For Firebase AI Logic, the Firebase console provides a guided UI for you to specify the contents of a template.
Server prompt templates use a Dotprompt-based syntax and format. On this page, you can find detailed descriptions of the template format and syntax, along with examples for both Gemini and Imagen.
Here are the most important components for an example request to a Gemini model:
---
model: 'gemini-2.5-flash'
---
{{role "system"}}
All output must be a clearly structured invoice document.
Use a tabular or clearly delineated list format for line items.
{{role "user"}}
Create an example customer invoice for a customer named {{customerName}}.
The top section within the triple-dashes contains the model name as well as optionally any model configuration, input validation, or schema you want to send in the request. It's written as key-value pairs and is commonly called YAML frontmatter.
The body of the template contains the prompt. It can also optionally include system instructions and input values (using Handlebars syntax).
This page provides detailed descriptions of the template format and syntax, along with examples, for the following:
Gemini
All the examples in this section show templates that use gemini-2.5-flash, but
you can use any Gemini model supported by Firebase AI Logic
(except for Gemini Live models).
Hello world
Here's a minimal example of a server prompt template:
Configuration (frontmatter)
---
model: 'gemini-2.5-flash'
---
Prompt and (as applicable) system instructions
Write a story about a magic backpack.
Model configuration
Set a model configuration to control how the model generates a response, like max output tokens, temperature, top-K, and top-P.
Configuration (frontmatter)
---
model: 'gemini-2.5-flash'
config:
candidateCount: 1
temperature: 0.9
topP: 0.1
topK: 16
maxOutputTokens: 200
stopSequences: ["red"]
---
Prompt and (as applicable) system instructions
Write a story about a magic backpack.
System instructions
Set system instructions to steer the behavior of the model. You include them as part of the prompt:
Specify the system instructions using the
{{role "system"}}syntax.Specify the text prompt using the
{{role "user"}}syntax.
Configuration (frontmatter)
---
model: 'gemini-2.5-flash'
---
Prompt and (as applicable) system instructions
{{role "system"}}
All output must be a clearly structured invoice document.
Use a tabular or clearly delineated list format for line items.
{{role "user"}}
Create an example customer invoice for a customer.
Input variables
Some prompts are static, but you often need to include some data from the user as part of the prompt.
You can include dynamic input variables in the prompt using
Handlebars expressions, which
are contained within {{ }} tags in the format
of {{variableName}} or
{{object.propertyName}} (for example,
Hello, {{name}} from {{address.city}}).
Configuration (frontmatter)
---
model: 'gemini-2.5-flash'
---
Prompt and (as applicable) system instructions
Create an example customer invoice for a customer named {{customerName}}.
You can provide a default value in the template, but the value of an input variable is usually provided by the client as part of the request.
|
Click your Gemini API provider to view provider-specific content and code on this page. |
Swift
// ...
// Initialize the Gemini Developer API backend service
// Create a `TemplateGenerativeModel` instance
let model = FirebaseAI.firebaseAI(backend: .googleAI()).templateGenerativeModel()
let customerName = "Jane"
do {
let response = try await model.generateContent(
// Specify your template ID
templateID: "my-first-template-v1-0-0",
// Provide the values for any input variables required by your template.
inputs: [
"customerName": customerName
]
)
if let text = response.text {
print("Response Text: \(text)")
}
} catch {
print("An error occurred: \(error)")
}
print("\n")
Kotlin
// ...
// Initialize the Gemini Developer API backend service
// Create a `TemplateGenerativeModel` instance
val model = Firebase.googleAI.templateGenerativeModel()
val customerName = "Jane"
val response = model.generateContent(
// Specify your template ID
"my-first-template-v1-0-0",
// Provide the values for any input variables required by your template.
mapOf(
"customerName" to customerName
)
)
val text = response.text
println(text)
Java
// ...
// Initialize the Gemini Developer API backend service
// Create a `TemplateGenerativeModel` instance
TemplateGenerativeModel generativeModel = FirebaseAI.getInstance().templateGenerativeModel();
TemplateGenerativeModelFutures model = TemplateGenerativeModelFutures.from(generativeModel);
String customerName = "Jane";
Future response = model.generateContent(
// Specify your template ID
"my-first-template-v1-0-0",
// Provide the values for any input variables required by your template.
mapOf("customerName", customerName)
);
addCallback(response,
new FutureCallback() {
public void onSuccess(GenerateContentResponse result) {
System.out.println(result.getText());
}
public void onFailure(Throwable t) {
reportError(t);
}
}
executor);
Web
// ...
