prompt-api-polyfill 0.1.0 → 0.3.0

package/prompt-api-polyfill.js

@@ -1,548 +1,586 @@
 /**
  * Polyfill for the Prompt API (`LanguageModel`)
- * Backend: Firebase AI Logic
+ * Backends:
+ * - Firebase AI Logic (via `firebase/ai`)
+ * - Google Gemini API (via `@google/generative-ai`)
+ * - OpenAI API (via `openai`)
+ *
  * Spec: https://github.com/webmachinelearning/prompt-api/blob/main/README.md
  *
- * * Instructions:
+ * Instructions:
  * 1. Include this script in your HTML type="module".
- * 2. Define window.FIREBASE_CONFIG with your Firebase configuration object BEFORE importing this.
+ * 2. Configure the backend:
+ *    - For Firebase: Define `window.FIREBASE_CONFIG`.
+ *    - For Gemini: Define `window.GEMINI_CONFIG`.
+ *    - For OpenAI: Define `window.OPENAI_CONFIG`.
  */
 
-import { initializeApp } from 'https://esm.run/firebase/app';
-import {
-  getAI,
-  getGenerativeModel,
-  GoogleAIBackend,
-  InferenceMode,
-} from 'https://esm.run/firebase/ai';
-
-import './async-iterator-polyfill.js'; // Still needed for Safari 26.2.
+import './async-iterator-polyfill.js';
 import MultimodalConverter from './multimodal-converter.js';
 import { convertJsonSchemaToVertexSchema } from './json-schema-converter.js';
 
-(() => {
-  if ('LanguageModel' in window) {
-    return;
+// --- Helper to convert initial History ---
+async function convertToHistory(prompts) {
+  const history = [];
+  for (const p of prompts) {
+    const role = p.role === 'assistant' ? 'model' : 'user';
+    let parts = [];
+
+    if (Array.isArray(p.content)) {
+      // Mixed content
+      for (const item of p.content) {
+        if (item.type === 'text') {
+          parts.push({ text: item.value || item.text || '' });
+        } else {
+          const part = await MultimodalConverter.convert(item.type, item.value);
+          parts.push(part);
+        }
+      }
+    } else {
+      // Simple string
+      parts.push({ text: p.content });
+    }
+    history.push({ role, parts });
   }
+  return history;
+}
 
-  const firebaseConfig = window.FIREBASE_CONFIG;
-  if (!firebaseConfig) {
-    console.error(
-      'Firebase Prompt API Polyfill: Missing configuration. Please set window.FIREBASE_CONFIG.'
-    );
-    return;
+/**
+ * Main LanguageModel Class
+ */
+export class LanguageModel extends EventTarget {
+  #backend;
+  #model;
+  #history;
+  #options;
+  #inCloudParams;
+  #destroyed;
+  #inputUsage;
+  #topK;
+  #temperature;
+  #onquotaoverflow;
+
+  constructor(backend, model, initialHistory, options = {}, inCloudParams) {
+    super();
+    this.#backend = backend;
+    this.#model = model;
+    this.#history = initialHistory || [];
+    this.#options = options;
+    this.#inCloudParams = inCloudParams;
+    this.#destroyed = false;
+    this.#inputUsage = 0;
+
+    this.#topK = options.topK;
+    this.#temperature = options.temperature;
   }
 
-  // Initialize Firebase
-  const app = initializeApp(firebaseConfig);
-  const ai = getAI(app, { backend: new GoogleAIBackend() });
-  const MODEL_NAME = firebaseConfig.modelName || 'gemini-2.5-flash-lite';
-
-  // Helper to convert initial History
-  async function convertToFirebaseHistory(prompts) {
-    const history = [];
-    for (const p of prompts) {
-      const role = p.role === 'assistant' ? 'model' : 'user';
-      let parts = [];
-
-      if (Array.isArray(p.content)) {
-        // Mixed content
-        for (const item of p.content) {
-          if (item.type === 'text') {
-            parts.push({ text: item.value || item.text || '' });
-          } else {
-            const part = await MultimodalConverter.convert(
-              item.type,
-              item.value
-            );
-            parts.push(part);
-          }
-        }
-      } else {
-        // Simple string
-        parts.push({ text: p.content });
-      }
-      history.push({ role, parts });
-    }
-    return history;
+  get inputUsage() {
+    return this.#inputUsage;
+  }
+  get inputQuota() {
+    return 1000000;
+  }
+  get topK() {
+    return this.#topK;
+  }
+  get temperature() {
+    return this.#temperature;
   }
 
