Level Up Your Customer Service: A Startup’s Step-by-Step Guide to Vertex AI Conversation’s Generative Playbooks

An illustration that depicts the personalized, targeted experience that a customer enjoys when interacting with a virtual agent that was created using Vertex AI Conversation.

Are you struggling to keep up with the constant flow of customer inquiries? Do you wish you could offer exceptional support around the clock, without sacrificing personalization or efficiency? Imagine an AI-powered ally that not only understands your customers’ needs but also delivers tailored responses, streamlines your support processes, and even uncovers hidden insights within every interaction. That’s the potential of Google Cloud’s Vertex AI Conversation.

Vertex AI Conversation is revolutionizing customer engagement for small businesses and startups, regardless of their stages (early or later). By harnessing the power of advanced natural language understanding and machine learning, it empowers you to create intelligent virtual agents that consistently provide top-notch customer experiences. In this blog post, we’ll delve into the remarkable features and tangible business value of Vertex AI Conversation. But we won’t stop there — we’ll also guide you through building your own conversational AI application from the ground up, step by step.

What is Vertex AI Conversation?

Vertex AI Conversation, a cornerstone feature within Vertex AI Agent Builder, is a fully managed platform that unlocks the potential of conversational AI for everyone. By removing the barriers of complex coding, it empowers individuals and organizations of all technical backgrounds to effortlessly design and implement intelligent conversational experiences.

Vertex AI Conversation is designed to be user-friendly and adaptable, offering a range of unique benefits:

• Scalable and Flexible: Easily scale your conversational AI solutions to meet the growing demands of your business, ensuring a smooth and consistent user experience.
• Personalized Customer Interactions: Tailor conversations to individual customer preferences and needs, fostering deeper engagement and loyalty.
• Multilingual Support: Cater to a global audience with virtual agents that can understand and respond in multiple languages.
• Data-Driven Insights: Uncover hidden patterns and trends in customer conversations to gain valuable insights into their preferences, pain points, and needs.
• Reduced Operational Costs: Automate repetitive tasks and free up human agents to focus on more complex issues, resulting in significant cost savings.
• Scalable and Flexible: Easily scale your conversational AI solutions to meet the growing demands of your business, ensuring a smooth and consistent user experience.
• 24/7 Customer Support: Provide round-the-clock support with virtual agents that never tire, ensuring customers receive assistance whenever they need it.

Vertex AI Conversation: Built for Everyone, Transforming Every Industry

• E-commerce, revolutionizing how your business sells: Imagine you own a thriving online store. Vertex AI Conversation can be your virtual sales associate, available 24/7 to answer questions, offer personalized product recommendations, and guide customers through the purchasing process.
• Healthcare, optimizing patient care: If you run a busy medical practice, Vertex AI Conversation can streamline operations by acting as a virtual receptionist. It can schedule appointments, answer common questions, triage patient needs, and even send appointment reminders, allowing your staff to focus on providing quality care.
• Financial Services, enhancing client experiences: For financial advisors, time is precious. Vertex AI Conversation can automate routine tasks like answering account balance inquiries, providing transaction history, and even offering basic financial guidance, freeing you up to focus on building relationships and providing high-value services to clients.
• Education, personalizing the learning journey: Imagine a virtual tutor available at any time to answer students’ questions, provide personalized feedback, and offer tailored support. Vertex AI Conversation can make this a reality, creating a more engaging and effective learning environment for students of all ages.
• Content Creation & Marketing, boosting audience engagement: In the digital age, audience engagement is key. Vertex AI Conversation can help you build chatbots that interact with your audience on social media platforms, answering questions, providing recommendations, and fostering a sense of community around your brand.

No matter your industry, Vertex AI Conversation empowers your business to engage with customers more effectively, streamline operations, and drive growth through intelligent, personalized interactions. It’s the AI-powered conversational tool that’s transforming industries and making every interaction count.’

Unlock the Power of Conversation: Build Your Own GenAI-Powered Virtual Agent

Vertex AI Conversation is engineered for accessibility. Its user-friendly interface and pre-configured elements enable rapid and straightforward development. In this in-depth tutorial, we’ll guide you through creating a conversational AI solution tailored for the tourism industry: a virtual travel agent that can answer questions about South Florida, offer personalized recommendations based on user preferences and budget, and even book the selected experience.

Disclaimer: The data used in this demonstration is publicly available and does not represent any affiliation with or endorsement by any specific organization or entity.

Prerequisites

• Learn more about Agent Builder and its underlying architecture here: Vertex AI Agent Builder
• A Google Cloud Platform project with billing enabled.
• Enable Vertex AI Agent Builder and Cloud Storage APIs.
• [Optional] If you want to retrieve your data from Google Cloud Storage, make sure to upload your files to a Google Cloud Storage Bucket before creating the agent.

Step 1:

On your Google Cloud Platform, Search for “Agent Builder” and select the “Agent Builder: A platform for search and conversation apps” option under Products & Pages.

