@firebase/ai 2.4.0 → 2.5.0-20251028194003

package/dist/index.node.mjs

@@ -4,7 +4,7 @@ import { FirebaseError, Deferred, getModularInstance } from '@firebase/util';
 import { Logger } from '@firebase/logger';
 
 var name = "@firebase/ai";
-var version = "2.4.0";
+var version = "2.5.0-20251028194003";
 
 /**
  * @license
@@ -323,6 +323,15 @@ const InferenceMode = {
     'ONLY_IN_CLOUD': 'only_in_cloud',
     'PREFER_IN_CLOUD': 'prefer_in_cloud'
 };
+/**
+ * Indicates whether inference happened on-device or in-cloud.
+ *
+ * @beta
+ */
+const InferenceSource = {
+    'ON_DEVICE': 'on_device',
+    'IN_CLOUD': 'in_cloud'
+};
 /**
  * Represents the result of the code execution.
  *
@@ -1269,7 +1278,7 @@ function hasValidCandidates(response) {
  * Creates an EnhancedGenerateContentResponse object that has helper functions and
  * other modifications that improve usability.
  */
-function createEnhancedContentResponse(response) {
+function createEnhancedContentResponse(response, inferenceSource = InferenceSource.IN_CLOUD) {
     /**
      * The Vertex AI backend omits default values.
      * This causes the `index` property to be omitted from the first candidate in the
@@ -1280,6 +1289,7 @@ function createEnhancedContentResponse(response) {
         response.candidates[0].index = 0;
     }
     const responseWithHelpers = addHelpers(response);
+    responseWithHelpers.inferenceSource = inferenceSource;
     return responseWithHelpers;
 }
 /**
@@ -1656,16 +1666,16 @@ const responseLineRE = /^data\: (.*)(?:\n\n|\r\r|\r\n\r\n)/;
  *
  * @param response - Response from a fetch call
  */
-function processStream(response, apiSettings) {
+function processStream(response, apiSettings, inferenceSource) {
     const inputStream = response.body.pipeThrough(new TextDecoderStream('utf8', { fatal: true }));
     const responseStream = getResponseStream(inputStream);
     const [stream1, stream2] = responseStream.tee();
     return {
-        stream: generateResponseSequence(stream1, apiSettings),
-        response: getResponsePromise(stream2, apiSettings)
+        stream: generateResponseSequence(stream1, apiSettings, inferenceSource),
+        response: getResponsePromise(stream2, apiSettings, inferenceSource)
     };
 }
-async function getResponsePromise(stream, apiSettings) {
+async function getResponsePromise(stream, apiSettings, inferenceSource) {
     const allResponses = [];
     const reader = stream.getReader();
     while (true) {
@@ -1675,12 +1685,12 @@ async function getResponsePromise(stream, apiSettings) {
             if (apiSettings.backend.backendType === BackendType.GOOGLE_AI) {
                 generateContentResponse = mapGenerateContentResponse(generateContentResponse);
             }
-            return createEnhancedContentResponse(generateContentResponse);
+            return createEnhancedContentResponse(generateContentResponse, inferenceSource);
         }
         allResponses.push(value);
     }
 }
-async function* generateResponseSequence(stream, apiSettings) {
+async function* generateResponseSequence(stream, apiSettings, inferenceSource) {
     const reader = stream.getReader();
     while (true) {
         const { value, done } = await reader.read();
@@ -1689,10 +1699,10 @@ async function* generateResponseSequence(stream, apiSettings) {
         }
         let enhancedResponse;
         if (apiSettings.backend.backendType === BackendType.GOOGLE_AI) {
-            enhancedResponse = createEnhancedContentResponse(mapGenerateContentResponse(value));
+            enhancedResponse = createEnhancedContentResponse(mapGenerateContentResponse(value), inferenceSource);
         }
         else {
-            enhancedResponse = createEnhancedContentResponse(value);
+            enhancedResponse = createEnhancedContentResponse(value, inferenceSource);
         }
         const firstCandidate = enhancedResponse.candidates?.[0];
         // Don't yield a response with no useful data for the developer.
