Skip to content

Making Edits to JSON of an Agentic AI Card

⚠ NOTE

This guide is intended for power users who want to manually configure Agentic AI Cards using JSON.
Regular users can use the Card Manager within the UI for an easier experience.

Introduction

This guide explains how to create and customize Agentic AI Cards via JSON templates. By following these steps, you can quickly set up a new card tailored to your needs.

Overview

An Agentic AI Cards defines a multi-step process for a specific task. It comprises two key sections:

  • Metadata: Configuration settings (IDs, title, icon, etc.).
  • Workflow Process: A sequence of steps executed by the agent.
  • Each step includes:
    • input_mode (user or assistant)
    • tool
    • message
    • parameters

1. Base Template

Below is the JSON template for a Create Assessments card, illustrating a simple document-based process:

{
  "bundle_id": "SMB/Consumer",
  "app_id": "DOCUMENT_SEARCH",
  "category": "Education",
  "icon": "/assets/cards/card-analyze.svg",
  "title": "Create Assessments",
  "description": "This workflow demonstrates how to create multiple-choice, fill-in-the-blank, and essay questions aligned with a given document or set of learning objectives.",
  "required": "llm",
  "workflow": [
    {
      "step": 1,
      "input_mode": "user",
      "tool": "pick_integration",
      "message": "Pick your integration",
      "parameters": { "integration": "" }
    },
    {
      "step": 2,
      "input_mode": "user",
      "tool": "pick_file",
      "message": "Pick the file containing learning objectives or source material",
      "parameters": {
        "file_path": "",
        "integration": "$1.output"
      }
    },
    {
      "step": 3,
      "input_mode": "user",
      "tool": "get_user_input",
      "message": "Clarify assessment scope",
      "parameters": {
        "question": "What specific topics or key objectives would you like the assessment to focus on?",
        "description": "Please provide any particular learning goals or areas of emphasis for the questions.",
        "input": ""
      }
    },
    {
      "step": 4,
      "input_mode": "assistant",
      "tool": "search_docs_vdb",
      "message": "Retrieving relevant details",
      "parameters": {
        "query": "$3.output",
        "integration": "$1.output",
        "source": "$2.output",
        "n_chunks": 5
      }
    },
    {
      "step": 5,
      "input_mode": "assistant",
      "tool": "generate_llm_response",
      "message": "Generating assessment questions",
      "parameters": {
        "query": "You are an AI assistant specialized in education. Using the provided learning objectives or source material, create multiple-choice, fill-in-the-blank, and essay questions that align with the user's specified topics: $3.output",
        "document": "$4.output",
        "custom_instructions": ""
      }
    }
  ],
  "output_format": "text"
}

2. Editing Metadata

Configure the following top-level fields:

  • Bundle ID - bundle_id
    Choose one or more: SMB, Consumer, Business (e.g., "SMB/Consumer/Business").
  • App ID - app_id
    Defines the feature: DOCUMENT_SEARCH or IMAGE_SEARCH.
  • Category - category
    Select from the Available Categories.
  • Title - title
    The card's display title.
  • Description - description
    A brief summary of the card's purpose.
  • Required - required
    Specify "llm" if a language model is required.

Available Categories

  • Education
  • Enterprise Communications
  • Events & Scheduling
  • Finance
  • Food & Recipes
  • General
  • Health & Wellness
  • Home & Lifestyle
  • Human Resources (HR)
  • IT & Security
  • Marketing
  • Operations
  • Personal Finance
  • Product Development
  • Research & Development
  • Sales & Customer Service
  • Shopping & Gifting
  • Social
  • System
  • Tech & Digital Tools
  • Travel & Adventure

3. Defining Workflow Steps

Each workflow step must include:

  1. Order (step): Execution sequence.
  2. Input Mode (input_mode): user or assistant.
  3. Tool (tool): Choose from built-in tools.
    • User Input Tools: pick_integration, pick_file, get_user_input
    • LLM Processing Tools: search_docs_vdb, generate_llm_response
    • Other Tools
  4. Message (message): UI prompt.
  5. Parameters (parameters): Tool configuration.
    • In some cases where the user input is required these parameters can be a query or message that is retrieved from the user side.
    • In other cases where the assistant (LLM) is required, parameters can contain multiple fields such as query, integration, document, number of chunks, message from the user.

4. Update & Test

  1. Open the Generate Source code in Visual Studio Code Editor.
  2. Install the Database Client extension.
  3. Connect to your database and clear the Workflow table.
  4. Copy your updated JSON to the appropriiate category under: \Generate AI Manager\data\workflow\cards.
  5. Restart the Generate application.
  6. Verify your new card appears on the homepage.

Other Available Tools

You might have noticed that each step contains a parameter called tool, which is a function that allows the user to perform a specific task. Below is the current list of available tools:

  • Get User Input
    Asks for user input.

  • Search Vector DB
    Runs a similarity search on the Vector Database.

  • Generate LLM Response
    Prompts the LLM to generate a context-based response based on the provided input information.

  • Write File
    Writes output to a file.

  • Pick Integration
    Allows the user to pick their choice of integration (OneDrive, SharePoint, Google Drive, Slack, etc.).

  • Pick File
    Allows the user to pick one or more documents from the selected integration.

  • Search Website
    Retrieves data from a website based on the provided URL and prompt.

  • Search YouTube
    Retrieves video links from YouTube.

  • Search Google
    Retrieves links from Google based on user input.

  • Combine Steps
    Combines results from previous steps.

  • Fetch YouTube Transcript
    Retrieves the transcript from YouTube videos.

Example Workflow Steps

  • RFP Response - Combine Sections

    From RFP Response, you can create sections to segment bigger documents and then combine them all in one step:

    {
"step": 13,
"input_mode": "assistant",
"tool": "generate_llm_response",
"message": "Generating Conclusion...",
"parameters": {
    "query": "Generate a concise Conclusion section for a RFP response document using the RFP Request and the supporting information, only focus on the Conclusion, be truthful and limit the information to that found in the provided information. RFP Request: $3.output",
    "document": "Supporting Information: $4.output",
    "custom_instructions": ""
}
},
{
"step": 14,
"input_mode": "assistant",
"tool": "combine_llm_outputs",
"message": "Combining all sections into a structured RFP response.",
"parameters": {
    "steps": "Executive Summary:\n\n'$6.output' [SEP]\n\nProposed Solution:\n\n'$8.output' [SEP]\n\nImplementation & Support:\n\n'$10.output' [SEP]\n\nPricing:\n\n'$12.output' [SEP]\n\nConclusion:\n\n'$13.output' [SEP]"
}
}

  • Search Website

    From Search Website, this allows you to search the web from a specific link based on a user’s query:

    {
    "step": 3,
    "input_mode": "assistant",
    "tool": "search_website",
    "message": "Fetching information from Website",
    "parameters": {
        "link": "$1.output",
        "query": "$2.output"
    }
}

  • Find YouTube Videos

    From Find Youtube Videos, it allows you to search for youtube videos within the Agentic AI Cards:

    {
    "step": 2,
    "input_mode": "assistant",
    "tool": "search_youtube",
    "message": "Searching YouTube",
    "parameters": {
        "query": "$1.output"
    }
}

Editing the JSON gives you full control over AI-powered workflows. Use the Card Manager UI for simpler tasks, or dive into JSON for advanced configurations. Enjoy the flexibility!