[AI] Hybrid

FirebaseAI/Sources/AILog.swift

@@ -35,6 +35,7 @@ enum AILog {
     case generativeModelInitialized = 1000
     case unsupportedGeminiModel = 1001
     case invalidSchemaFormat = 1002
+    case unsupportedConfig = 1003
 
     // Imagen Model Configuration
     case unsupportedImagenModel = 1200

FirebaseAI/Sources/FirebaseAI.swift

@@ -52,8 +52,10 @@ public final class FirebaseAI: Sendable {
     )
     // Verify that the `FirebaseAI` instance is always configured with the production endpoint since
     // this is the public API surface for creating an instance.
-    assert(instance.apiConfig.service.endpoint == .firebaseProxyProd)
-    assert(instance.apiConfig.version == .v1beta)
+    if case let .cloud(config) = instance.apiConfig {
+      assert(config.service.endpoint == .firebaseProxyProd)
+      assert(config.version == .v1beta)
+    }
     return instance
   }
 
@@ -86,8 +88,9 @@ public final class FirebaseAI: Sendable {
                               systemInstruction: ModelContent? = nil,
                               requestOptions: RequestOptions = RequestOptions())
     -> GenerativeModel {
-    if !modelName.starts(with: GenerativeModel.geminiModelNamePrefix)
-      && !modelName.starts(with: GenerativeModel.gemmaModelNamePrefix) {
+    if case .cloud = apiConfig,
+       !modelName.starts(with: GenerativeModel.geminiModelNamePrefix)
+       && !modelName.starts(with: GenerativeModel.gemmaModelNamePrefix) {
       AILog.warning(code: .unsupportedGeminiModel, """
       Unsupported Gemini model "\(modelName)"; see \
       https://firebase.google.com/docs/vertex-ai/models for a list supported Gemini model names.
@@ -281,11 +284,16 @@ public final class FirebaseAI: Sendable {
       """)
     }
 
-    switch apiConfig.service {
-    case let .vertexAI(endpoint: _, location: location):
-      return vertexAIModelResourceName(modelName: modelName, location: location)
-    case .googleAI:
-      return developerModelResourceName(modelName: modelName)
+    switch apiConfig {
+    case let .cloud(config):
+      switch config.service {
+      case let .vertexAI(endpoint: _, location: location):
+        return vertexAIModelResourceName(modelName: modelName, location: location)
+      case .googleAI:
+        return developerModelResourceName(modelName: modelName)
+      }
+    case .onDevice:
+      return modelName
     }
   }
 
@@ -304,7 +312,11 @@ public final class FirebaseAI: Sendable {
   }
 
   private func developerModelResourceName(modelName: String) -> String {
-    switch apiConfig.service.endpoint {
+    guard case let .cloud(config) = apiConfig else {
+      fatalError("Developer API is only supported for cloud configurations.")
+    }
+
+    switch config.service.endpoint {
     case .firebaseProxyProd:
       return "projects/\(firebaseInfo.projectID)/models/\(modelName)"
     #if DEBUG

FirebaseAI/Sources/GenerateContentRequest.swift

@@ -74,7 +74,13 @@ extension GenerateContentRequest: GenerativeAIRequest {
   typealias Response = GenerateContentResponse
 
   func getURL() throws -> URL {
-    let modelURL = "\(apiConfig.service.endpoint.rawValue)/\(apiConfig.version.rawValue)/\(model)"
+    guard case let .cloud(config) = apiConfig else {
+      throw AILog.makeInternalError(
+        message: "URL generation not supported for on-device models",
+        code: .unsupportedConfig
+      )
+    }
+    let modelURL = "\(config.service.endpoint.rawValue)/\(config.version.rawValue)/\(model)"
     let urlString: String
     switch apiMethod {
     case .generateContent:

