Alibaba Cloud Model Studio:Workflow application

Start/End

• When to use

  • When you design a workflow, you must define the structure and content of the input and output parameters in the Start and End nodes.

• How to use

  • Start node

    Component	Description
    Variable name	The name of the input parameter. Subsequent nodes can call the variable by its name.
    Type	The data type of the input parameter. Currently, only String, Boolean, and Number are supported. To pass an Array or Object, you must parse it yourself before it enters the workflow.
    Description	The description of the input parameter, used to explain the purpose of the parameter.

    Note

    In addition to custom variables, the application has several built-in utility variables:

    • query

      The user query variable. It is used to receive the prompt variable from API calls or text sent by users in the test window.

    • historyList (only available in conversational workflows)

      This is the conversation history list. The application uses it to automatically maintain conversation history records and provide a multi-round conversation experience.

    • imageList (only available in conversational workflows)

      This is the image list. It stores images that users upload in the test window. Currently, only a single image can be stored.

  • End node

    Component	Description
    Output mode	The format of the output parameters. Text Output and JSON Output are supported.
    Text Box (in text output mode)	You can arrange the paragraph format of the output content. This field supports text input and variable references and is suitable for unstructured content.
    Variable name/Variable value (only effective in JSON mode)	You must arrange the output content in JSON format. You can customize variable names. This field supports text input and variable references and is suitable for structured content.
    Return result	This parameter is effective only when the application is called using an API. It determines whether to output the node content. For more information about the purpose of this component, see Control node content using Return Result.

Knowledge Base

• Why use it

  This node helps you extract the most relevant content from a large information source, such as a repository of documents, FAQs, and product information. This prepares the data for subsequent AI Q&A.

• Functions and features and usage

  • You can search one or more knowledge bases to retrieve relevant document segments. This allows the AI to access more context and provide more professional or accurate answers. This node is often used with LLM nodes.

  • It is especially useful for building conversational workflows and intelligent customer service systems.

• Parameter configuration

  Parameter name	Parameter description
  Input	content: You can directly enter text or reference variables output from previous nodes.
  	imageList: Supports image search. You can directly enter image links or reference variables output from previous nodes.
  Select knowledge base	Specify the knowledge bases to search in. Multiple selections are supported.
  Output	Name the result variable of this node, such as kbResult.

• Sample output structure:

  {
    "rewriteQuery": "...",
    "chunkList": [
      {
        "score": 0.36,
        "documentName": "Document Name",
        "title": "Document Title",
        "content": "Related Content Segment",
        "imagesUrl": "Image URL"
      }
    ]
  }

  The chunkList is the most commonly used part of the output. It contains document segments and their similarity scores.

LLM

• Why use it

  This is the "intelligent brain" of the entire workflow. It can understand language, generate text, analyze images, and participate in multi-round conversations. You can use it to write copy, summarize text, or even analyze image content if you use a VL series model.

• Functions and features

  • It supports processing a single input at a time or processing a large amount of data in batches.

  • You can configure different LLMs, such as Qwen-Plus. You can select the appropriate model based on its performance, speed, or other features.

• Node parameter configuration

  Parameter name	Parameter description
  Mode selection	Single processing mode: This mode performs a fast search using a lower search ratio and not using query rewriting.
  	Batch processing mode: In batch processing mode, the node runs multiple times. Each time it runs, an item from the list is sequentially assigned to the batch processing variable. This process continues until all items in the list have been processed or the specified maximum number of batch processing times is reached. Batch processing configuration: Maximum number of batch processing times (range: 1-100, default for regular user: 100): The upper limit for batch processing runs. Note The actual number of batch processing times is the minimum length of the user-input arrays. If no input variable is specified, the number of times is determined by the number of batches in the configuration. Number of parallel runs (range: 1-10): The concurrency limit for batch processing. If you set this parameter to 1, all tasks are executed sequentially.
  Model configuration	Select a suitable LLM and adjust the model parameters. For more information about the supported models, see Supported models. When you select a VL model: Model input parameters: vlImageUrl can reference parameters or input image links. Image source: You can select Image Set or Video Frame. Image Set: The model considers the uploaded images to be independent and matches the corresponding images for understanding based on the question. Video Frame: The model considers the uploaded images to be from the same video and understands the images as a whole in sequence. At least four video frames are required.
  Parameter configuration	Temperature coefficient: Used to adjust the diversity of the generated content. A higher temperature value increases the randomness of the generated text, producing more unique outputs. A lower temperature value makes the generated content more conservative and consistent. DeepSeek R1 do not support this configuration.
  	Maximum reply length: Limits the maximum length of the text generated by the model, excluding the prompt. This limit varies by model type, and the specific maximum value may differ.
  System Prompt	You can use this parameter to define the model's role, task, and output format. For example: "You are a math expert who specializes in solving math problems. Output the problem-solving process and results in the specified format."
  User Prompt	Configure the prompt template. Variable insertion is supported. The LLM processes and generates content based on the prompt configuration.
  Output	The variable name for the processing result of this node. Subsequent nodes use this variable to identify and process the result. DeepSeek R1 support outputting the deep thinking process (reasoningContent).