Step 2:

At the top of the page, click on the “+ Create App” button.

Step 3:

Note: Select “Agent” as your app type. Please note that the “Preview” icon will appear next to “Agent” until your project is allowlisted access to generative features.

Agent Builder also offers three additional app types, including the following:

1. Search: Create information retrieval applications quickly and easily. Our platform delivers high-quality results from the start, with the flexibility to customize the search experience to fit your unique requirements.
2. Chat: Create intelligent conversational AI agents that can understand and respond to complex questions out-of-the-box.
3. Recommendations: Create a state-of-the-art content recommendation engine that suggests relevant items to users based on their preferences and behavior, leveraging unstructured data (PDFs), structured data (NDJSON), or website data (URLs) from your sources.

Please ensure that you pick the app type that best aligns with your business use case and technical requirements.

Step 4:

Give your agent a meaningful name. Choose a name that reflects the role your agent will play. In the example above, “South Florida Travel Agent” is a fitting name for an agent that specializes in answering questions about and booking experiences in South Florida.

Select your agent’s operating region. This determines where your agent’s data will be processed. “Global” is the most common choice, offering the fastest response times. However, if your organization has specific data residency requirements, choose the region that best aligns with your needs. Remember, this choice cannot be changed later.

Step 5:

Agent Name: Think of this as a concise nickname for your agent. It serves as a quick signal to both you and the underlying Large Language Model (LLM) about the agent’s core purpose. Keep it short and sweet — typically 5 words or less.

Agent Goal: This is where you define the agent’s mission statement. What overarching objective should the agent be working towards? The goal you set here provides crucial context for the LLM, guiding its responses and actions to align with the desired outcome.

By defining both name and goal, you create a clear framework that helps the LLM understand the agent’s identity and purpose, ultimately leading to more relevant and effective interactions.

Step 6:

Now that you’ve defined your agent’s purpose, it’s time to chart the course it will take to achieve its goal. This involves creating a clear, step-by-step plan, known as the instructional execution path.

Think of it like a roadmap: each step should logically follow the next, guiding the agent towards the desired outcome. Here are some tips for crafting an effective execution path:

• Break it down: Divide the overall goal into smaller, manageable tasks.
• Be specific: Clearly define the actions the agent needs to take at each step.
• Consider tools (optional): If your agent will utilize external datastores or resources, you can reference them directly in your plan using the following syntax: ${TOOL:tool name}. For example, ${TOOL:South Florida Travel Docs} might indicate accessing a datastore that holds PDFs related to South Florida travel. We’ll explore how to connect datastores to your agent later.
• Collaborate with other agents (optional): If your project involves multiple agents working together, you can specify their interactions in the execution path using this syntax: ${AGENT:agent name}. For example, ${AGENT:Data Analyst} could signal that another agent will be responsible for analyzing data. We’ll delve into agent collaboration in more detail soon.

Remember, a well-structured execution path is essential for ensuring your agent operates effectively and achieves its intended purpose.

Step 7 (Optional):

Simply click the wrench icon on the left side of your screen. You’ll then see a “+Create” button at the top — click that to begin building your tool.

A note on Code Interpreter: You might be wondering what Code Interpreter is. Code Interpreter is a built in tool that is hosted and managed by Google, offering a seamless way to add functionality without any manual setup. The Code Interpreter is particularly versatile, capable of handling tasks like data analysis (with libraries like Pandas), generating visualizations (with libraries like Matplotlib), processing text, and even solving complex mathematical equations.

Let’s bring your tool to life! Start by giving it a descriptive name, then choose from these three powerful tool types:

1. OpenAPI Tools: These tools allow your agent to tap into the power of external APIs using the standardized OpenAPI specification. By providing the API’s schema (a blueprint of its structure and available endpoints), you can easily integrate with a wide range of third-party services and data sources, like weather APIs, payment gateways, or social media platforms.
2. Datastore Tools: Datastore tools provide a direct connection between your agent and both structured (e.g., databases, spreadsheets) and unstructured (e.g., documents, images) data sources. This enables your agent to answer user questions by referencing information stored in your knowledge base, ensuring accurate and up-to-date responses.
3. Function Tools: Function tools are a unique way to extend your agent’s capabilities by integrating with custom code that you define within your own application or environment. This is particularly valuable when you need to access functionalities or perform actions that aren’t easily handled by built-in tools or external APIs.

In this article, we’ll focus specifically on how to leverage the data store tool to enhance your agent’s capabilities.

To help the LLM make the most of your datastore, consider providing a short description outlining what kind of information it contains and how it’s organized.

Next, create the data store. Do so by clicking “Create a data store”.

You’ll be taken to a page (above) where you can choose how to populate your datastore:

• Website Content: Automatically crawl and extract information from public websites you specify. Note: If you enable advanced website indexing for faster and more comprehensive crawling, you’ll need to verify that you own the domains of the websites you want to index. Without verification, these websites cannot be indexed.
• Cloud Storage: Seamlessly import data that’s already stored in your Google Cloud Storage bucket.
• API: If you prefer a more hands-on approach, you can manually import data through the API.