FirebaseAI/Sources/GenerativeModel.swift

@@ -221,6 +221,41 @@ public final class GenerativeModel: Sendable {
                               generationConfig: GenerationConfig?) async throws
     -> GenerateContentResponse {
     try content.throwIfError()
+
+    if case .onDevice = apiConfig {
+      #if canImport(FoundationModels)
+        if #available(iOS 26.0, macOS 26.0, *) {
+          let prompt = content.compactMap { modelContent in
+            modelContent.parts.compactMap { ($0 as? TextPart)?.text }.joined()
+          }.joined(separator: "\n")
+
+          let instructions = systemInstruction?.parts.compactMap { ($0 as? TextPart)?.text }
+            .joined()
+
+          // TODO: Map `tools` to FoundationModels tools when supported.
+          let session = LanguageModelSession(model: .default, instructions: instructions)
+          let response = try await session.respond(to: prompt)
+
+          return GenerateContentResponse(candidates: [Candidate(
+            content: ModelContent(role: "model", parts: [TextPart(response.content)]),
+            safetyRatings: [],
+            finishReason: .stop,
+            citationMetadata: nil
+          )], promptFeedback: nil, usageMetadata: nil)
+        } else {
+          throw AILog.makeInternalError(
+            message: "Foundation Models require iOS 26.0+",
+            code: .unsupportedConfig
+          )
+        }
+      #else
+        throw AILog.makeInternalError(
+          message: "Foundation Models not available",
+          code: .unsupportedConfig
+        )
+      #endif
+    }
+
     let response: GenerateContentResponse
     let generateContentRequest = GenerateContentRequest(
       model: modelResourceName,
@@ -389,36 +424,48 @@ public final class GenerativeModel: Sendable {
   /// - Returns: The results of running the model's tokenizer on the input; contains
   /// ``CountTokensResponse/totalTokens``.
   public func countTokens(_ content: [ModelContent]) async throws -> CountTokensResponse {
-    let requestContent = switch apiConfig.service {
-    case .vertexAI:
-      content
-    case .googleAI:
-      // The `role` defaults to "user" but is ignored in `countTokens`. However, it is erroneously
-      // erroneously counted towards the prompt and total token count when using the Developer API
-      // backend; set to `nil` to avoid token count discrepancies between `countTokens` and
-      // `generateContent` and the two backend APIs.
-      content.map { ModelContent(role: nil, parts: $0.parts) }
+    let requestContent: [ModelContent]
+    switch apiConfig {
+    case let .cloud(config):
+      switch config.service {
+      case .vertexAI:
+        requestContent = content
+      case .googleAI:
+        // The `role` defaults to "user" but is ignored in `countTokens`. However, it is erroneously
+        // erroneously counted towards the prompt and total token count when using the Developer API
+        // backend; set to `nil` to avoid token count discrepancies between `countTokens` and
+        // `generateContent` and the two backend APIs.
+        requestContent = content.map { ModelContent(role: nil, parts: $0.parts) }
+      }
+    case .onDevice:
+      requestContent = content
     }
 
     // When using the Developer API via the Firebase backend, the model name of the
     // `GenerateContentRequest` nested in the `CountTokensRequest` must be of the form
     // "models/model-name". This field is unaltered by the Firebase backend before forwarding the
     // request to the Generative Language backend, which expects the form "models/model-name".
-    let generateContentRequestModelResourceName = switch apiConfig.service {
-    case .vertexAI:
-      modelResourceName
-    case .googleAI(endpoint: .firebaseProxyProd):
-      "models/\(modelName)"
-    #if DEBUG
-      case .googleAI(endpoint: .firebaseProxyStaging):
-        "models/\(modelName)"
-      case .googleAI(endpoint: .googleAIBypassProxy):
-        modelResourceName
-      case .googleAI(endpoint: .vertexAIStagingBypassProxy):
-        fatalError(
-          "The Vertex AI staging endpoint does not support the Gemini Developer API (Google AI)."
-        )
-    #endif // DEBUG
+    let generateContentRequestModelResourceName: String
+    switch apiConfig {
+    case let .cloud(config):
+      switch config.service {
+      case .vertexAI:
+        generateContentRequestModelResourceName = modelResourceName
+      case .googleAI(endpoint: .firebaseProxyProd):
+        generateContentRequestModelResourceName = "models/\(modelName)"
+      #if DEBUG
+        case .googleAI(endpoint: .firebaseProxyStaging):
+          generateContentRequestModelResourceName = "models/\(modelName)"
+        case .googleAI(endpoint: .googleAIBypassProxy):
+          generateContentRequestModelResourceName = modelResourceName
+        case .googleAI(endpoint: .vertexAIStagingBypassProxy):
+          fatalError(
+            "The Vertex AI staging endpoint does not support the Gemini Developer API (Google AI)."
+          )
+      #endif // DEBUG
+      }
+    case .onDevice:
+      generateContentRequestModelResourceName = modelResourceName
     }
 
     let generateContentRequest = GenerateContentRequest(

FirebaseAI/Sources/TemplateGenerateContentRequest.swift

@@ -44,9 +44,16 @@ extension TemplateGenerateContentRequest: GenerativeAIRequest {
   typealias Response = GenerateContentResponse
 