-  /**
-   * Main LanguageModel Class
-   */
-  class LanguageModel extends EventTarget {
-    #model;
-    #history;
-    #options;
-    #inCloudParams;
-    #destroyed;
-    #inputUsage;
-    #topK;
-    #temperature;
-    #onquotaoverflow;
-
-    constructor(model, initialHistory, options = {}, inCloudParams) {
-      super();
-      this.#model = model;
-      this.#history = initialHistory || [];
-      this.#options = options;
-      this.#inCloudParams = inCloudParams;
-      this.#destroyed = false;
-      this.#inputUsage = 0;
-
-      this.#topK = options.topK;
-      this.#temperature = options.temperature;
-    }
+  get onquotaoverflow() {
+    return this.#onquotaoverflow;
+  }
 
-    get inputUsage() {
-      return this.#inputUsage;
+  set onquotaoverflow(handler) {
+    if (this.#onquotaoverflow) {
+      this.removeEventListener('quotaoverflow', this.#onquotaoverflow);
     }
-    get inputQuota() {
-      return 1000000;
+    this.#onquotaoverflow = handler;
+    if (typeof handler === 'function') {
+      this.addEventListener('quotaoverflow', handler);
     }
-    get topK() {
-      return this.#topK;
+  }
+
+  static async availability(options = {}) {
+    await LanguageModel.#validateOptions(options);
+    const backendClass = await LanguageModel.#getBackendClass();
+    return backendClass.availability(options);
+  }
+
+  static #backends = [
+    {
+      config: 'FIREBASE_CONFIG',
+      path: './backends/firebase.js',
+    },
+    {
+      config: 'GEMINI_CONFIG',
+      path: './backends/gemini.js',
+    },
+    {
+      config: 'OPENAI_CONFIG',
+      path: './backends/openai.js',
+    },
+    {
+      config: 'TRANSFORMERS_CONFIG',
+      path: './backends/transformers.js',
+    },
+  ];
+
+  static #getBackendInfo() {
+    for (const b of LanguageModel.#backends) {
+      const config = window[b.config];
+      if (config && config.apiKey) {
+        return { ...b, configValue: config };
+      }
+    }
+    throw new DOMException(
+      'Prompt API Polyfill: No backend configuration found. Please set window.FIREBASE_CONFIG, window.GEMINI_CONFIG, or window.OPENAI_CONFIG.',
+      'NotSupportedError'
+    );
+  }
+
+  static async #getBackendClass() {
+    const info = LanguageModel.#getBackendInfo();
+    return (await import(/* @vite-ignore */ info.path)).default;
+  }
+
+  static async #validateOptions(options = {}) {
+    const { maxTemperature, maxTopK } = await LanguageModel.params();
+
+    const hasTemperature = Object.prototype.hasOwnProperty.call(
+      options,
+      'temperature'
+    );
+    const hasTopK = Object.prototype.hasOwnProperty.call(options, 'topK');
+
+    if (hasTemperature !== hasTopK) {
+      throw new DOMException(
+        'Initializing a new session must either specify both topK and temperature, or neither of them.',
+        'NotSupportedError'
+      );
     }
-    get temperature() {
-      return this.#temperature;
+
+    // If neither temperature nor topK are provided, nothing to validate.
+    if (!hasTemperature && !hasTopK) {
+      return;
     }
 
-    get onquotaoverflow() {
-      return this.#onquotaoverflow;
+    const { temperature, topK } = options;
+
+    if (
+      typeof temperature !== 'number' ||
+      Number.isNaN(temperature) ||
+      typeof topK !== 'number' ||
+      Number.isNaN(topK)
+    ) {
+      throw new DOMException(
+        'The provided temperature and topK must be numbers.',
+        'NotSupportedError'
+      );
     }
 