@@ -1862,31 +1872,52 @@ const errorsCausingFallback = [
  */
 async function callCloudOrDevice(request, chromeAdapter, onDeviceCall, inCloudCall) {
     if (!chromeAdapter) {
-        return inCloudCall();
+        return {
+            response: await inCloudCall(),
+            inferenceSource: InferenceSource.IN_CLOUD
+        };
     }
     switch (chromeAdapter.mode) {
         case InferenceMode.ONLY_ON_DEVICE:
             if (await chromeAdapter.isAvailable(request)) {
-                return onDeviceCall();
+                return {
+                    response: await onDeviceCall(),
+                    inferenceSource: InferenceSource.ON_DEVICE
+                };
             }
             throw new AIError(AIErrorCode.UNSUPPORTED, 'Inference mode is ONLY_ON_DEVICE, but an on-device model is not available.');
         case InferenceMode.ONLY_IN_CLOUD:
-            return inCloudCall();
+            return {
+                response: await inCloudCall(),
+                inferenceSource: InferenceSource.IN_CLOUD
+            };
         case InferenceMode.PREFER_IN_CLOUD:
             try {
-                return await inCloudCall();
+                return {
+                    response: await inCloudCall(),
+                    inferenceSource: InferenceSource.IN_CLOUD
+                };
             }
             catch (e) {
                 if (e instanceof AIError && errorsCausingFallback.includes(e.code)) {
-                    return onDeviceCall();
+                    return {
+                        response: await onDeviceCall(),
+                        inferenceSource: InferenceSource.ON_DEVICE
+                    };
                 }
                 throw e;
             }
         case InferenceMode.PREFER_ON_DEVICE:
             if (await chromeAdapter.isAvailable(request)) {
-                return onDeviceCall();
+                return {
+                    response: await onDeviceCall(),
+                    inferenceSource: InferenceSource.ON_DEVICE
+                };
             }
-            return inCloudCall();
+            return {
+                response: await inCloudCall(),
+                inferenceSource: InferenceSource.IN_CLOUD
+            };
         default:
             throw new AIError(AIErrorCode.ERROR, `Unexpected infererence mode: ${chromeAdapter.mode}`);
     }
@@ -1916,8 +1947,8 @@ async function generateContentStreamOnCloud(apiSettings, model, params, requestO
     /* stream */ true, JSON.stringify(params), requestOptions);
 }
 async function generateContentStream(apiSettings, model, params, chromeAdapter, requestOptions) {
-    const response = await callCloudOrDevice(params, chromeAdapter, () => chromeAdapter.generateContentStream(params), () => generateContentStreamOnCloud(apiSettings, model, params, requestOptions));
-    return processStream(response, apiSettings); // TODO: Map streaming responses
+    const callResult = await callCloudOrDevice(params, chromeAdapter, () => chromeAdapter.generateContentStream(params), () => generateContentStreamOnCloud(apiSettings, model, params, requestOptions));
+    return processStream(callResult.response, apiSettings); // TODO: Map streaming responses
 }
 async function generateContentOnCloud(apiSettings, model, params, requestOptions) {
     if (apiSettings.backend.backendType === BackendType.GOOGLE_AI) {
@@ -1927,9 +1958,9 @@ async function generateContentOnCloud(apiSettings, model, params, requestOptions
     /* stream */ false, JSON.stringify(params), requestOptions);
 }
 async function generateContent(apiSettings, model, params, chromeAdapter, requestOptions) {
-    const response = await callCloudOrDevice(params, chromeAdapter, () => chromeAdapter.generateContent(params), () => generateContentOnCloud(apiSettings, model, params, requestOptions));
-    const generateContentResponse = await processGenerateContentResponse(response, apiSettings);
-    const enhancedResponse = createEnhancedContentResponse(generateContentResponse);
+    const callResult = await callCloudOrDevice(params, chromeAdapter, () => chromeAdapter.generateContent(params), () => generateContentOnCloud(apiSettings, model, params, requestOptions));
+    const generateContentResponse = await processGenerateContentResponse(callResult.response, apiSettings);
+    const enhancedResponse = createEnhancedContentResponse(generateContentResponse, callResult.inferenceSource);
     return {
         response: enhancedResponse
     };
@@ -2503,75 +2534,104 @@ class LiveSession {
         this.webSocketHandler.send(JSON.stringify(message));
     }
     /**
-     * Sends realtime input to the server.
+     * Sends text to the server in realtime.
      *
-     * @param mediaChunks - The media chunks to send.
+     * @example
+     * ```javascript
+     * liveSession.sendTextRealtime("Hello, how are you?");
+     * ```
+     *
+     * @param text - The text data to send.
      * @throws If this session has been closed.
      *
      * @beta
      */
-    async sendMediaChunks(mediaChunks) {
+    async sendTextRealtime(text) {
         if (this.isClosed) {
             throw new AIError(AIErrorCode.REQUEST_ERROR, 'This LiveSession has been closed and cannot be used.');
         }
-        // The backend does not support sending more than one mediaChunk in one message.
-        // Work around this limitation by sending mediaChunks in separate messages.