  Note

  To integrate the application into your business using an API, see Application call.

• LLM node in a conversational workflow

  • Unlike the LLM node in a task-based workflow, this node supports multi-round conversation configuration. The model can use selected historical conversation information as input.

    Multi-round conversation configuration: The conversational application collects the variables required in the "Context" from the previous conversation rounds and passes them as input parameters to the LLM.

    

  • Context: Declares the context input that the LLM requires. The default ${System Variable.historyList} represents the application input and output from previous conversation rounds. Other parameters refer to application parameters from previous conversation rounds.

    

• Text model node example

  On the Test interface, enter chip engineer in the query parameter:

  

  End node output:

  

• Image model node example

  The LLM supports single or multiple images as input in both URL and base64 formats.

  Note

  A single image can be passed directly. For example: https://****.com/****.jpg.

  Multiple images can be passed as a list. For example: ["URL","URL","URL"].

  On the Test interface, enter https://****.com/****.jpg in the query parameter.

  

  End node output:

  

API

• Definition

  You can call custom API services using POST, GET, PUT, PATCH, and DELETE methods and output the API call results.

  Call method	Purpose
  POST	Used to submit data to the server to create a new resource.
  GET	Used to retrieve a representation of a resource without modifying data on the server.
  PUT	Used to update the representation of a specified resource on the server or create a new resource on the server.
  PATCH	Used to partially update a resource on the server.
  DELETE	Used to delete a specified resource from the server.

• Parameter configuration

  Parameter name	Parameter description
  API address	Enter the API address to be called. You can select POST, GET, PUT, PATCH, or DELETE.
  Header settings	Set the Header parameters by setting KEY and VALUE.
  Param settings	Set the Param parameters by setting KEY and VALUE.
  Body settings	You can select none, form-data, raw, or JSON.
  Output	The variable name for the processing result of this node, used for subsequent nodes to identify and process the result of this node.

  Note

  To integrate the application into your business using an API, see Application call.

• Node example

  Use the POST method to call an API.

  

Intent Classification

• Definition

  Intelligently classifies and matches user input based on intent descriptions and selects a branch to execute.

• Parameter configuration

  Parameter name	Parameter description
  Input	The variable to be processed by this node, used to identify the content to be processed. You can reference variables from ancestor or Start nodes or directly enter variable values.
  Model configuration	Model selection: Qwen-Plus.
  Intent configuration	Configure different intents and enter intent descriptions. The model matches subsequent links based on the intent descriptions, such as "For calculating math problems" and "For Q&A about weather-related knowledge".
  Other intents	If no intent is matched, this link is matched.
  Intent mode	Single-selection mode: The LLM selects the most appropriate intent from the existing intent configurations as the output. Multi-selection mode: The LLM selects all matching intents from the existing intent configurations as the output.
  Thinking mode	Fast mode: This mode can avoid outputting complex inference processes, thereby improving processing speed. It is suitable for simple scenarios. Effect mode: This mode thinks step by step to more accurately match the corresponding classification.
  Advanced configuration	The advanced configuration content is provided to the model as an additional prompt. Here, you can enter more conditions or provide more examples to make the model's classification results more in line with your requirements. Example Suppose you are developing a customer service system for an e-commerce platform, and users may ask various questions about order inquiries, returns and exchanges, and payments. To ensure that the model classifies accurately, you can add relevant prompts and examples in the advanced configuration. Please classify the intent based on the following examples: Case 1: User enters "I want to return the coat I just bought", classified as "Return/Exchange". Case 2: User enters "Please help me check the shipping status of my order", classified as "Order Inquiry". Condition: Only process queries related to orders, and ignore payment and technical issues. Effect: User input: "When can the book I ordered on your website last week be delivered to my home?" Classification result: "Order Inquiry" In this instance, the advanced configuration guides the model to classify "query delivery time" as the "Order Inquiry" intent by providing specific classification cases, while also limiting the classification scope and excluding other irrelevant questions.
  Context	After the context capability is enabled, the system automatically records historical conversation information in Message format. When the model is called, the context is passed in, and the model generates content based on the context. This configuration item is only available in the Intent Classification node of conversational workflows. Note If you enable context, the variable type you pass to this node must be List.
  Output	The variable name for the processing result of this node, used for subsequent nodes to identify and process the result of this node.