-    set onquotaoverflow(handler) {
-      if (this.#onquotaoverflow)
-        this.removeEventListener('quotaoverflow', this.#onquotaoverflow);
-      this.#onquotaoverflow = handler;
-      if (typeof handler === 'function')
-        this.addEventListener('quotaoverflow', handler);
+    if (temperature < 0 || temperature > maxTemperature || topK > maxTopK) {
+      throw new DOMException(
+        'The provided temperature or topK is outside the supported range.',
+        'NotSupportedError'
+      );
     }
+  }
 
-    static async availability(options = {}) {
-      await LanguageModel.#validateOptions(options);
-      return 'available';
+  static async params() {
+    return {
+      // Values from https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/2-5-flash-lite#:~:text=%2C%20audio/webm-,Parameter%20defaults,-tune.
+      defaultTemperature: 1.0,
+      defaultTopK: 64,
+      maxTemperature: 2.0,
+      maxTopK: 64, // Fixed
+    };
+  }
+
+  static async create(options = {}) {
+    const availability = await LanguageModel.availability(options);
+    if (availability === 'downloadable' || availability === 'downloading') {
+      throw new DOMException(
+        'Requires a user gesture when availability is "downloading" or "downloadable".',
+        'NotAllowedError'
+      );
     }
 
-    static async #validateOptions(options = {}) {
-      const { maxTemperature, maxTopK } = await LanguageModel.params();
+    // --- Backend Selection Logic ---
+    const info = LanguageModel.#getBackendInfo();
+
+    const BackendClass = await LanguageModel.#getBackendClass();
+    const backend = new BackendClass(info.configValue);
+
+    const defaults = {
+      temperature: 1.0,
+      topK: 3,
+    };
 
-      const hasTemperature = Object.prototype.hasOwnProperty.call(
-        options,
-        'temperature'
+    const resolvedOptions = { ...defaults, ...options };
+
+    const inCloudParams = {
+      model: backend.modelName,
+      generationConfig: {
+        temperature: resolvedOptions.temperature,
+        topK: resolvedOptions.topK,
+      },
+    };
+
+    let initialHistory = [];
+
+    if (
+      resolvedOptions.initialPrompts &&
+      Array.isArray(resolvedOptions.initialPrompts)
+    ) {
+      const systemPrompts = resolvedOptions.initialPrompts.filter(
+        (p) => p.role === 'system'
+      );
+      const conversationPrompts = resolvedOptions.initialPrompts.filter(
+        (p) => p.role !== 'system'
       );
-      const hasTopK = Object.prototype.hasOwnProperty.call(options, 'topK');
 
-      if (hasTemperature !== hasTopK) {
-        throw new DOMException(
-          'Initializing a new session must either specify both topK and temperature, or neither of them.',
-          'NotSupportedError'
-        );
+      if (systemPrompts.length > 0) {
+        inCloudParams.systemInstruction = systemPrompts
+          .map((p) => p.content)
+          .join('\n');
       }
+      // Await the conversion of history items (in case of images in history)
+      initialHistory = await convertToHistory(conversationPrompts);
+    }
 
-      // If neither temperature nor topK are provided, nothing to validate.
-      if (!hasTemperature && !hasTopK) {
-        return;
-      }
+    const model = backend.createSession(resolvedOptions, inCloudParams);
+
+    // If a monitor callback is provided, simulate simple downloadprogress events
+    if (typeof resolvedOptions.monitor === 'function') {
+      const monitorTarget = new EventTarget();
 
-      const { temperature, topK } = options;
-
-      if (
-        typeof temperature !== 'number' ||
-        Number.isNaN(temperature) ||
-        typeof topK !== 'number' ||
-        Number.isNaN(topK)
-      ) {
-        throw new DOMException(
-          'The provided temperature and topK must be numbers.',
-          'NotSupportedError'
-        );
+      try {
+        resolvedOptions.monitor(monitorTarget);
+      } catch (e) {
+        console.error('Error in monitor callback:', e);
       }
 