   func getURL() throws -> URL {
+    guard case let .cloud(config) = apiConfig else {
+      throw AILog.makeInternalError(
+        message: "Templates not supported on-device",
+        code: .unsupportedConfig
+      )
+    }
+
     var urlString =
-      "\(apiConfig.service.endpoint.rawValue)/\(apiConfig.version.rawValue)/projects/\(projectID)"
-    if case let .vertexAI(_, location) = apiConfig.service {
+      "\(config.service.endpoint.rawValue)/\(config.version.rawValue)/projects/\(projectID)"
+    if case let .vertexAI(_, location) = config.service {
       urlString += "/locations/\(location)"
     }
 

FirebaseAI/Sources/TemplateImagenGenerationRequest.swift

@@ -41,9 +41,16 @@ struct TemplateImagenGenerationRequest<ImageType: ImagenImageRepresentable>: Sen
 @available(iOS 15.0, macOS 12.0, macCatalyst 15.0, tvOS 15.0, watchOS 8.0, *)
 extension TemplateImagenGenerationRequest: GenerativeAIRequest where ImageType: Decodable {
   func getURL() throws -> URL {
+    guard case let .cloud(config) = apiConfig else {
+      throw AILog.makeInternalError(
+        message: "Templates not supported on-device",
+        code: .unsupportedConfig
+      )
+    }
+
     var urlString =
-      "\(apiConfig.service.endpoint.rawValue)/\(apiConfig.version.rawValue)/projects/\(projectID)"
-    if case let .vertexAI(_, location) = apiConfig.service {
+      "\(config.service.endpoint.rawValue)/\(config.version.rawValue)/projects/\(projectID)"
+    if case let .vertexAI(_, location) = config.service {
       urlString += "/locations/\(location)"
     }
     urlString += "/templates/\(template):\(ImageAPIMethod.generateImages.rawValue)"

FirebaseAI/Sources/Types/Internal/APIConfig.swift

@@ -12,8 +12,17 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-/// Configuration for the generative AI backend API used by this SDK.
-struct APIConfig: Sendable, Hashable, Encodable {
+/// Configuration for the generative AI backend.
+enum APIConfig: Sendable, Hashable, Encodable {
+  /// Configuration for cloud-based inference (Vertex AI, Google AI).
+  case cloud(CloudConfig)
+
+  /// Configuration for on-device inference (Apple Foundation Models).
+  case onDevice
+}
+
+/// Configuration for cloud-based backend APIs.
+struct CloudConfig: Sendable, Hashable, Encodable {
   /// The service to use for generative AI.
   ///
   /// This controls which backend API is used by the SDK.
@@ -22,7 +31,7 @@ struct APIConfig: Sendable, Hashable, Encodable {
   /// The version of the selected API to use, e.g., "v1".
   let version: Version
 
-  /// Initializes an API configuration.
+  /// Initializes a cloud configuration.
   ///
   /// - Parameters:
   ///   - service: The API service to use for generative AI.
@@ -33,7 +42,7 @@ struct APIConfig: Sendable, Hashable, Encodable {
   }
 }
 
-extension APIConfig {
+extension CloudConfig {
   /// API services providing generative AI functionality.
   ///
   /// See [Vertex AI and Google AI
@@ -66,7 +75,7 @@ extension APIConfig {
   }
 }
 
-extension APIConfig.Service {
+extension CloudConfig.Service {
   /// Network addresses for generative AI API services.
   // TODO: maybe remove the https:// prefix and just add it as needed? websockets use these too.
   enum Endpoint: String, Encodable {
@@ -98,8 +107,8 @@ extension APIConfig.Service {
   }
 }
 
-extension APIConfig {
-  /// Versions of the configured API service (`APIConfig.Service`).
+extension CloudConfig {
+  /// Versions of the configured API service (`CloudConfig.Service`).
   enum Version: String, Encodable {
     /// The beta channel for version 1 of the API.
     case v1beta

