The workflow provides users with an intuitive visual canvas, supporting orchestration of complex business processes through multiple nodes such as large language model (LLM) node, code node, and parameter extraction node. It aims to achieve stable and controllable business effects, ensuring accuracy and interpretability for each process node.
Main Features
Visual Canvas
Intuitive operation: Users can easily build and manage complex business processes by dragging, dropping and connecting nodes. The canvas supports mouse and touchpad operation and provides multiple shortcut keys to improve workflow configuration efficiency.
Simple UI: The intuitive UI design makes workflow modification simple. The node list, canvas, and node detail are categorized for display, making the page more concise and clear.
Clear Node Design
Information Collection Nodes: These nodes collect information based on multi-round dialogue, webpage UI, and other methods, and transmit the information backward. For example, the parameter collection node collects necessary parameter information by performing multi-round dialogue with users.
Information Processing Nodes: These nodes only process information and no longer collect it. For example, the LLM node leverages LLM capabilities to process information by editing prompt.
Variable Processing Nodes: These nodes support operations such as converting, assigning value, and aggregating variables. For example, JSON serialization, array sorting, and variable aggregation.
Basic Nodes: These nodes form the foundation of workflow operation, implementing functions such as workflow result output and logical judgments. They include nodes like reply node and conditional judgment node.
Flexible Interactive Dialogue
Multi-round Dialogue: The workflow operating mechanism supports interactive multi-round dialogue with users. During information collection, the system performs self-service analysis to check whether required fields are complete. If incomplete, it automatically generates counter-questions to guide users in providing required information.
Flexible Redirection: Built-in Agent monitoring node redirection globally, no need to distinguish between "dialogue flow" and "workflow", enabling both flexible dialogue and precise execution process. The built-in Agent has the following capabilities:
Responsible for Dialogue Takeover: ALL workflow-related dialogues transit this Agent for unified processing.
Control Node Redirection: The Agent intelligently navigates to the corresponding node according to the user conversation content.
Identify Global Intent: Globally recognize exit intents like "do not continue executing workflow" and small talk intents like "OK" or "no problem", then intelligently reply to user.
Optimized Supportive Features
Global Debug: The workflow supports multi-round dialogue debug, showing the handling path and detailed information of each round, suitable for testing after workflow orchestration completion.
Node Debug:
Support end-to-end testing for workflows, suitable for fully debugging workflows during the testing stage. In the canvas, dynamically display the running process, tokens and time consumed by node operation, node input and output details, etc.
Some workflow nodes support single-point debugging, suitable for small-scale testing during workflow development, including: Parameter Extraction Node, Option Node, File Collection Node, Text Collection Node, LLM Node, LLM Intent Recognizer Node, LLM Knowledge Q&A Node, LLM Label Extractior Node, Knowledge Retrieval Node, Plugin Node, Tool Node, Code Node, Message Queue Node and Workflow Node.
Quick feedback: provide users with one-click issue feedback feature, support reporting the problem against nodes of different types.
Key Concepts
Variables
Variables are for hosting the data of workflow circulation. Workflow supports the use of application-level variables and workflow-level variables (see Variable Description for definition). Within the workflow, you can refer to variables from preceding nodes or application-level variables, workflow-level variables through the "Input Variables" of a node.
Node
Nodes are the basic building units in the workflow. Each node represents a specific operation or feature. Nodes connect to form a complete business process. According to functionality and purpose, Tencent Cloud Agent Development Platform (Tencent Cloud ADP) workflow nodes can be divided into the following categories:
1. Information Collection Nodes: Used to collect necessary information from user dialogue or UI interaction, and transmit this information to subsequent nodes. Information collection nodes include:
2. Information Processing Nodes: Used to process collected information and no longer collect new information. Information Processing Nodes include:
3. Variable Processing Nodes: Support operations such as converting, assigning value, and aggregating variables. For example, JSON serialization, array sorting, and variable aggregation. Variable Processing Nodes include:
4. Basic Nodes: The foundation of workflow operation, responsible for entering or exiting workflows, conditional branch judgment, serial/parallel execution, and content output. Basic nodes include:
Connecting Line
A connecting line is a directed line segment that connects two workflow nodes, describes the flow direction of workflow execution, and connects nodes under constraints.
Limitations
Constraints hinder connecting lines between nodes, and under specific circumstances, connection between two nodes is not allowed. Specific constraints include:
At least one end node is required in the workflow for the entire workflow to execute normally.
The input of the current node must connect with the output of the previous node. Inputs cannot connect with each other, nor can outputs.
There is no fixed sequential relationship between information collection nodes and information processing nodes.
The output of information collection nodes, information processing nodes, and start nodes can be connected to multiple subsequent nodes, supporting parallel processing.
Connections between multi-nodes must comply with the following validation logic:
loop verification
Definition: Circular dependency between nodes is not allowed in the workflow, meaning no loops can occur.
Example: Node A -> Node B -> Node C -> Node D -> Node A ring-shaped connection.
Parallel execution verification
Definition: Information collection nodes cannot appear in parallel execution branches.
Example: Parameter extraction node 2 and parameter extraction node 3 are in different parallel branches.
Use Cases
Workflow is suitable for rigorous business scenarios that require high accuracy and strong interpretability, including but not limited to:
Invoice scenario: Collect user invoice title, tax number and other information, normalize the information through LLM, code and other methods, call tools to verify the information, and finally generate the invoice through tools and return to user.
Case handling scenario: Collect key information in the case, perform different handling based on logical judgment, and finally enter the final case handling result.
Patient information query scenario: Combine patient identity information to provide quick and accurate medical record query and management, summarize medical records, and perform QA based on the information.
Tencent Cloud Agent Development Platform (Tencent Cloud ADP)'s workflow function enhances business process configuration efficiency and accuracy to cater to various complex business scenarios.