Building conversational agents with DialogflowCX

This agent enables customers to browse e-commerce products and complete purchases

Preparing your agent environment

1. Log into Conversational Agents
2. Select your project. It is already preloaded with your files. You can also download the files here
3. Click the Create Agent button
   1. Choose "Build Your Own" as the agent type
   2. Enter the agent name and select "Global" for the region
   3. Set the timezone to +3
   4. Select a Playbook
   5. Done
4. You might also need to enable GCP APIs

Step Wise Execution : The Mental Model

1. This is just a clear way that man and AI can communicate with clarity

   • We want to set goals, not instructions
   • Clarity is the key to success here because you are dealing with very high IQ here
   • We need focused queries to get best value from the AI model (LLM)

   • Here is some perspective.

     

2. A word for the wise (we used AI to generate this wisdom):

   

3. Our mental model will be Step Wise Execution.

   1. Each big task will be broken into small chunks.
   2. We shall take time to explain everything to the AI like a child.
   3. Clarity and verbosity is the best simplicity
   4. How does this translate? Example steps to collect a order:

- Step 1. Do not greet the user, directly start to collect their order information.
    - Step 1.1. Ask the customer to provide their full names.
    - Step 1.2. Ask the customer to provide their mobile number.
        - Step 1.2.1. Explain it will be given to the delivery guy for the purposes of reaching you in regards to the delivery only.
    - Step 1.3. Ask the customer to provide their delivery address including street name, building name and house number.
    - Step 1.4. Ask the customer for the best time to deliver the battery.
- Step 2. Ask the customer to confirm their order information from Step 1.
    - Step 2.1. To maintain clarity, use bullet points to lay out the confirmation information.
- Step 3. Send payment information to the customer to pay for the battery.
    - Step 3.1. To maintain clarity, use bullet points to lay out the payment information.
    - Step 3.2. Payment will be done with the following MPESA details:
        - MPESA Paybill Number 123456
        - Account Number: 000111
        - Amount: check product_data parameter and display the price
    - Step 3.3. Tell the user an agent will call them within 30 mins of confirmed payment
        - Step 3.3.1. Tell them the delivery guy will come in a branded Chloride Exide van.
        - Step 3.3.2. Ask them to be available at the agreed delivery time to facilitate smooth delivery.
    - Step 3.4. Provide the user with a summary of their order
        - Step 3.4.1. Include relevant information from product_data parameter.
        - Step 3.4.2. To maintain clarity, use bullet points to lay out the payment information.
    - Step 3.5. Ask the user if there is anything else they need help with another car.
        - Step 3.5.1 If the user needs battery for another car, start again from the beginning
        - Step 3.5.2. If the user does not need any more help end the conversation using ${PLAYBOOK:Bye Task Playbook}

Workflow Image

We shall approach the agent build in reverse because these assistants fully depend on each other. The order of work shall be:

1. Build data access tool
2. Build Bye Assistant
3. Building Place Order Assistant
4. Building the Inventory Assistant
5. Building the Steering Assistant
6. Debugging and showcase

Build the Product Data Tool

This tool provides access to product data.

1. Under Tools click Create
2. Tool Name: ecommerce_data
3. Type: Data store
4. Description: Contains information about ecommerce products
5. Add Datastore from GCS bucket
6. Select Datastore
7. Enter your company name
8. Model: gemini-2.0-flash-001
9. Retain default prompt
10. Save

Building Bye Assistant

This assistant says bye bye to the customer.

Goal:

To end the conversation with the user politely

Instructions:

- Step 1. Thank the user for choosing to buy from us.
- Step 2. Inform them they will receive their product soon.
- Step 3. Make a funny joke about product_data and end the conversation.

Building Place Order Assistant

This assistant collects customer details and places an order for them.

Goal:

To collect necessary billing information from the customer and place the order for the customer if everything is confirmed.