FirebaseAI/Sources/Types/Internal/Imagen/ImagenGenerationRequest.swift

@@ -40,8 +40,14 @@ extension ImagenGenerationRequest: GenerativeAIRequest where ImageType: Decodabl
   typealias Response = ImagenGenerationResponse<ImageType>
 
   func getURL() throws -> URL {
+    guard case let .cloud(config) = apiConfig else {
+      throw AILog.makeInternalError(
+        message: "Imagen not supported on-device",
+        code: .unsupportedConfig
+      )
+    }
     let urlString =
-      "\(apiConfig.service.endpoint.rawValue)/\(apiConfig.version.rawValue)/\(model):predict"
+      "\(config.service.endpoint.rawValue)/\(config.version.rawValue)/\(model):predict"
     guard let url = URL(string: urlString) else {
       throw AILog.makeInternalError(message: "Malformed URL: \(urlString)", code: .malformedURL)
     }

FirebaseAI/Sources/Types/Internal/Live/LiveSessionService.swift

@@ -346,12 +346,25 @@ actor LiveSessionService {
   ///
   /// Will apply the required app check and auth headers, as the backend expects them.
   private nonisolated func createWebsocket() async throws -> AsyncWebSocket {
-    let host = apiConfig.service.endpoint.rawValue.withoutPrefix("https://")
-    let urlString = switch apiConfig.service {
+    guard case let .cloud(config) = apiConfig else {
+      throw NSError(
+        domain: "\(Constants.baseErrorDomain).\(Self.self)",
+        code: AILog.MessageCode.unsupportedConfig.rawValue,
+        userInfo: [
+          NSLocalizedDescriptionKey: "The Live API is not supported for on-device foundation models.",
+        ]
+      )
+    }
+
+    let host = config.service.endpoint.rawValue.withoutPrefix("https://")
+    let urlString: String
+    switch config.service {
     case let .vertexAI(_, location: location):
-      "wss://\(host)/ws/google.firebase.vertexai.\(apiConfig.version.rawValue).LlmBidiService/BidiGenerateContent/locations/\(location)"
+      urlString =
+        "wss://\(host)/ws/google.firebase.vertexai.\(config.version.rawValue).LlmBidiService/BidiGenerateContent/locations/\(location)"
     case .googleAI:
-      "wss://\(host)/ws/google.firebase.vertexai.\(apiConfig.version.rawValue).GenerativeService/BidiGenerateContent"
+      urlString =
+        "wss://\(host)/ws/google.firebase.vertexai.\(config.version.rawValue).GenerativeService/BidiGenerateContent"
     }
     guard let url = URL(string: urlString) else {
       throw NSError(

FirebaseAI/Sources/Types/Internal/Requests/CountTokensRequest.swift

@@ -30,8 +30,14 @@ extension CountTokensRequest: GenerativeAIRequest {
   var apiConfig: APIConfig { generateContentRequest.apiConfig }
 
   func getURL() throws -> URL {
-    let version = apiConfig.version.rawValue
-    let endpoint = apiConfig.service.endpoint.rawValue
+    guard case let .cloud(config) = apiConfig else {
+      throw AILog.makeInternalError(
+        message: "URL generation not supported for on-device models",
+        code: .unsupportedConfig
+      )
+    }
+    let version = config.version.rawValue
+    let endpoint = config.service.endpoint.rawValue
     let urlString = "\(endpoint)/\(version)/\(modelResourceName):countTokens"
     guard let url = URL(string: urlString) else {
       throw AILog.makeInternalError(message: "Malformed URL: \(urlString)", code: .malformedURL)
@@ -66,7 +72,14 @@ extension CountTokensRequest: Encodable {
   }
 
   func encode(to encoder: any Encoder) throws {
-    switch apiConfig.service {
+    guard case let .cloud(config) = apiConfig else {
+      throw AILog.makeInternalError(
+        message: "Encoding not supported for on-device models",
+        code: .unsupportedConfig
+      )
+    }
+
+    switch config.service {
     case .vertexAI:
       try encodeForVertexAI(to: encoder)
     case .googleAI:

FirebaseAI/Sources/Types/Public/Backend.swift

@@ -25,20 +25,20 @@ public struct Backend {
   ///     for a list of supported locations.
   public static func vertexAI(location: String = "us-central1") -> Backend {
     return Backend(
-      apiConfig: APIConfig(
+      apiConfig: .cloud(CloudConfig(
         service: .vertexAI(endpoint: .firebaseProxyProd, location: location),
         version: .v1beta
-      )
+      ))
     )
   }
 
   /// Initializes a `Backend` configured for the Google Developer API.
   public static func googleAI() -> Backend {
     return Backend(
-      apiConfig: APIConfig(
+      apiConfig: .cloud(CloudConfig(
         service: .googleAI(endpoint: .firebaseProxyProd),
         version: .v1beta
-      )
+      ))
     )
   }