Building conversations

1. Building conversations

So far, we've explored how to generate single responses. We'll now look at how to create a model that remembers previous interactions and maintains context.

2. Maintaining context

Imagine an AI assistant developed for a travel booking company that helps customers plan vacations. A user could start by asking about popular destinations in France

3. Maintaining context

, then, based on the AI-generated response, follow up with questions

4. Maintaining context

about more destinations, flights, hotels, and local attractions. Instead of treating each query as a separate request, the model should

5. Maintaining context

remember what was previously discussed and use

6. Maintaining context

that information in its responses. This means we need a structured way to track conversation history and maintain a consistent dialog, which can be done with Llama by implementing a conversation class.

7. Conversation class

Through this class we'll organize the code so that it can store a history of prior messages, and new messages will be added to the list as the conversation with Llama continues. The class starts with a method called 'init'. This is the constructor method that runs every time we create a new instance of the class. It allows us to initialize attributes: in this case, we store an instance of the Llama model and setting up a system prompt. This defines how the assistant should behave throughout the conversation. The system message is stored at the start of the history list.

8. Conversation class

Next, we define create_completion(), which: Adds the user input to the conversation history; calls create_chat_completion() to generate a response; and finally, extracts the assistant's reply and adds it to the conversation history. With this setup, every new input builds on previous exchanges, and this helps to ensure a more continuous conversation rather than a series of separate responses.

9. Running a multi-turn conversation

Let's have a look at how to interact with our conversation class. First, we load our model and initialize the class with a system prompt, such as asking to act as a virtual travel assistant helping with planning trips. We then pass a prompt to the create_completion method, asking for some destinations in France for a short weekend break. Llama returns some options, and we follow up with another question about destinations in Spain.

10. Running a multi-turn conversation

Because conversation history is tracked, Llama understands the reference to the weekend

11. Running a multi-turn conversation

break and provides a relevant response. Multi-turn conversations make interactions feel more natural;

12. Let's practice!

let's see some in action.

Create Your Free Account

By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.