// Initialize the Gemini Developer API backend service
const ai = getAI(app, { backend: new GoogleAIBackend() });
// Create a `TemplateGenerativeModel` instance
const model = getTemplateGenerativeModel(ai);
const customerName = 'Jane';
const result = await model.generateContent(
// Specify your template ID
'my-first-template-v1-0-0',
// Provide the values for any input variables required by your template
{
customerName: customerName,
}
);
const response = response.result;
const text = response.text();
Dart
The Flutter plugin will support server prompt templates soon!
Unity
The Unity package will support server prompt templates soon!
Control flows (loops & conditionals)
To write more complex prompts, you can use conditional blocks (like #if ,
else, and #unless) and iteration (#each).
You can provide additional contextual information as variables with a special
@ prefix:
@first: true when iterating the first item of an#eachblock.@last: true when iterating the last item of an#eachblock.@index: gives the (zero-based) index position of the current element.
See the Handlebars documentation for information on all of the built-in logical helpers.
Configuration (frontmatter)
---
model: 'gemini-2.5-flash'
---
Prompt and (as applicable) system instructions
Create an example customer invoice for a customer named {{customerName}}.
Include entries for each of the following products
{{#each productNames}}
{{#if @first}}
Include line items for the following purchases
{{/if}}
- {{this}}
{{/each}}
{{#if isVipCustomer}}
Give the customer a 5% discount.
{{/if}}
Note that conditionals only accept a variable reference, not any type of expression, for example:
- The following works:
{{#if isVipCustomer}} ... {{/if}} - The following does not work:
{{#if customer.type == 'vip'}} ... {{/if}}
If the variable is a boolean, then the conditional works as you would expect. If the variable is not a boolean then the conditional is effectively an "is-not-null" check. This can be useful for handling optional inputs, for example:
{{#if customerName}}
Hello {{customerName}}
{{else}}
Hello Guest
{{/if}}
Input validation and schema
If you have data coming from the client, we strongly recommend using input schema to help protect against prompt injection as well as ensure that the data passed in the request matches your expectations.
You can provide default values in case the client doesn't provide a value.
The schema supports scalar types
string,integer,number,boolean, andobject. Objects, arrays, and enums are denoted by a parenthetical after the field name.All properties are considered required unless you denote it as optional with
?. When a property is marked as optional, it's also made nullable to provide more leniency for LLMs to return null instead of omitting a field.
Here's a basic example for providing input schema. You can find a more advanced schema just below.
Configuration (frontmatter)
---
model: 'gemini-2.5-flash'
input:
default:
isVipCustomer: false
schema:
customerName: string, the customers name # string, number, and boolean types are defined like this
productNames?(array, list of products to include in the invoice): string # optional fields are marked with a ?
isVipCustomer?: boolean, whether or not the customer is a VIP
---
Prompt and (as applicable) system instructions
Create an example customer invoice for a customer named {{customerName}}.
Include entries for each of the following products
{{#each productNames}}
{{#if @first}}
Include line items for the following purchases
{{/if}}
- {{this}}
{{/each}}
{{#if isVipCustomer}}
Give the customer a 5% discount.
{{/if}}
Configuration (frontmatter)
---
model: gemini-2.5-flash
input:
schema:
customer:
type: object
description: details about the customer
required: [lastName, address]
properties:
firstName:
type: string
minLength: 2
maxLength: 50
lastName:
type: string
minLength: 2
maxLength: 50
address:
type: object
description: customer's address
required: [street, city, country]
properties:
street:
type: string
city:
type: string
country:
type: string
isVip:
type: boolean
description: whether or not the customer is a VIP
purchases?:
type: array
description: list of things the customer purchased
items:
type: object
required: [itemName, sku, cost, quantity]
properties:
itemName:
type: string
sku:
type: string
minLength: 7
maxLength: 19
pattern: "^[0-9]{1,3}-[a-zA-Z]{3,10}-[0-9]{1,4}$"
category:
enum: ["ELECTRONICS", "APPAREL", "HOME", "MISC" ]
cost:
type: number
minimum: 0
quantity:
type: number
minimum: 1
maximum: 200
---
Prompt and (as applicable) system instructions
Create an example customer invoice for a customer.
