Application Scenarios

Currently, two types of agents can be built based on the agent development platform. One is the knowledge-based agent for text generation and text retrieval, such as the search Q&A assistant and code generation assistant. This type of agent is implemented the large model. The other is the process-based agent for complex workflow scenarios, such as financial analysis assistant and network detection assistant.

• Agent: uses a large model as the core of task execution. Users can configure information such as prompts and knowledge bases to implement tool planning and invocation. The advantage is that the tool can be developed without code and the dialog process is more intelligent. The disadvantage is that when the large model is restricted by input, so it is difficult to execute complex processes with long chains.
  • Scenarios
    1. Intelligent customer service
       • Scenario: e-commerce pre-sales consultation and after-sales Q&A
       • Application mode: By configuring prompts and knowledge bases, the model can directly understand user questions (such as return policies and offering parameter query), invoke plug-ins to obtain order information, and quickly generate natural language replies.
       • Advantages: No complex process design is required. Open-ended questions can be handled, and the conversation experience is closer to that of real-world conversion with humans.
    2. Education and coaching
       • Scenario: personalized Q&A and knowledge point explanation
       • Application mode: Based on the subject knowledge base and model inference capability, students can ask questions (such as mathematical problem analysis) through dialogs. The model generates problem solving steps in real time and provides feedback.
       • Advantages: No code is required to quickly cover requirements of multiple disciplines and flexibly adapt to different question modes.
    3. Medical consultation
       • Scenario: preliminary symptom analysis and health suggestions
       • Application mode: After a user describes a symptom, the model generates possible etiology based on the medical knowledge base and invokes the plug-in to obtain drug information or recommend a medical department.
       • Advantages: The interaction is natural, and the knowledge base can be quickly iterated and updated to adapt to the latest medical guide.

• Workflow: functions as the core of task execution. You can drag and drop nodes on the canvas to set up a task process. Scenario-specific nodes include the LLM node, IntentDetection node, Questioner node, Plugin node, Branch node, Code node, and Message node. The workflow's advantage lies in its high scalability and support for low-code development. However, it has low intelligence in dialog interactions, and maintaining numerous branches in complex scenarios can be challenging.
  • Scenarios
    1. Financial risk control
       • Scenario: loan approval and anti-fraud review
       • Workflow design: Nodes are connected in series to implement multi-step review (for example, calling the credit investigation API -> using a large model to analyze the proof of income -> triggering manual review by the Branch node).
       • Advantages: The process is transparent and controllable, and branch conditions are clear (for example, different approval paths are triggered by different quotas), preventing the hallucination of large models from interfering with key decision-making.
    2. Supply chain management
       • Scenario: inventory warning, automatic replenishment, and logistics exception handling
       • Workflow design: The system interconnects with the ERP system through the Code node, identifies the inventory threshold through the Branch node, and triggers the replenishment through the Plugin or Message node.
       • Advantages: Complex rules (such as supplier priority and logistics delay policy) can be accurately implemented through branch nodes, and the stability is higher than that of a pure model-driven model.
    3. Government approval
       • Scenario: enterprise qualification application and public benefit subsidy distribution
       • Workflow design: The intent recognition node (for classifying user requirements), questioner node (for supplementing materials), and code node (for verifying data compliance) are connected in series. The message node returns the result.
       • Advantages: The standardized process reduces errors caused by human factors. Branch nodes can meet differentiated policy requirements (for example, subsidy standards in different regions).