-      if (temperature < 0 || temperature > maxTemperature || topK > maxTopK) {
-        throw new DOMException(
-          'The provided temperature or topK is outside the supported range.',
-          'NotSupportedError'
-        );
+      try {
+        const startEvent = new ProgressEvent('downloadprogress', {
+          loaded: 0,
+          total: 1,
+        });
+        const endEvent = new ProgressEvent('downloadprogress', {
+          loaded: 1,
+          total: 1,
+        });
+        monitorTarget.dispatchEvent(startEvent);
+        monitorTarget.dispatchEvent(endEvent);
+      } catch (e) {
+        console.error('Error dispatching downloadprogress events:', e);
       }
     }
 
-    static async params() {
-      return {
-        // Values from https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/2-5-flash-lite#:~:text=%2C%20audio/webm-,Parameter%20defaults,-tune.
-        defaultTemperature: 1.0,
-        defaultTopK: 64,
-        maxTemperature: 2.0,
-        maxTopK: 64, // Fixed
-      };
+    return new LanguageModel(
+      backend,
+      model,
+      initialHistory,
+      resolvedOptions,
+      inCloudParams
+    );
+  }
+
+  // Instance Methods
+
+  async clone(options = {}) {
+    if (this.#destroyed) {
+      throw new DOMException('Session is destroyed', 'InvalidStateError');
     }
 
-    static async create(options = {}) {
-      const availability = await LanguageModel.availability(options);
-      // This will be relevant when the implementation is backed by a local
-      // model that needs downloading and simulates the Prompt API's behavior.
-      if (availability === 'downloadable' || availability === 'downloading') {
-        throw new DOMException(
-          'Requires a user gesture when availability is "downloading" or "downloadable".',
-          'NotAllowedError'
-        );
-      }
-      const defaults = {
-        temperature: 1.0,
-        topK: 3,
-      };
-
-      const resolvedOptions = { ...defaults, ...options };
-
-      const inCloudParams = {
-        model: MODEL_NAME,
-        generationConfig: {
-          temperature: resolvedOptions.temperature,
-          topK: resolvedOptions.topK,
-        },
-      };
-
-      let initialHistory = [];
-      let systemInstruction = undefined;
-
-      if (
-        resolvedOptions.initialPrompts &&
-        Array.isArray(resolvedOptions.initialPrompts)
-      ) {
-        const systemPrompts = resolvedOptions.initialPrompts.filter(
-          (p) => p.role === 'system'
-        );
-        const conversationPrompts = resolvedOptions.initialPrompts.filter(
-          (p) => p.role !== 'system'
-        );
-
-        if (systemPrompts.length > 0) {
-          inCloudParams.systemInstruction = systemPrompts
-            .map((p) => p.content)
-            .join('\n');
-        }
-        // Await the conversion of history items (in case of images in history)
-        initialHistory = await convertToFirebaseHistory(conversationPrompts);
-      }
+    const historyCopy = JSON.parse(JSON.stringify(this.#history));
+    const mergedOptions = { ...this.#options, ...options };
+    const mergedInCloudParams = { ...this.#inCloudParams };
 
-      const model = getGenerativeModel(ai, {
-        mode: InferenceMode.ONLY_IN_CLOUD,
-        inCloudParams,
-      });
+    if (options.temperature !== undefined) {
+      mergedInCloudParams.generationConfig.temperature = options.temperature;
+    }
+    if (options.topK !== undefined) {
+      mergedInCloudParams.generationConfig.topK = options.topK;
+    }
 
-      // If a monitor callback is provided, simulate simple downloadprogress events
-      if (typeof resolvedOptions.monitor === 'function') {
-        const monitorTarget = new EventTarget();
+    // Re-create the backend for the clone since it now holds state (#model)
+    const BackendClass = await LanguageModel.#getBackendClass();
+    const info = LanguageModel.#getBackendInfo();
+    const newBackend = new BackendClass(info.configValue);
+    const newModel = newBackend.createSession(
+      mergedOptions,
+      mergedInCloudParams
+    );
 
