google--adk-python/contributing/workflow_samples/request_input_advanced

ADK Workflow Request Input Advanced Sample

Overview

This sample demonstrates advanced features for requesting Human-in-the-Loop (HITL) input dynamically during an ADK Workflow execution.

Specifically, it highlights how to pass structured data to the client UI using the payload parameter, and how to mandate a structured response type using the response_schema parameter on the yielded RequestInput event.

In this scenario, an employee requests time off by providing a natural language description of their request (e.g., "I need next Monday off to go to the dentist").

• An LLM agent (process_request) parses the natural language into a structured Pydantic model containing the number of days and a reason.
• A python node (evaluate_request) evaluates the parsed request:
  • If days <= 1, it yields a TimeOffDecision approving the request.
  • If days > 1, it yields a RequestInput to a manager. It attaches the request details to the payload so the client UI can render it. It enforces that the manager must respond with a JSON object containing an approved boolean and an optional approved_days integer by specifying response_schema with a valid Pydantic JSON schema.

Sample Inputs

Start the workflow by providing the initial time off request in natural language:

• I'm feeling under the weather and need to take today off.

  Parses as 1 day, auto-approves.

• Taking my family to Disney World, I'll be out for 5 days next week.

  Parses as 5 days, routes to manager review.

When the terminal prompts you as the manager, provide valid JSON matching the schema:

• {"approved": true, "approved_days": 5}
• {"approved": false, "approved_days": 0}

Graph

            [ START ]
                |
                v
        [ process_request ] (LLM Agent)
                |
                v
       [ evaluate_request ]
                | (Yields TimeOffDecision OR RequestInput event)
                v
       [ process_decision ]
                |
                v (implicit end)

How To

1. Define the Response Schema: Use a Pydantic model's model_json_schema() to get a standard layout of what the human should return.

   from typing import Optional
   from pydantic import BaseModel, Field
   
   class TimeOffDecision(BaseModel):
       approved: bool = Field(...)
       approved_days: Optional[int] = Field(None)
   

2. Yield a RequestInput: Pass the schema and optionally a payload for the client to display.

   def evaluate_request(request: TimeOffRequest):
       # ... logic to check if manager review is needed ...
       yield RequestInput(
           interrupt_id="manager_approval",
           message="Please review this time off request.",
           payload=request,
           response_schema=TimeOffDecision.model_json_schema()
       )
   

3. Parse the Resumed Input: When the workflow resumes, the node_input to the next node will be the parsed Pydantic model implicitly (if type-hinted).

   def process_decision(request: TimeOffRequest, node_input: TimeOffDecision):
       if node_input.approved:
           # ...