Best practices and considerations for templates

This page describes some best practices and considerations for server prompt templates, including not-yet-supported features. Note that many of these features are only not available at the initial release of server prompt templates, so check the release notes for updates!

Best practices

Many of these best practices are also detailed in Manage your templates.

Version your templates

• Create template IDs appended with a version that uses semantic versioning (semver).

• Use Firebase Remote Config so that you can easily change the template and other values in your request.

Protect your template

• Lock your template before going to production. And avoid editing templates that are used in production.

  • Locking a template acts as protection against unintentional editing, but locking does not entirely block editing. A project member with the appropriate permissions can always unlock a template to edit it.

  • We recommend locking templates that are actively being used by code – especially production code.

• Write strong input validation for your input variables, which can help with the following:

  • Can help protect against prompt injection.
  • Can help ensure requests succeed and responses are as expected.

Not-yet-supported features

Many of these not-yet-supported features are only not available at the initial release, so check the release notes for updates!

Not-yet-supported capabilities of Firebase AI Logic

Server prompt templates do not yet support the following capabilities of Firebase AI Logic. Several of these capabilities are coming soon!

• Chat
• Using tools (including function calling and grounding with Google Search)
• Iterative editing of images (chat) with Gemini models (requires chat)
• Editing images with Imagen models
• Bidirectional streaming (Gemini Live API)
• Hybrid on-device
• Constraining the output to a list of enums
• Configuring thinking-related settings
• Configuring safety settings

Also note that if you use AI monitoring in the Firebase console, the template ID is not yet populated in any record.

Not-yet-supported common elements of Dotprompt

Server prompt templates will likely soon support the following common elements of Dotprompt:

• Using the JSON schema specification for input and output schemas in your frontmatter.

  • Note that this spec isn't supported in non-server template requests either. We currently only support the OpenAPI schema specification.

• Declaring a wildcard field definition using *.

• Using @key or generally iterating over the fields of an object input (which is the case where @key is relevant).

• Using @root, which lets you reference the root variable context regardless of current this.

Server prompt templates will likely not support the following common elements of Dotprompt:

• Using partials, which are reusable template snippets that can be included in other templates.

Other considerations

• By default, when you use the guided UI in the Firebase console, we provision the template in all available regions for Firebase AI Logic. If you use the Vertex AI Gemini API and if your use case requires location-based restrictions, you can specify the location for your template using the REST API.

• If you want to provide a server prompt template as a file (instead of using the Firebase console's guided UI), you can use the REST API. Note that server prompt templates do not support schema defined in app code and passed into the template.

• Even though the template is on the server, it cannot directly interact with your Firebase project's other server-side resources (like a database) except for Cloud Storage for Firebase URLs (which can be provided as input variables).