-        // Let the caller attach listeners
-        try {
-          resolvedOptions.monitor(monitorTarget);
-        } catch (e) {
-          console.error('Error in monitor callback:', e);
-        }
+    return new LanguageModel(
+      newBackend,
+      newModel,
+      historyCopy,
+      mergedOptions,
+      mergedInCloudParams
+    );
+  }
 
-        // Fire two fake downloadprogress events: first with loaded = 0, then loaded = 1
-        try {
-          const startEvent = new ProgressEvent('downloadprogress', {
-            loaded: 0,
-            total: 1,
-          });
-          const endEvent = new ProgressEvent('downloadprogress', {
-            loaded: 1,
-            total: 1,
-          });
-          // The `ProgressEvent`'s `currentTarget`, `srcElement` and `target`
-          // properties are `EventTarget`, not `CreateMonitor`, when using the
-          // polyfill. Hopefully developers won't rely on these properties.
-          monitorTarget.dispatchEvent(startEvent);
-          monitorTarget.dispatchEvent(endEvent);
-        } catch (e) {
-          console.error('Error dispatching downloadprogress events:', e);
-        }
-      }
+  destroy() {
+    this.#destroyed = true;
+    this.#history = null;
+  }
 
-      return new LanguageModel(
-        model,
-        initialHistory,
-        resolvedOptions,
-        inCloudParams
-      );
+  async prompt(input, options = {}) {
+    if (this.#destroyed) {
+      throw new DOMException('Session is destroyed', 'InvalidStateError');
+    }
+    if (options.signal?.aborted) {
+      throw new DOMException('Aborted', 'AbortError');
     }
 
-    // Instance Methods
-
-    async clone(options = {}) {
-      if (this.#destroyed)
-        throw new DOMException('Session is destroyed', 'InvalidStateError');
-      // Clone private history
-      const historyCopy = JSON.parse(JSON.stringify(this.#history));
-      return new LanguageModel(
-        this.#model,
-        historyCopy,
-        {
-          ...this.#options,
-          ...options,
-        },
-        this.#inCloudParams
+    if (options.responseConstraint) {
+      // Update Schema
+      const schema = convertJsonSchemaToVertexSchema(
+        options.responseConstraint
       );
-    }
+      this.#inCloudParams.generationConfig.responseMimeType =
+        'application/json';
+      this.#inCloudParams.generationConfig.responseSchema = schema;
 
-    destroy() {
-      this.#destroyed = true;
-      this.#history = null;
+      // Re-create model with new config/schema (stored in backend)
+      this.#model = this.#backend.createSession(
+        this.#options,
+        this.#inCloudParams
+      );
     }
 
-    async prompt(input, options = {}) {
-      if (this.#destroyed)
-        throw new DOMException('Session is destroyed', 'InvalidStateError');
-      if (options.signal?.aborted)
-        throw new DOMException('Aborted', 'AbortError');
-
-      if (options.responseConstraint) {
-        const vertexSchema = convertJsonSchemaToVertexSchema(
-          options.responseConstraint
-        );
-        this.#inCloudParams.generationConfig.responseMimeType =
-          'application/json';
-        this.#inCloudParams.generationConfig.responseSchema = vertexSchema;
-        this.#model = getGenerativeModel(ai, {
-          mode: InferenceMode.ONLY_IN_CLOUD,
-          inCloudParams: this.#inCloudParams,
-        });
-      }
+    // Process Input (Async conversion of Blob/Canvas/AudioBuffer)
+    const parts = await this.#processInput(input);
+    const userContent = { role: 'user', parts: parts };
 
-      // Process Input (Async conversion of Blob/Canvas/AudioBuffer)
-      const parts = await this.#processInput(input);
-      const userContent = { role: 'user', parts: parts };
+    try {
+      // Estimate usage
+      const totalTokens = await this.#backend.countTokens([
+        { role: 'user', parts },
+      ]);
 
