorization
Unstructured data supports tag categorization. Click Create and Use a Knowledge Base for details.
Application Center: Add retrieval configuration to agent applications
In agent applications, enable knowledge retrieval augmentation to access retrieval configuration. Set answer scope and whether to show answer sources. Click Create and Use a Knowledge Base for full details.
Data Management: Structured data capability updates
Structured data supports incremental import and configurable column data types. Try it in Data Management.
Application Center: Add custom parameter invocation instructions
Pass custom parameters when invoking workflow and agent orchestration applications. For more information, see Application Invocation.
Application Center: Add custom parameter pass-through examples
Pass custom parameters when invoking workflow and agent orchestration applications. See "Request Parameters (Parameter Pass-through for Workflows and Agent Orchestration Applications)" for details.
Application Center: Add custom parameter passing instructions
Pass custom parameters when invoking workflow and agent orchestration applications. Click Pass Plugin Parameters via API for full details.
Application Center: Add application type
Add Flow Orchestration Application. Try it at Flow Orchestration Application.
Application Marketplace: Add media/retail article style and format learning and e-commerce retail promotion copywriting applications
Try these in the Application Marketplace. Click Media/Retail Article Style and Format Learning and E-commerce Copywriting Intelligent and Controllable Generation for full details.
Application Center: Update application model display
After selecting a model in My Applications, display maximum-length data.
Application Center: Agent application capability updates
Agent applications support multiple function invocations and support interrupting loop invocations.
Application Center: Agent application capability updates
Agent applications support multiple function invocations and support interrupting recursive invocations.
Application Center: Add testing capabilities to application test window
The application test window supports long-term memory user insights and user personas. Click My Applications for full details.
Plugin Management: Add built-in plugin types
Add QR code generation and GitHub search plugins.
Application Center: Create custom applications supporting qwen2 models
- Create custom applications supporting qwen2-72b-instruct and qwen2-57b-a14b-instruct models.
Data Center: Knowledge index updates
- Knowledge indexes support hit testing for recall. Click Knowledge Index for details.
Application Marketplace: Custom applications support sharing
- Published applications support sharing on web (PC/H5).
- API invocation: Uses existing API invocation methods. Applications publish regardless of configuration.
- Web (PC/H5): Built using existing Alibaba Cloud Model Studio pages and Magic Pen templates.
- Official sharing channel: Provides a test web page for your application. Supports only small-scale sharing for experience testing.
- Magic Pen sharing channel: Magic Pen is a low-code development platform offering multiple templates. Choose flexibly, customize user interfaces and custom domains, and easily publish production-grade web applications.
Data Management: OSS authorization method updates
- Document data in Data Management supports local upload and OSS import. After associating an OSS bucket, OSS authorization and cross-domain access are enabled. Click Import Data from OSS for details.
Application Center: Plugin services
- Application Center supports custom parameter pass-through. Click Pass Plugin Parameters via API for details.
Developer Guide: Add permission control to knowledge indexes
- RAM users must be authorized by the Alibaba Cloud account or RAM administrator to use Alibaba Cloud Model Studio knowledge index capabilities. Click Custom Policies for Knowledge Indexes for full details.
Data Management: Support uploading Excel files
- Support structured data import and custom column information setup. Click Data Import for details.
Application Center: Support building knowledge bases and RAG online inference using structured data
- Support building and linking knowledge bases with structured and unstructured data for knowledge retrieval. Click Knowledge Index for details.
Application Marketplace: Application Center
New applications support test versions. Editing an application saves it as a test version by default.
Application Marketplace: Process Management
Process canvases support business parameter pass-through. Click Flow Orchestration (Legacy, Pending Deprecation) for details.
Data Center: Add data augmentation capability
Data augmentation supports enhancing data through parameter and prompt configuration to improve processing results.
Data Center: Data Management supports multiple file formats
Data Management now supports importing and parsing MD, TXT, PPT, and PPTX files, plus previewing pages.
Data Application: Knowledge indexes support chunk management
Manage chunks (enable/disable, edit content).
My Applications: Add multi-model application creation
A new application creation flow based on the assistant API mechanism for RAG, workflows, and plugin scheduling.
Select from max, plus, and turbo models. Configure effects and costs across multiple dimensions.
A new long-term memory mechanism supports long-term memory storage and personalized assistant effects. Click My Applications for details.
Data Management: Add unified platform data management module
A new unified data management capability supports local data uploads and OSS imports, plus bulk import of unstructured documents.
One-click IDP parsing converts data into usable formats. Supports multi-category management.
Knowledge Index: Create and manage knowledge base indexes for RAG applications, referencing the Data Center uniformly
Uses LlamaIndex to provide complete knowledge index configuration, including chunking and vector settings.
Supports multiple recall index configurations to meet RAG effectiveness requirements at different granularities.
Application Tools: Add official plugin capability
Launch a new plugin center with four official plugins: Image Generation, Quark Search, Python Code Interpreter, and Calculator. Click Plugin Center for details.
Application Marketplace: Agent API applications support API invocation
Add assistant API and Python SDK interfaces. Click Best practices for details.
Enterprise Knowledge Base: Add enterprise retrieval pipeline
Add ES vector database. Knowledge base name is Default Knowledge Base. Retrieval performance improved. View details.
Enterprise Knowledge Base: Support uploading FAQ-style files
Add Excel file upload. View details at Configure Enterprise Knowledge Base.
Enterprise Knowledge Base: Add upload formats
Support uploading documents from OSS. View details at Configure Enterprise Knowledge Base.
Prompt Engineering: Add prompt template types
Add 110 prompt templates.
Flow Orchestration: Add flow orchestration version
Flow orchestration supports one-click layout adjustment to optimize canvas interaction experience.
Application Management: Add session record backup
Back up session records using ADB-PG. Support manual enable/disable.
Flow Orchestration: Add flow canvas version
Script nodes provide simple code development for developers. Add Python language support.
Flow Orchestration: Add large language model node to flow canvas
Large language model nodes link to official models, SFT models, and others to leverage large language model capabilities. View details at LLM Node Description.
API-related: Fix missing document upload field
Related fields have been updated. View details at Application Invocation SDK Reference (Legacy).