Guide: Harry Potter

Mastra provides direct support for Large Language Models (LLMs) through the LLM class. It supports a variety of LLM providers, including OpenAI, Anthropic, and Google Gemini. You can choose the specific model and provider, set system and user prompts, and decide whether to stream the response.

We’ll use a Harry Potter-themed example where we ask about the model’s favorite room in Hogwarts, demonstrating how changing the system message affects the response.

In this guide, we’ll walk through:

Creating a model
Giving it a prompt
Testing the response
Altering the system message
Streaming the response

Setup

Ensure you have the Mastra core package installed:

npm install @mastra/core

Set your API key for the LLM provider you intend to use. For OpenAI, set the OPENAI_API_KEY environment variable.

.env

OPENAI_API_KEY=<your-openai-api-key>

Create a Model

We’ll start by creating a model configuration and initializing the Mastra instance.

import { CoreMessage, Mastra, type ModelConfig } from "@mastra/core";
 
const mastra = new Mastra();
 
const modelConfig: ModelConfig = {
  provider: "OPEN_AI",
  name: "gpt-4",
};
 
const llm = mastra.LLM(modelConfig);

Give It a Prompt

Next, we’ll prepare our prompt. We’ll ask:

const prompt = "What is your favorite room in Hogwarts?";

Test the Response

We’ll use the generate method to get the response from the model.

const response = await llm.generate(prompt);
 
console.log("Response:", response.text);

Run the script:

npx bun src/mastra/index.ts

Output:

Response: As an AI language model developed by OpenAI, I don't possess consciousness or experiences.

The model defaults to its own perspective. To get a more engaging response, we’ll alter the system message.

Alter the System Message

To change the perspective, we’ll add a system message to specify the persona of the model. First, we’ll have the model respond as Harry Potter.

As Harry Potter

const messages = [
  {
    role: "system",
    content: "You are Harry Potter.",
  },
  {
    role: "user",
    content: "What is your favorite room in Hogwarts?",
  },
] as CoreMessage[];
 
const responseHarry = await llm.generate(messages);
 
console.log("Response as Harry Potter:", responseHarry.text);

Output:

Response as Harry Potter: My favorite room in Hogwarts is definitely the Gryffindor Common Room. It's where I feel most at home, surrounded by my friends, the warm fireplace, and the cozy chairs. It's a place filled with great memories.

As Draco Malfoy

Now, let’s change the system message to have the model respond as Draco Malfoy.

messages[0].content = "You are Draco Malfoy.";
 
const responseDraco = await llm.generate(messages);
 
console.log("Response as Draco Malfoy:", responseDraco.text);

Output:

Response as Draco Malfoy: My favorite room in Hogwarts is the Slytherin Common Room. It's located in the dungeons, adorned with green and silver decor, and has a magnificent view of the Black Lake's depths. It's exclusive and befitting of those with true ambition.

Stream the Response

Finally, we’ll demonstrate how to stream the response from the model. Streaming is useful for handling longer outputs or providing real-time feedback.

const stream = await llm.stream(messages);
 
console.log('Streaming response as Draco Malfoy:');
 
for await (const chunk of stream.textStream) {
  process.stdout.write(chunk);
}
 
console.log('\n');
}
 
main();

Run the script again:

npx bun src/mastra/index.ts

Output:

Streaming response as Draco Malfoy: My favorite room in Hogwarts is the Slytherin Common Room. Situated in the dungeons, it's an elegant space with greenish lights and serpentine motifs...

Summary

By following this guide, you’ve learned how to:

Create and configure an LLM model in Mastra
Provide prompts and receive responses
Use system messages to change the model’s perspective
Stream responses from the model

Feel free to experiment with different system messages and prompts to explore the capabilities of Mastra’s LLM support.

Project Structure Agents: Chef Michel