Let’s import data from a Cloud Storage bucket.

Data: Select the type of data you are importing. The options are:

• Unstructured documents (PDF, HTML, TXT, and more)
• Linked unstructured documents (JSONL, with metadata)
• Structured FAQ data for a chat application (CSV)

Select a folder or a file you want to import: Click on the “Browse” button and navigate to the folder or file within your Cloud Storage bucket that you want to import into your datastore.

Enter the GCS location: In the text field below “Browse,” provide the full Google Cloud Storage (GCS) path to the folder or file you selected. This should look something like: gs://your-bucket-name/your-folder-or-file

Continue: Click the “Continue” button to proceed to the next step in the data import process.

Datastore Name: Give your datastore a unique and descriptive name. Remember, this name cannot be changed later, so choose carefully.

Access Control Information (Optional): If this datastore contains sensitive or restricted data, check the box to indicate that it requires access control measures.

Document Processing Options (If Applicable): If you’re uploading unstructured documents, you may have additional options here for processing the data (like splitting large documents into smaller chunks for easier analysis).

Create: Once you’ve filled in all the required fields and configured any optional settings, click the “Create” button to create your datastore.

Make sure you select the right datastore type based on the kind of information you’ve stored. In our case, since we’re working with text documents in a Cloud Storage bucket, we’ll choose “Unstructured documents” as the most suitable format. This ensures that the data is handled and processed in a way that optimizes its usefulness for your agent.

Step 8:

In Step 6, we saw how one agent can seamlessly pass the torch to another. In our example, the “South Florida FAQ Agent” effectively references the “Activity Booking Agent” in line three of its instructions. This means that when a user expresses the intent to book an activity, the South Florida FAQ Agent gracefully hands off control to the Activity Booking Agent. From that point on, the user’s journey is guided by the instructions and logic defined within the Activity Booking playbook, ensuring a smooth and specialized booking experience. Please note that agents have the flexibility to reference and collaborate with other agents at any stage within their instructional execution path.

In the Activity Booking Agent, users are presented with tailored recommendations for South Florida activities based on their preferences. To ensure seamless booking, relevant user information is collected. Finally, the interaction is routed to a structured conversational flow (CF) where the data gathered during the conversation is stored in BigQuery. This step showcases how Agent Builder can be seamlessly integrated with Dialogflow CX, leveraging Python code to streamline data extraction and storage. While we won’t delve into this specific integration in this article, keep an eye out for a future piece that explores this exciting topic in depth.

Step 9 (Optional);

Examples (optional) are sample conversations that demonstrate the desired interaction between a user and your agent. These “few-shot prompts” help guide the agent’s responses and actions. By providing diverse examples, you teach your agent to understand user intent, respond appropriately, and even perform specific tasks.

Display name: Think of this as the title for your conversation scenario. It’s a way to label different examples to keep things organized.

Basic Settings: Here’s where you control how each example is used.

• Selection strategy: This is like choosing the difficulty level for your AI. “Auto” is the default and lets the system decide if an example is needed. “Always” forces the AI to use the example, while “Disabled” ignores it completely.

Input & Output: This is the heart of the training process. Here, you write out a back-and-forth between a user and your agent. It’s like scripting a scene in a movie, but instead of actors, you’ve got your AI learning the lines.

End example with output information: This is where you tell the AI how the conversation should wrap up.

• Summary: A brief overview of what happened in the conversation. This helps the AI understand the key takeaways.
• Conversation state: Did the agent achieve its goal (“OK”), did the user leave (“CANCELLED”), was there an error (“FAILED”), or was a human needed (“ESCALATED”)?

Adding Actions: Each example consists of a series of actions (user input, agent response, tool use). You can manually add these (by simply clicking the + sign found at the bottom of the image below) or have the system generate them automatically based on your instructions.

Let’s Test!

Key Takeaways

Scalable and Flexible: Easily scale your conversational AI solutions to meet the growing demands of your business, ensuring a smooth and consistent user experience.

Versatility Across Industries: Vertex AI Conversation isn’t limited to one sector. It’s revolutionizing e-commerce, healthcare, finance, education, and content creation by automating tasks, enhancing customer experiences, and freeing up human resources.

Data-Driven Insights: Beyond customer service, Vertex AI provides valuable insights into customer behavior, preferences, and pain points through analysis of conversations. These insights can inform business strategies and improvements.

Call to Action

Some additional resources are included below and each can serve as a quick value add for your business:

• Elevate your document processing with Document AI. Extract structured data from your documents effortlessly, saving valuable time and resources when dealing with large volumes of information.
• Unlock the power of Google-quality search with Vertex AI Search . Build RAG-powered search applications out-of-the-box, providing your users with accurate and relevant results.
• Tap into Google’s cutting-edge AI with Model Garden. Access high-performance foundational models tailored to your unique business requirements, propelling your machine learning initiatives to new heights.