Agent Input and Output
Agent input and output configuration is essential for implementing intelligent interactions. By properly configuring input variables, Agents can receive diverse parameters and contextual information. By defining output result formats, Agents can return structured data for programmatic use. Through streaming output, you can obtain real-time Agent runtime status and results. This article provides detailed instructions on these configuration methods and usage techniques.
Configuring input variables
Each Agent receives a default parameter when called: User Input. Developers can also customize additional input variables and reference these variables in system prompts or program logic.
Developers can switch to the Input Variables tab in the Agent visual editor, click the Add Input Variable button, fill in the variable title and select the variable data type in the popup dialog, then click the Confirm button.
Input variables support JitAi data types and their configuration options, including single-line text, multi-line text, numbers, amounts, dictionaries, and more.
Using variables in prompts
When calling an Agent, in addition to passing user input content, you can also dynamically pass in variables. These variables can be referenced in prompts, enabling more flexible context customization and information filling, allowing the Agent to dynamically adjust behavior and response content based on actual business scenarios.
Developers can trigger the variable selection list by typing { where they need to insert variables, then click on the corresponding variable to insert it into the prompt. During Agent runtime, variables are automatically replaced with actual values.
By default, only the User Input variable is available, but developers can add more through the input variables configuration.
Configuring output results
Agents support two output modes: JSON structured output and Plain text output. Each mode is described below:
JSON structured output
In this mode, the agent outputs results in JSON format. By default, an Agent outputs a dictionary containing only one output text field after completion. If you need to call the Agent in program logic or orchestrate the Agent with other nodes in AI Assistant, the Agent's output results must be structured data that can be parsed programmatically, rather than just a string.
Developers can switch to the Output Result tab in the Agent visual editor to modify the data type and title of the output field, or click the Add Output Result button to add new output fields.
Output results are automatically converted to JSON Schema that large language models can accurately understand and sent to the large language model service as part of the system prompt. If there are no special requirements, developers do not need to write output format related content in the system prompts.
The title and name of each output result field need to be carefully considered to be self-explanatory, so that the large language model can understand correctly.
Output result data types
Output results support various data types of the JitAi platform, including:
- Single-line Text: Suitable for short text output, such as titles, names, etc.
- Multi-line Text: Suitable for longer text content, such as descriptions, summaries, etc.
- Number: Suitable for numerical results, such as calculation results, statistical data, etc.
- Amount: Suitable for currency-related output
- Dictionary: Suitable for complex structured data output
- Boolean: Suitable for yes/no type judgment results
- Date Time: Suitable for time-related output
Choosing the appropriate data type helps the large language model understand output requirements more accurately and generate results that meet expected formats.
Plain text output
In this mode, the agent outputs the final result as plain text to the assistant conversation. Upon completion, it returns a dictionary containing only an output text field with the final result as its value. This field can be used when invoking the agent programmatically or when orchestrating the flow between the agent and other nodes in the AI Assistant.
Agent streaming output
During Agent execution, it continuously receives information such as large language model reasoning progress, text content, and tool calls. Developers can obtain and process this data in real-time using streaming methods.
AI Assistant already supports streaming output when calling an Agent. If developers call the Agent's run method through code, they can also pass in a custom streaming callback function to subscribe to and process streaming output messages in real-time. Please refer to Streaming Callback Processing in the reference manual for detailed usage.
Calling Agent in frontend functions
In actual business scenarios, it's often necessary to call AI capabilities in real-time based on user operations on pages (such as clicking buttons, filling forms, switching options, etc.) to implement intelligent recommendations, auto-completion, intelligent Q&A, and other functions. To meet these needs, the platform supports directly calling Agents in frontend function logic (page functions, page event functions, component event functions) and flexibly passing parameters, enabling AI capabilities to be seamlessly integrated into various frontend interactions, improving user experience and business intelligence levels.
As shown in the figure above, developers can insert a Basic Statement in the function logic, select to call the specified Agent's Run function, fill in user instructions in the parameter input box, then insert another Basic Statement to refresh the current page of the table. This achieves the functionality where clicking a button triggers the Agent to generate content based on user instructions and refresh the current page of the table.
Agent execution usually takes some time. If users wait too long, it will affect the experience. It's recommended to combine with AI Assistant to optimize the interaction flow and achieve continuous interaction between AI and users, improving the overall user experience.
Calling Agent in backend service functions
Backend service functions are commonly used for processing business flows, data processing, and other scenarios. When you need to introduce AI capabilities (such as intelligent analysis, content generation, automatic Q&A, etc.) into these processes, you can directly call Agents in service functions to combine AI with business logic, such as automatic approval, intelligent recommendations, content review, etc.
Developers can insert a Basic Statement in the service function editor, select to call the specified AI Agent's Run function, fill in user instructions in the parameter input box to implement the functionality of service functions calling the Agent. Switching to source code mode shows that only one line of code is needed to complete the call.
Testing Agent in page assistant
During Agent development, frequent debugging is required. Developers can create a simple AI Assistant that is internally orchestrated to directly call the target Agent, then integrate the assistant into a page. In the AI Assistant's dialog box, you can see the Agent's execution process and output results.