  Note

  • This node supports context in conversational workflows.

  • Running this node consumes tokens, and the consumption amount is displayed at runtime.

Text Conversion

• Definition

  This node is used for text content conversion and processing, such as extracting specific content and converting formats. It supports a template mode.

• Parameter configuration

  Parameter name	Parameter description
  Output mode	Text Output and JSON Output are supported.
  Input	Use an LLM to specify a processing method to convert the content to be processed into a specific format for subsequent nodes. You can reference the processing results of ancestor nodes through variable configuration. Text Output: Enter / to insert a variable. JSON Output: Variable Name | Reference/Input | Variable.

• Node example

  The following is a simple example of a Text Conversion node. The workflow logic is as follows: First, the user enters a keyword. Then, this keyword is passed to the Text Conversion node, where it is processed internally to generate a corresponding output reply. Finally, the reply is output through the End node, completing the entire process.

  On the Test interface, enter mathematics in the query parameter:

  

  End node output:

  

Script Conversion

• Definition

  This node processes input content using script code and converts it into a template or a specific output format. This process includes parsing, converting, and formatting the input data to ensure consistency and readability.

• Node example

  This is a Python script conversion example: The city and date variables are passed in from an ancestor node and stored in the params key-value pair. After conversion, a JSON object is output, containing result, result.key0, and result.key1. The JSON Schema of the code's return value must be consistent with the node's defined output.

  Note

  What is a JSON Schema?

  A JSON Schema is a data structure specification. It allows other nodes in the platform to clearly understand which fields (such as result and key1) and data types this node will output. This makes it easier for you to reference them in descendant nodes.

  

• Parameter configuration

  Parameter name	Parameter description
  Input	Defines the input data for this node. You can provide input in two ways: Static value (Input): Directly fill in a fixed value in the left-side input panel, such as Beijing. Dynamic variable (Reference): Reference the output of an ancestor node. For example, if the previous node is named node_a and its output has a field named city_name, you can select node_a.output.city_name to reference it.
  Output	The result produced by the node's code logic. The dictionary returned by the return statement in the code will be the output of this node. For example, if {'result': 'Processing successful'} is returned, descendant nodes can obtain the string "Processing successful" using this_node_name.result.
  Code	Write the core logic code. Obtain input: Use the built-in params object to obtain input parameters. Return output: The processing function main must return a dictionary/object, whose key-value pairs will constitute the node's output.

Conditional Judgement

• Definition

  This node sets conditional branches. When a variable meets a condition, the flow proceeds to the corresponding subsequent branch. It supports AND/OR condition configurations. Multiple conditions are executed sequentially from top to bottom.

• Parameter configuration

  Parameter name	Parameter description
  Conditional branch	Enter the conditional judgement statement.
  Other	Outputs that do not require conditional judgement can be output from here.

• Node example

  The following is a simple example of a Conditional Judgement node. The workflow logic is as follows: The user first enters two parameters, which are then passed to the Conditional Judgement node. The node performs a conditional judgement on the parameters and then generates an output reply through different branches of the Text Conversion node. Finally, the End node outputs the generated reply.

  On the Test interface, enter 12345 in the scert parameter and admin in the admin parameter:

  

  End node output:

  

Function Compute

• Definition

  This node authorizes the Alibaba Cloud Function Compute service and calls custom services in Function Compute.

• Parameter configuration

  Parameter name	Parameter description
  Input	The variable to be processed by this node, used to identify the content to be processed. You can reference variables from ancestor or Start nodes or directly enter variable values.
  Region	Select a region: Singapore, Kuala Lumpur, Jakarta.
  Service configuration	Select the service configuration.
  Output	The variable name for the processing result of this node, used for subsequent nodes to identify and process the result of this node.

Plugin

• Definition

  You can configure plug-in nodes in your workflow application to expand its capabilities and perform more complex tasks. Alibaba Cloud Model Studio provides a series of official plug-ins, such as Quark Search, Calculator, and Python Code Interpreter. You can also create custom plug-ins based on specific requirements.

  For more information, see Plugin overview.

Workflow application related

Node related