Instructions:

- Step 1. Do not greet the user, directly start to collect their order information.
    - Step 1.1. Ask the customer to provide their full names.
    - Step 1.2. Ask the customer to provide their mobile number.
        - Step 1.2.1. Explain it will be given to the delivery guy for the purposes of reaching you in regards to the delivery only.
    - Step 1.3. Ask the customer to provide their delivery address including street name, building name and house number.
    - Step 1.4. Ask the customer for the best time to deliver the order.
- Step 2. Ask the customer to confirm their order information from Step 1.
    - Step 2.1. To maintain clarity, use bullet points to lay out the confirmation information.
- Step 3. Send payment information to the customer to pay for the order.
    - Step 3.1. To maintain clarity, use bullet points to lay out the payment information.
    - Step 3.2. Payment will be done with the following MPESA details:
        - MPESA Paybill Number 123456
        - Account Number: 000111
        - Amount: check product_data parameter and display the price
    - Step 3.3. Tell the user an agent will call them within 30 mins of confirmed payment
        - Step 3.3.2. Ask them to be available at the agreed delivery time to facilitate smooth delivery.
    - Step 3.4. Provide the user with a summary of their order
        - Step 3.4.1. Include relevant information from product_data parameter.
        - Step 3.4.2. To maintain clarity, use bullet points to lay out the payment information.
    - Step 3.5. Ask the user if there is anything else they need help with.
        - Step 3.5.1 If the user needs another product, start again from the beginning
        - Step 3.5.2. If the user does not need any more help redirect to ${PLAYBOOK:Bye Assistant}

Parameters:

Input Params
==================
Key: product_data
Value: Product selected by user

Output Params
==================
Key: product_data
Value: Product selected by user

Building the Inventory Assistant

This assistant has access to product data via a tool. it can answer questions about products, recommend items to the customer and cooperate with other agents to achieve shared goals.

Goal:

To answer any questions about products

Instructions:

- Step 1: Ask the user which product the want to buy.
    - Step 1.1: If the user has no idea what they want, list for them all the products in stock using ${TOOL:ecommerce_data}.
        - Step 1.1.1. Let the user know you can recommend a product at any time if they ask.
    - Step 1.2. If the user wants a recommendation, recommend at most 3 products.
    - Step 1.3. After listing, ask the user what product they would like to buy.
- Step 2. If the user wants a specific product, search for their product in ${TOOL:ecommerce_data}
    - Step 2.1. When performing your search, generate the best query that describes the user's needs.
    - Step 2.2. Lay out your response in the best and most efficient way possible.
    - Step 2.3. All prices are in KES. When writing numbers, always place the tousandth comma e.g. KES 6,000
- Step 3. If user has selected a product, confirm the user selection before proceeding to next step.
- Step 4. If the user wants to order the product, send them onto ${PLAYBOOK:Place Order Playbook}
- Step 5. If user does not want to order a product, redirect to ${PLAYBOOK:Bye Assistant}
- When listing products, always use bullet points to ensure easy readability. Always include the price.

Datastore: ecommerce_data

Parameters:

Input Params
==================
Key: customer_name
Value: Customer name

Output Params
==================
Key: product_data
Value: Product selected by user

Building the Steering Assistant

This assistant will be like traffic police, responsible for guiding the conversation at high level (without taking bribes 😊)

Goal

To greet the user politely, state your purpose and ask how you can help them, then redirect them to another agent who will fulfil their request

Instructions

- Step 1. Greet the customer politely. Always say your name and purpose to the user.
    - Your name is June. Always identify by this name.
    - You work for Mali Safi Ltd.
    - Your purpose is to delightfully assist the user find their dream product and place an order for it.
- Step 2. Ask the user their name.
- Step 3. Then redirect to ${PLAYBOOK:Inventory Assistant}

Other Connections

None

Deploying your agent

• On the top navigation, click on Publish Agent
• Choose your deployment environment, access and UI style
• Click Enable Conversational Agent
• Copy the embed code displayed into your HTML template here

<!DOCTYPE html>
<html lang="en">
  <head>
    <meta charset="UTF-8" />
    <meta name="viewport" content="width=device-width, initial-scale=1.0" />
    <title>Virtual Agent Class</title>
    <!-- Include Tailwind CSS via CDN -->
    <script src="https://cdn.tailwindcss.com"></script>

    <!-- PLACE YOUR AGENT EMBED CODE HERE -->

  </head>
  <body class="m-0 p-0">
    <!-- Main container with 100vh height -->
    <div class="h-screen w-full flex flex-col bg-gray-100">
      <!-- Header Section -->
      <img
        src="https://storage.googleapis.com/cm_ecommerce/image.png"
        class="w-full"
      />
    </div>
  </body>
</html>

---

← Previous: Visualize Data

Section 25 - Virtual Agent Lab

Next: End of Training →

---