How to Build AI Agents

1. Go to the Raven home page. You’ll see your previous agents there. If it’s your first agent, click on ‘Create Custom Agent’

1. There are different sections in this page. You can refer to the left bar.

Getting Started: This is where you share the details and background information for your agent.

Actions: This is where you add the necessary resources to train your agent.

Test: This is where you test your AI agent, see how it responds to different inputs and make adjustments as needed.

Deploy: This is where you deploy your AI agent across your social and messaging channels.

Let’s get into this details of each section:

1. Getting Started:

Name of your agent: Give a name of your agent that is relevant to its purpose and helps you identify among multiple agents. Example: Customer Care Agent

What your agent will do: This is where you shape the AI agent’s purpose, behavior and personality. It sets the tone for how the AI interacts with users by offering initial guidelines or instructions. This can involve defining the AI’s role, preferred communication style, how to handle specific types of queries, how to use the resources that the agent will be trained on etc. Customizing this section helps ensure the AI aligns with your brand’s voice and delivers responses that meet your business objectives.

Best Practices:

1. Write in clear, natural language.

2. Use short, straightforward sentences.

3. Add examples to illustrate what a good response looks like.

4. Emphasize what you want the bot to do, rather than what to avoid.

5. Begin prompts with action verbs like generate, create, or provide to guide behavior effectively.

Sample Prompt:

You're a customer support agent for ABCD Airways International. Customers will ask you different questions about AAI, their services, different policies etc. Respond concisely and use the information provided to you.

You might get questions in Arabic where you should respond in Arabic. Your tone should be friendly and welcoming, similar to how air hostesses speak. Attach relevant pages from AAI’s website where applicable.

At any point, if you get stuck or don't have enough information, pass the conversation to a human agent with proper declaration to the customer.

4. Actions: The Actions tab is your control center for managing the content that fuels your AI agent. From here, you can upload documents, crawl websites or sitemaps, connect to external APIs, and route conversations to human agents when needed. These tools give you complete flexibility over the data your agent uses, helping it deliver accurate, relevant, and timely responses to your users.

Adding an Action: Click on ‘Add Action’ and select the action you’d like to add.

Each action will have a section where you need to specify how and when this particular action should be used by the agent. Consider this as a standalone prompt for your agent in order to help it differentiate between multiple actions and make sure it doesn’t hallucinate. Keep it short, concise and relevant to the action only.

Get Info from Website: Provide the URL of the website and click on ‘Fetch Links’

This will import all the pages available in that website.

It will automatically crawl and show all the pages available in that website. You can now select the pages you want to train your agent with and click on ‘Train’

It might take a few seconds for the training to complete. You’ll see a ‘Trained’ badge once the training is complete.

Get Info from Documents: Accepted file types are .txt, .pdf, .doc and .docx up to 50 MB in size.

You can upload multiple documents and select the ones that you’d like to use for training.

Once the training is complete, you’ll see the ‘Trained’ badge against the document(s).

Make API Call: If you want to connect external systems with your agent in order to fetch or pass information between these systems, you can connect APIs. Add a unique name for you to differentiate between multiple APIs and give it a description as to how the API should be used by the agent.

Get the details of the API structure and add it here.

Assign Ticket: If you want your human agents to be in the mix, so that your AI agent can effortlessly pass on a conversation to them, you can use the Assign Agent action.

Give a description for when conversations should be passed on to agents. And select the agent/group to whom you want the AI agent to assign the conversation.

5. Test: Testing your AI agent and tuning it to your liking is one of the most important steps of building an AI agent.

Put yourself in the shoes of your customers and ask the AI agent questions that it may receive.

See how it’s responding, whether it is using the actions as and when you’ve instructed it to and following the guidelines you had mentioned at the beginning. Any deviance, you can go back to the respective section and make changes as needed. Remember to clear the chat history before testing any new changes you have done.