-        mediaChunks.forEach(mediaChunk => {
-            const message = {
-                realtimeInput: { mediaChunks: [mediaChunk] }
-            };
-            this.webSocketHandler.send(JSON.stringify(message));
-        });
+        const message = {
+            realtimeInput: {
+                text
+            }
+        };
+        this.webSocketHandler.send(JSON.stringify(message));
     }
     /**
-     * Sends function responses to the server.
+     * Sends audio data to the server in realtime.
      *
-     * @param functionResponses - The function responses to send.
+     * @remarks The server requires that the audio data is base64-encoded 16-bit PCM at 16kHz
+     * little-endian.
+     *
+     * @example
+     * ```javascript
+     * // const pcmData = ... base64-encoded 16-bit PCM at 16kHz little-endian.
+     * const blob = { mimeType: "audio/pcm", data: pcmData };
+     * liveSession.sendAudioRealtime(blob);
+     * ```
+     *
+     * @param blob - The base64-encoded PCM data to send to the server in realtime.
      * @throws If this session has been closed.
      *
      * @beta
      */
-    async sendFunctionResponses(functionResponses) {
+    async sendAudioRealtime(blob) {
         if (this.isClosed) {
             throw new AIError(AIErrorCode.REQUEST_ERROR, 'This LiveSession has been closed and cannot be used.');
         }
         const message = {
-            toolResponse: {
-                functionResponses
+            realtimeInput: {
+                audio: blob
             }
         };
         this.webSocketHandler.send(JSON.stringify(message));
     }
     /**
-     * Sends a stream of {@link GenerativeContentBlob}.
+     * Sends video data to the server in realtime.
      *
-     * @param mediaChunkStream - The stream of {@link GenerativeContentBlob} to send.
+     * @remarks The server requires that the video is sent as individual video frames at 1 FPS. It
+     * is recommended to set `mimeType` to `image/jpeg`.
+     *
+     * @example
+     * ```javascript
+     * // const videoFrame = ... base64-encoded JPEG data
+     * const blob = { mimeType: "image/jpeg", data: videoFrame };
+     * liveSession.sendVideoRealtime(blob);
+     * ```
+     * @param blob - The base64-encoded video data to send to the server in realtime.
      * @throws If this session has been closed.
      *
      * @beta
      */
-    async sendMediaStream(mediaChunkStream) {
+    async sendVideoRealtime(blob) {
         if (this.isClosed) {
             throw new AIError(AIErrorCode.REQUEST_ERROR, 'This LiveSession has been closed and cannot be used.');
         }
-        const reader = mediaChunkStream.getReader();
-        while (true) {
-            try {
-                const { done, value } = await reader.read();
-                if (done) {
-                    break;
-                }
-                else if (!value) {
-                    throw new Error('Missing chunk in reader, but reader is not done.');
-                }
-                await this.sendMediaChunks([value]);
-            }
-            catch (e) {
-                // Re-throw any errors that occur during stream consumption or sending.
-                const message = e instanceof Error ? e.message : 'Error processing media stream.';
-                throw new AIError(AIErrorCode.REQUEST_ERROR, message);
+        const message = {
+            realtimeInput: {
+                video: blob
             }
+        };
+        this.webSocketHandler.send(JSON.stringify(message));
+    }
+    /**
+     * Sends function responses to the server.
+     *
+     * @param functionResponses - The function responses to send.
+     * @throws If this session has been closed.
+     *
+     * @beta
+     */
+    async sendFunctionResponses(functionResponses) {
+        if (this.isClosed) {
+            throw new AIError(AIErrorCode.REQUEST_ERROR, 'This LiveSession has been closed and cannot be used.');
         }
+        const message = {
+            toolResponse: {
+                functionResponses
+            }
+        };
+        this.webSocketHandler.send(JSON.stringify(message));
     }
     /**
      * Yields messages received from the server.
@@ -2629,6 +2689,62 @@ class LiveSession {
             await this.webSocketHandler.close(1000, 'Client closed session.');
         }
     }
+    /**
+     * Sends realtime input to the server.
+     *
+     * @deprecated Use `sendTextRealtime()`, `sendAudioRealtime()`, and `sendVideoRealtime()` instead.
+     *
+     * @param mediaChunks - The media chunks to send.
+     * @throws If this session has been closed.
+     *
+     * @beta
+     */
+    async sendMediaChunks(mediaChunks) {
+        if (this.isClosed) {
+            throw new AIError(AIErrorCode.REQUEST_ERROR, 'This LiveSession has been closed and cannot be used.');
+        }
+        // The backend does not support sending more than one mediaChunk in one message.
+        // Work around this limitation by sending mediaChunks in separate messages.