-      try {
-        // Estimate usage before request to fire quota events if needed
-        const { totalTokens } = await this.#model.countTokens({
-          contents: [{ role: 'user', parts }],
-        });
-        if (this.#inputUsage + totalTokens > this.inputQuota)
-          this.dispatchEvent(new Event('quotaoverflow'));
+      if (this.#inputUsage + totalTokens > this.inputQuota) {
+        this.dispatchEvent(new Event('quotaoverflow'));
+      }
 
-        const requestContents = [...this.#history, userContent];
+      const requestContents = [...this.#history, userContent];
 
-        const result = await this.#model.generateContent({
-          contents: requestContents,
-        });
+      const { text, usage } =
+        await this.#backend.generateContent(requestContents);
 
-        // Exact usage update from Backend response
-        if (result.response.usageMetadata?.totalTokenCount) {
-          this.#inputUsage = result.response.usageMetadata.totalTokenCount;
-        }
+      if (usage) {
+        this.#inputUsage = usage;
+      }
 
-        const responseText = result.response.text();
+      this.#history.push(userContent);
+      this.#history.push({ role: 'model', parts: [{ text }] });
 
-        this.#history.push(userContent);
-        this.#history.push({ role: 'model', parts: [{ text: responseText }] });
+      return text;
+    } catch (error) {
+      console.error('Prompt API Polyfill Error:', error);
+      throw error;
+    }
+  }
 
-        return responseText;
-      } catch (error) {
-        console.error('Firebase AI Logic Error:', error);
-        throw error;
-      }
+  promptStreaming(input, options = {}) {
+    if (this.#destroyed) {
+      throw new DOMException('Session is destroyed', 'InvalidStateError');
+    }
+    if (options.signal?.aborted) {
+      throw new DOMException('Aborted', 'AbortError');
     }
 
-    promptStreaming(input, options = {}) {
-      if (this.#destroyed)
-        throw new DOMException('Session is destroyed', 'InvalidStateError');
-      if (options.signal?.aborted)
-        throw new DOMException('Aborted', 'AbortError');
-
-      const _this = this; // Capture 'this' to access private fields in callback
-
-      if (options.responseConstraint) {
-        const vertexSchema = convertJsonSchemaToVertexSchema(
-          options.responseConstraint
-        );
-        this.#inCloudParams.generationConfig.responseMimeType =
-          'application/json';
-        this.#inCloudParams.generationConfig.responseSchema = vertexSchema;
-        this.#model = getGenerativeModel(ai, {
-          mode: InferenceMode.ONLY_IN_CLOUD,
-          inCloudParams: this.#inCloudParams,
-        });
-      }
+    const _this = this; // Capture 'this' to access private fields in callback
 
-      const signal = options.signal;
+    const signal = options.signal;
 
-      return new ReadableStream({
-        async start(controller) {
-          const abortError = new DOMException('Aborted', 'AbortError');
+    return new ReadableStream({
+      async start(controller) {
+        const abortError = new DOMException('Aborted', 'AbortError');
 
-          // If already aborted before the stream starts, error the stream.
-          if (signal?.aborted) {
+        if (signal?.aborted) {
+          controller.error(abortError);
+          return;
+        }
+
+        let aborted = false;
+        const onAbort = () => {
+          aborted = true;
+          try {
             controller.error(abortError);
-            return;
+          } catch {
+            // Ignore
           }
+        };
 
-          let aborted = false;
-          const onAbort = () => {
-            aborted = true;
-            try {
-              controller.error(abortError);
-            } catch {
-              // Controller might already be closed/errored; ignore.
-            }
-          };
+        if (signal) {
+          signal.addEventListener('abort', onAbort);
+        }
 
-          if (signal) {
-            signal.addEventListener('abort', onAbort);
+        try {
+          if (options.responseConstraint) {
+            const schema = convertJsonSchemaToVertexSchema(
+              options.responseConstraint
+            );
+            _this.#inCloudParams.generationConfig.responseMimeType =
+              'application/json';
+            _this.#inCloudParams.generationConfig.responseSchema = schema;
+            _this.#model = _this.#backend.createSession(
+              _this.#options,
+              _this.#inCloudParams
+            );
           }
 
