Dotprompt is an executable prompt template file format for Generative AI. It is designed to be agnostic to programming language and model provider to allow for maximum flexibility in usage. Dotprompt extends the popular Handlebars templating language with GenAI-specific features.
An executable prompt template is a file that contains not only the text of a prompt but also metadata and instructions for how to use that prompt with a generative AI model. Here's what makes Dotprompt files executable:
-
Metadata Inclusion: Dotprompt files include metadata about model configuration, input requirements, and expected output format. This information is typically stored in a YAML frontmatter section at the beginning of the file.
-
Self-Contained Entity: Because a Dotprompt file contains all the necessary information to execute a prompt, it can be treated as a self-contained entity. This means you can "run" a Dotprompt file directly, without needing additional configuration or setup in your code.
-
Model Configuration: The file specifies which model to use and how to configure it (e.g., temperature, max tokens).
-
Input Schema: It defines the structure of the input data expected by the prompt, allowing for validation and type-checking.
-
Output Format: The file can specify the expected format of the model's output, which can be used for parsing and validation.
-
Templating: The prompt text itself uses Handlebars syntax, allowing for dynamic content insertion based on input variables.
This combination of features makes it possible to treat a Dotprompt file as an executable unit, streamlining the process of working with AI models and ensuring consistency across different uses of the same prompt.
Here's an example of a Dotprompt file that extracts structured data from provided text:
This Dotprompt file:
- Specifies the use of the
googleai/gemini-1.5-promodel. - Defines an input schema expecting a
textstring. - Specifies that the output should be in JSON format.
- Provides a schema for the expected output, including fields for name, age, and occupation.
- Uses Handlebars syntax (
{{text}}) to insert the input text into the prompt.
When executed, this prompt would take a text input, analyze it using the specified AI model, and return a structured JSON object with the extracted information.
The remainder of this getting started guide will use the reference Dotprompt implementation included as part of the Firebase Genkit GenAI SDK. To use other implementations of Dotprompt, see the list of Implementations.
First, install the necessary packages using NPM. Here we'll be using the Gemini API from Google as our model implementation:
npm i genkit @genkit-ai/googleaiAfter installation, you'll need to set up your environment and initialize the Dotprompt system. Here's a basic setup:
import { genkit } from "genkit";
import { googleAI } from "@genkit-ai/googleai";
// Configure Genkit with the GoogleAI provider and Dotprompt plugin
const ai = genkit({
plugins: [googleAI()],
// promptDir: 'prompts', /* this is the default value */
});
// Now you're ready to use Dotprompt!Note: You will need to set your Google AI API key to the GOOGLE_API_KEY
environment variable or pass it as an option to the googleAI() plugin
configuration.
With this setup, you can now create .prompt files in your project and use them
in your code. For example, if you have a file named extractInfo.prompt with
the content from the earlier example, you can use it like this:
const extractInfo = ai.prompt("extractInfo");
const { output } = await extractInfo({
text: "John Doe is a 35-year-old software engineer living in New York.",
});
console.log(output);
// Output: { "name": "John Doe", "age": 35, "occupation": "software engineer" }This setup allows you to leverage the power of Dotprompt, making your AI interactions more structured, reusable, and maintainable.
By following these steps, you'll have a basic Dotprompt setup ready to go. From here, you can create more complex prompts, integrate them into your application, and start harnessing the full power of generative AI in a structured, template-driven way.