The customer's address is
{{customer.firstName}} {{customer.lastName}}
{{customer.address.street}}
{{customer.address.city}}
{{customer.address.country}}
{{#each purchases}}
{{#if @first}}
Include line items for the following purchaes
{{/if}}
- {{quantity}} x {{itemName}} ({{sku}}) @ ${{cost}}, category {{category}}
{{/each}}
{{#if customer.isVip}}
Give the customer a 5% discount.
{{/if}}
Output schema
If you want the model to generate
structured JSON output,
then you can specify an output schema. By specifying format: json, you're
constraining the model to always return a JSON response that follows the
specified schema.
The schema supports scalar types
string,integer,number,boolean, andobject. Objects, arrays, and enums are denoted by a parenthetical after the field name.All properties are considered required unless you denote it as optional with
?. When a property is marked as optional, it's also made nullable to provide more leniency for LLMs to return null instead of omitting a field.
Here's a basic example for generating structured JSON output. You can find a more advanced schema just below.
Configuration (frontmatter)
---
model: gemini-2.5-flash
output:
format: json
schema:
invoiceId: string
invoiceFile(object, an invoice file):
url?: string
contents: string
mimeType: string
---
Prompt and (as applicable) system instructions
Create an example customer invoice.
Configuration (frontmatter)
---
model: gemini-2.5-flash
output:
format: json
schema:
title: string # string, number, and boolean types are defined like this
subtitle?: string # optional fields are marked with a ?
draft?: boolean, true when in draft state
status?(enum, approval status): [PENDING, APPROVED]
date: string, the date of publication for example '2024-04-09' # descriptions follow a comma
tags(array, relevant tags for article): string # arrays are denoted via parentheses
authors(array):
name: string
email?: string
metadata?(object): # objects are also denoted via parentheses
updatedAt?: string, ISO timestamp of last update
approvedBy?: integer, id of approver
extra?: any, arbitrary extra data
(*): string, wildcard field
---
Prompt and (as applicable) system instructions
Store all the specified information about the provided article.
Multimodal input
Multimodal prompts sent to a Gemini model can include multiple types of input, including files (like text along with images, PDFs, plain-text files, audio, and video).
Provide a file using its URL with the
{{media url}}syntax.Provide an inline file with the
{{media type="mime_type" data="contents"}}syntax.
Here's a basic example for providing multimodal input. You can find a more complex example just below.
Configuration (frontmatter)
---
model: 'gemini-2.5-flash'
---
Prompt and (as applicable) system instructions
Describe this image
{{media type="mimeType" data="imageData"}}
If your template has input that's base64-encoded data, here's how you can include it in the testing experience of the Firebase console:
{
"mimeType": "image/png",
"imageData": "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"
}
Configuration (frontmatter)
---
model: gemini-2.5-flash
input:
schema:
image_urls?(array, urls of external images): string
inline_images?(array, inline image data):
type: object
properties:
mime_type: string
contents: string # inline data must be base64-encoded
---
Prompt and (as applicable) system instructions
{{role "system"}}
Use the following image as the basis for comparisons
{{media url="http://example.com/reference_img.bmp"}}
{{role "user"}}
What do the following images have in common?
{{#each image_urls}}
{{media url="this"}}
{{/each}}
{{#each inline_images}}
{{media type="mime_type" data="contents"}}
{{/each}}
Imagen (image generation)
With the initial release, server prompt templates support generating images using Imagen models and a text-only prompt. Check back soon for more support, including editing images with Imagen (when using the Vertex AI Gemini API).
Basic
This example shows a basic template for generating images with Imagen, with input variables and input validation similar to Gemini.
Configuration (frontmatter)
---
model: 'imagen-4.0-generate-001'
input:
schema:
prompt: 'string'
---
Prompt and (as applicable) system instructions
Create an image containing {{prompt}}
Advanced
This example shows how to add model configuration and use more advanced features in the prompt, like input variables, input validation, and control flows similar to Gemini.
Configuration (frontmatter)
---
model: 'imagen-4.0-fast-generate-001'
config:
sampleCount: 1
aspectRatio: "16:9"
personGeneration: dont_allow
includeRaiReason: true
input:
schema:
style(enum, The style of image): [photo, sketch, painting]
subject: string, The object or animal or scenery to generate.
context?: string, Optional background or context description.
default:
style: photo
---
Prompt and (as applicable) system instructions
A {{style}} of {{subject}}{{#if context}}{{context}}{{/if}}.