What Can You Build with the AI Agent Builder?

Overview of the AI Agent Builder

The AI Agent Builder offers an intuitive interface with key sections that allow you to configure your agent. Here’s a breakdown of each section:

• Description: Give your agent a name (e.g: "Support Bot") and a short description of its purpose (e.g: "Helps customers with common questions"). This keeps things organized and the information can be used to determine which agent is suitable during the handoff process.
  
  
• AI Model: Choose the AI “brain” that powers your agent. Different models are suited for different tasks, like chatting or analyzing data. Beginners can start with a default model.
  
  
• Knowledge Sources: Tell your agent where to find information. You can point it to any app content like pages, products, or tasks. For example, a support agent might use Documentation pages.
  
  
• Knowledge Selection: Select the number of relevant records (1,2,3 or 5) to provide to the agent.
  
  
• System Prompt: Write a short instruction that defines your agent’s role and tone. For instance, "You’re a friendly assistant helping users navigate the site.". You may also put common knowledge information here instead of retrieving from the knowledge sources.
  

• Agent Memory: Decide how much of a conversation your agent remembers. This helps it stay on track during chats. You may choose to remember user messages only or both user and agent's output messages. This settings can be changed on site level (each site has its own settings).
  
  
• Access Controls: Set who can use your agent (per site):
  
  
• Staff: Internal team only.
  
  
• Clients: Registered users.
  
  
• Public: Everyone visiting the site.
  

SGI Token - The gas to power the AI Agents

Building Your First AI Agent: A Beginner’s Guide

• What It Does:
  This feature lets your agent understand the meaning behind questions, not just exact words. It’s great for handling varied or tricky queries. The agent will use the following record fields to look for relevant information: record name, title, description, content, collection, status (if set), product variant's ID, SKU, price, options (if available). Depending on the Knowledge Selection number (n), it will take the top "n" most relevant records as the references for generating output. The more records are selected, the more SGI token is required. Without Vector Embeddings, the agent will extract the keyword from the user query and search records' name for that keyword to find relevant information. 
• Example:
  A user asks, "How fast is delivery?" The agent, using vector embeddings, can also respond to "When will my order arrive?" by recognizing similar meanings.
• How to Use:
  Turn on vector embeddings in the settings. Depending on the amount of data, it will take a while to turn your information to vector embeddings to be used for semantic searching. When completed, the agent will be set to use vector embeddings (Status: Ready)

• What It Does:
  Include the user (and optionally agent messages) in the query to provide the context for the agent to understand what the conversation is about. Without memory, the agent won't know what the user asks in the previous message(s).
• Example:
  The user may ask for a pair of shoe size 10 in the first message. The agent will show the user some pairs. The user will then ask the agent to show black leather shoes. With memory, the agent will understand the user wants to see black leather shoes in size 10 and show them. Without memory, the agent will show black leather shoes in any size.
• How to Use:
  You can choose to include 6/12/18 or all messages. You can choose whether to include agent's output messages or just user messages, the latter choice does not provide the full context but helps the agent understand what the user is looking for while reducing the number of SGI token needed for processing. Please keep in mind that the more messages to include, the more SGI token is required.  

• What It Does:
  The user query will be examined and may be passed to more relevant agent specified in the handoff list. This will ensure the query is handled by the most appropriate agent. This is ideal for layered support systems.
• Example:
  A general support chatbot gets a query like "Fix my server error." It hands off to a Technical Support Agent you’ve built, configured for deeper troubleshooting.
• How to Use:
  You can list one or more agents as the target agents for handoff, just make sure those agent's description clearly describes what they can do. When the user query is received, the AI will examine the query, the context (using memory) and the listed agents' description to determine if passing/handing off to one of those agents is more appropriate. If it is, then the query will be handed off/passed to the chosen agent for handling, bypassing the original agent for that turn. 

• Content Curator: Uses a translator agent through automation to translate CMS articles from the main language to other languages and include all of them in the same article to support multiple languages.
• Sales Assistant: Hands off complex support questions to a specialized agent while handling basic product info itself.

• How It Works: Agents can pull data from CMS pages, products, or any applications as their knowledge base. This keeps responses fresh and accurate.
• Example: A Product Info Agent uses CMS product descriptions to answer questions like "What’s this item made of?" without manual updates.

• How It Works: Embed your agent as a chatbot on your site. It can chat with users, answer questions, or guide them—all powered by CMS data.
• Example: A Help Desk Chatbot greets visitors, pulls CMS troubleshooting guides, and answers queries like "How do I reset my password?"

• How It Works: Beyond chatbots, agents can automate internal processes using automation. Just pass any query to an agent, it will return the output in $response.output.content
• Example: A Translator Agent translates blog posts to other languages automatically whenever you write a blog.

Security Considerations and Best Practices

• Set Access Controls:
  Limit who can use the agent. A staff-only Admin Assistant Agent should never be public.
• Set Personal Knowledge:
  Limit agent to access personal/private knowledge only. That means the agent cannot access knowledge created by other users.
• Control Memory:
  Don’t let the agent store too many messages—it wastes resources and could leak info. Stick to what’s needed (e.g., 5 messages).

• Test with "Preview" before launching to catch issues.
• Update knowledge sources regularly for accurate responses.
• Start simple (e.g., a basic chatbot) and add features as you learn.