+        mediaChunks.forEach(mediaChunk => {
+            const message = {
+                realtimeInput: { mediaChunks: [mediaChunk] }
+            };
+            this.webSocketHandler.send(JSON.stringify(message));
+        });
+    }
+    /**
+     * @deprecated Use `sendTextRealtime()`, `sendAudioRealtime()`, and `sendVideoRealtime()` instead.
+     *
+     * Sends a stream of {@link GenerativeContentBlob}.
+     *
+     * @param mediaChunkStream - The stream of {@link GenerativeContentBlob} to send.
+     * @throws If this session has been closed.
+     *
+     * @beta
+     */
+    async sendMediaStream(mediaChunkStream) {
+        if (this.isClosed) {
+            throw new AIError(AIErrorCode.REQUEST_ERROR, 'This LiveSession has been closed and cannot be used.');
+        }
+        const reader = mediaChunkStream.getReader();
+        while (true) {
+            try {
+                const { done, value } = await reader.read();
+                if (done) {
+                    break;
+                }
+                else if (!value) {
+                    throw new Error('Missing chunk in reader, but reader is not done.');
+                }
+                await this.sendMediaChunks([value]);
+            }
+            catch (e) {
+                // Re-throw any errors that occur during stream consumption or sending.
+                const message = e instanceof Error ? e.message : 'Error processing media stream.';
+                throw new AIError(AIErrorCode.REQUEST_ERROR, message);
+            }
+        }
+    }
 }
 
 /**
@@ -2689,13 +2805,18 @@ class LiveGenerativeModel extends AIModel {
         else {
             fullModelPath = `projects/${this._apiSettings.project}/locations/${this._apiSettings.location}/${this.model}`;
         }
+        // inputAudioTranscription and outputAudioTranscription are on the generation config in the public API,
+        // but the backend expects them to be in the `setup` message.
+        const { inputAudioTranscription, outputAudioTranscription, ...generationConfig } = this.generationConfig;
         const setupMessage = {
             setup: {
                 model: fullModelPath,
-                generationConfig: this.generationConfig,
+                generationConfig,
                 tools: this.tools,
                 toolConfig: this.toolConfig,
-                systemInstruction: this.systemInstruction
+                systemInstruction: this.systemInstruction,
+                inputAudioTranscription,
+                outputAudioTranscription
             }
         };
         try {
@@ -3401,7 +3522,7 @@ class AudioConversationRunner {
                 mimeType: 'audio/pcm',
                 data: base64
             };
-            void this.liveSession.sendMediaChunks([chunk]);
+            void this.liveSession.sendAudioRealtime(chunk);
         };
     }
     /**
@@ -3796,5 +3917,5 @@ function registerAI() {
 }
 registerAI();
 
-export { AIError, AIErrorCode, AIModel, AnyOfSchema, ArraySchema, Backend, BackendType, BlockReason, BooleanSchema, ChatSession, FinishReason, FunctionCallingMode, GenerativeModel, GoogleAIBackend, HarmBlockMethod, HarmBlockThreshold, HarmCategory, HarmProbability, HarmSeverity, ImagenAspectRatio, ImagenImageFormat, ImagenModel, ImagenPersonFilterLevel, ImagenSafetyFilterLevel, InferenceMode, IntegerSchema, Language, LiveGenerativeModel, LiveResponseType, LiveSession, Modality, NumberSchema, ObjectSchema, Outcome, POSSIBLE_ROLES, ResponseModality, Schema, SchemaType, StringSchema, URLRetrievalStatus, VertexAIBackend, getAI, getGenerativeModel, getImagenModel, getLiveGenerativeModel, startAudioConversation };
+export { AIError, AIErrorCode, AIModel, AnyOfSchema, ArraySchema, Backend, BackendType, BlockReason, BooleanSchema, ChatSession, FinishReason, FunctionCallingMode, GenerativeModel, GoogleAIBackend, HarmBlockMethod, HarmBlockThreshold, HarmCategory, HarmProbability, HarmSeverity, ImagenAspectRatio, ImagenImageFormat, ImagenModel, ImagenPersonFilterLevel, ImagenSafetyFilterLevel, InferenceMode, InferenceSource, IntegerSchema, Language, LiveGenerativeModel, LiveResponseType, LiveSession, Modality, NumberSchema, ObjectSchema, Outcome, POSSIBLE_ROLES, ResponseModality, Schema, SchemaType, StringSchema, URLRetrievalStatus, VertexAIBackend, getAI, getGenerativeModel, getImagenModel, getLiveGenerativeModel, startAudioConversation };
 //# sourceMappingURL=index.node.mjs.map