-          try {
-            // Access private methods/fields via captured _this
-            const parts = await _this.#processInput(input);
-            const userContent = { role: 'user', parts: parts };
+          const parts = await _this.#processInput(input);
+          const userContent = { role: 'user', parts: parts };
 
-            // Estimate usage before request to fire quota events if needed
-            const { totalTokens } = await _this.#model.countTokens({
-              contents: [{ role: 'user', parts }],
-            });
-            if (_this.#inputUsage + totalTokens > this.inputQuota)
-              this.dispatchEvent(new Event('quotaoverflow'));
+          const totalTokens = await _this.#backend.countTokens([
+            { role: 'user', parts },
+          ]);
 
-            const requestContents = [..._this.#history, userContent];
+          if (_this.#inputUsage + totalTokens > _this.inputQuota) {
+            _this.dispatchEvent(new Event('quotaoverflow'));
+          }
 
-            const result = await _this.#model.generateContentStream({
-              contents: requestContents,
-            });
+          const requestContents = [..._this.#history, userContent];
 
-            let fullResponseText = '';
-
-            for await (const chunk of result.stream) {
-              if (aborted) {
-                // Try to cancel the underlying iterator; ignore any abort-related errors.
-                if (typeof result.stream.return === 'function') {
-                  try {
-                    await result.stream.return();
-                  } catch (e) {
-                    // Ignore cancellation errors (including AbortError).
-                  }
-                }
-                return;
-              }
-              if (chunk.usageMetadata?.totalTokenCount) {
-                _this.#inputUsage += chunk.usageMetadata.totalTokenCount;
+          const stream =
+            await _this.#backend.generateContentStream(requestContents);
+
+          let fullResponseText = '';
+
+          for await (const chunk of stream) {
+            if (aborted) {
+              // Try to cancel if supported
+              if (typeof stream.return === 'function') {
+                await stream.return();
               }
-              const chunkText = chunk.text();
-              fullResponseText += chunkText;
-              controller.enqueue(chunkText);
+              return;
             }
 
-            if (!aborted) {
-              _this.#history.push(userContent);
-              _this.#history.push({
-                role: 'model',
-                parts: [{ text: fullResponseText }],
-              });
+            const chunkText = chunk.text();
+            fullResponseText += chunkText;
 
-              controller.close();
-            }
-          } catch (error) {
-            // If we aborted, we've already signaled an AbortError; otherwise surface the error.
-            if (!aborted) {
-              controller.error(error);
-            }
-          } finally {
-            if (signal) {
-              signal.removeEventListener('abort', onAbort);
+            if (chunk.usageMetadata?.totalTokenCount) {
+              _this.#inputUsage = chunk.usageMetadata.totalTokenCount;
             }
+
+            controller.enqueue(chunkText);
           }
-        },
-      });
-    }
 
-    async append(input, options = {}) {
-      if (this.#destroyed)
-        throw new DOMException('Session is destroyed', 'InvalidStateError');
-      if (options.signal?.aborted)
-        throw new DOMException('Aborted', 'AbortError');
+          if (!aborted) {
+            _this.#history.push(userContent);
+            _this.#history.push({
+              role: 'model',
+              parts: [{ text: fullResponseText }],
+            });
 
-      const parts = await this.#processInput(input);
-      const content = { role: 'user', parts: parts };
+            controller.close();
+          }
+        } catch (error) {
+          if (!aborted) {
+            controller.error(error);
+          }
+        } finally {
+          if (signal) {
+            signal.removeEventListener('abort', onAbort);
+          }
+        }
+      },
+    });
+  }
 
-      try {
-        // Try to get accurate count first
-        const { totalTokens } = await this.#model.countTokens({
-          contents: [...this.#history, content],
-        });
-        this.#inputUsage = totalTokens;
-      } catch {
-        // Do nothing.
-      }
+  async append(input, options = {}) {
+    if (this.#destroyed) {
+      throw new DOMException('Session is destroyed', 'InvalidStateError');
+    }
+    if (options.signal?.aborted) {
+      throw new DOMException('Aborted', 'AbortError');
+    }
+
+    const parts = await this.#processInput(input);
+    const content = { role: 'user', parts: parts };
+
+    try {
+      const totalTokens = await this.#backend.countTokens([
+        ...this.#history,
+        content,
+      ]);
+      this.#inputUsage = totalTokens;
+    } catch {
+      // Do nothing.
+    }
 
-      this.#history.push(content);
+    this.#history.push(content);
 
-      if (this.#inputUsage > this.inputQuota) {
-        this.dispatchEvent(new Event('quotaoverflow'));
-      }
+    if (this.#inputUsage > this.inputQuota) {
+      this.dispatchEvent(new Event('quotaoverflow'));
     }
+  }
 
-    async measureInputUsage(input) {
-      if (this.#destroyed)
-        throw new DOMException('Session is destroyed', 'InvalidStateError');
+  async measureInputUsage(input) {
+    if (this.#destroyed) {
+      throw new DOMException('Session is destroyed', 'InvalidStateError');
+    }
 
-      try {
-        const parts = await this.#processInput(input);
-        const { totalTokens } = await this.#model.countTokens({
-          contents: [{ role: 'user', parts }],
-        });
-        return totalTokens;
-      } catch (e) {
-        // The API can't reject, so just return 0 if we don't know.
-        console.warn(
-          'The underlying API call failed, quota usage (0) is not reported accurately.'
-        );
-        return 0;
-      }
+    try {
+      const parts = await this.#processInput(input);
+      const totalTokens = await this.#backend.countTokens([
+        { role: 'user', parts },
+      ]);
+      return totalTokens || 0;
+    } catch (e) {
+      console.warn(
+        'The underlying API call failed, quota usage (0) is not reported accurately.'
+      );
+      return 0;
     }
+  }
 
-    // Private Helper to process diverse input types
-    async #processInput(input) {
-      if (typeof input === 'string') {
-        return [{ text: input }];
-      }
+  // Private Helper to process diverse input types
+  async #processInput(input) {
+    if (typeof input === 'string') {
+      return [{ text: input }];
+    }
 
-      if (Array.isArray(input)) {
-        if (input.length > 0 && input[0].role) {
-          let combinedParts = [];
-          for (const msg of input) {
-            if (typeof msg.content === 'string') {
-              combinedParts.push({ text: msg.content });
-              if (msg.prefix) {
-                console.warn(
-                  "The `prefix` flag isn't supported and was ignored."
-                );
-              }
-            } else if (Array.isArray(msg.content)) {
-              for (const c of msg.content) {
-                if (c.type === 'text') combinedParts.push({ text: c.value });
-                else {
-                  const part = await MultimodalConverter.convert(
-                    c.type,
-                    c.value
-                  );
-                  combinedParts.push(part);
-                }
+    if (Array.isArray(input)) {
+      if (input.length > 0 && input[0].role) {
+        let combinedParts = [];
+        for (const msg of input) {
+          if (typeof msg.content === 'string') {
+            combinedParts.push({ text: msg.content });
+            if (msg.prefix) {
+              console.warn(
+                "The `prefix` flag isn't supported and was ignored."
+              );
+            }
+          } else if (Array.isArray(msg.content)) {
+            for (const c of msg.content) {
+              if (c.type === 'text') {
+                combinedParts.push({ text: c.value });
+              } else {
+                const part = await MultimodalConverter.convert(c.type, c.value);
+                combinedParts.push(part);
               }
             }
           }
-          return combinedParts;
         }
-        return input.map((s) => ({ text: String(s) }));
+        return combinedParts;
       }
-
-      return [{ text: JSON.stringify(input) }];
+      return input.map((s) => ({ text: String(s) }));
     }
+
+    return [{ text: JSON.stringify(input) }];
   }
+}
 
+if (!('LanguageModel' in window) || window.__FORCE_PROMPT_API_POLYFILL__) {
   // Attach to window
   window.LanguageModel = LanguageModel;
+  LanguageModel.__isPolyfill = true;
   console.log(
-    'Polyfill: window.LanguageModel is now backed by Firebase AI Logic.'
+    'Polyfill: window.LanguageModel is now backed by the Prompt API polyfill.'
   );
-})();
+}