prompt-api-polyfill 1.17.0 → 1.18.0

package/README.md

@@ -8,6 +8,7 @@ supporting dynamic backends:
 - **Google Gemini API** (cloud)
 - **OpenAI API** (cloud)
 - **Transformers.js** (local after initial model download, **default backend**)
+- **WebLLM** (local after initial model download)
 
 When loaded in the browser, it defines a global:
 
@@ -57,6 +58,13 @@ configured, the polyfill will use Transformers.js with the default model.
 - **Model**: Uses default if not specified (see
   [`backends/defaults.js`](backends/defaults.js)).
 
+### WebLLM (local after initial model download)
+
+- **Uses**: `@mlc-ai/web-llm` SDK (MLC engine, runs via WebGPU).
+- **Select by setting**: `window.WEBLLM_CONFIG`.
+- **Model**: Uses default if not specified (see
+  [`backends/defaults.js`](backends/defaults.js)).
+
 ---
 
 ## Installation
@@ -166,6 +174,28 @@ npm install prompt-api-polyfill
 </script>
 ```
 
+### Backed by WebLLM (local after initial model download)
+
+1. **Only a dummy API Key required** (runs locally in the browser via WebGPU).
+2. **Provide configuration** on `window.WEBLLM_CONFIG`.
+3. **Import the polyfill**.
+
+```html
+<script type="module">
+  // Set WEBLLM_CONFIG to select the WebLLM backend
+  window.WEBLLM_CONFIG = {
+    apiKey: 'dummy', // Required for now by the loader
+    modelName: 'Llama-3.2-3B-Instruct-q4f32_1-MLC', // Optional: override default
+  };
+
+  if (!('LanguageModel' in window)) {
+    await import('prompt-api-polyfill');
+  }
+
+  const session = await LanguageModel.create();
+</script>
+```
+
 ---
 
 ## Configuration
@@ -322,13 +352,22 @@ Then open `.env.json` and fill in the values.
 }
 ```
 
+**For WebLLM:**
+
+```json
+{
+  "apiKey": "dummy",
+  "modelName": "Llama-3.2-3B-Instruct-q4f32_1-MLC"
+}
+```
+
 ### Field-by-field explanation
 
 - `apiKey`:
   - **Firebase AI Logic**: Your Firebase Web API key.
   - **Gemini**: Your Gemini API Key.
   - **OpenAI**: Your OpenAI API Key.
-  - **Transformers.js**: Use `"dummy"`.
+  - **Transformers.js** / **WebLLM**: Use `"dummy"`.
 - `projectId` / `appId`: **Firebase AI Logic only**.
 - `geminiApiProvider`: **Firebase AI Logic only**. Either `"developer"` (Gemini
   Developer API, default) or `"vertex"` (Vertex AI).
@@ -344,6 +383,10 @@ Then open `.env.json` and fill in the values.
 - `env` (optional): **Transformers.js only**. A flexible object to override
   [Transformers.js environment variables](https://huggingface.co/docs/transformers.js/api/env).
   This is useful for specifying local `wasmPaths` or proxy settings.
+- **WebLLM** only requires `apiKey: "dummy"` and an optional `modelName`. Model
+  IDs must be valid
+  [MLC model identifiers](https://github.com/mlc-ai/web-llm?tab=readme-ov-file#list-available-models)
+  (e.g., `"Llama-3.2-3B-Instruct-q4f32_1-MLC"`).
 
 - `modelName` (optional): The model ID to use. If not provided, the polyfill
   uses the defaults defined in [`backends/defaults.js`](backends/defaults.js).

package/dist/backends/firebase.js

@@ -1,4 +1,4 @@
-import { n as e, t } from "../chunks/defaults-B5W7MP9T.js";
+import { n as e, t } from "../chunks/defaults-DMD-8IKq.js";
 import { initializeApp as n } from "firebase/app";
 import { ReCaptchaEnterpriseProvider as r, initializeAppCheck as i } from "firebase/app-check";
 import { GoogleAIBackend as a, InferenceMode as o, Schema as s, VertexAIBackend as c, getAI as l, getGenerativeModel as u } from "firebase/ai";

package/dist/backends/gemini.js

@@ -1,4 +1,4 @@
-import { n as e, t } from "../chunks/defaults-B5W7MP9T.js";
+import { n as e, t } from "../chunks/defaults-DMD-8IKq.js";
 import { GoogleGenAI as n } from "@google/genai";
 //#region backends/gemini.js
 var r = class extends e {

package/dist/backends/openai.js

@@ -1,4 +1,4 @@
-import { n as e, t } from "../chunks/defaults-B5W7MP9T.js";
+import { n as e, t } from "../chunks/defaults-DMD-8IKq.js";
 import n from "openai";
 //#region backends/openai.js
 var r = class extends e {

package/dist/backends/transformers.js

@@ -1,4 +1,4 @@
-import { n as e, t } from "../chunks/defaults-B5W7MP9T.js";
+import { n as e, t } from "../chunks/defaults-DMD-8IKq.js";
 import { TextStreamer as n, env as r, pipeline as i } from "@huggingface/transformers";
 //#region backends/transformers.js
 var a = class extends e {

package/dist/backends/webllm.js

@@ -0,0 +1,107 @@
+import { n as e, t } from "../chunks/defaults-DMD-8IKq.js";
+import { CreateMLCEngine as n, prebuiltAppConfig as r } from "@mlc-ai/web-llm";
+//#region backends/webllm.js
+var i = class extends e {
+	#e;
+	#t;
+	#n = 0;
+	constructor(e = {}) {
+		super(e.modelName || t.webllm.modelName);
+	}
+	async #r(e) {
+		if (!this.#e) {
+			let t = (t) => {
+				if (!e) return;
+				let n = 1 / 65536, r = Math.floor(t / n) * n;
+				r <= e.__lastProgressLoaded || (e.dispatchEvent(new ProgressEvent("downloadprogress", {
+					loaded: r,
+					total: 1,
+					lengthComputable: !0
+				})), e.__lastProgressLoaded = r);
+			};
+			t(0);
+			let i = {
+				...r,
+				cacheBackend: "cross-origin"
+			};
+			this.#e = await n(this.modelName, {
+				appConfig: i,
+				initProgressCallback: (e) => {
+					t(e.progress);
+				}
+			}), t(1);
+		}
+		return this.#e;
+	}
+	static availability(e) {
+		if (e?.expectedInputs && Array.isArray(e.expectedInputs)) {
+			for (let t of e.expectedInputs) if (t.type === "audio" || t.type === "image") return "unavailable";
+		}
+		return "available";
+	}
+	async createSession(e, t, n) {
+		return await this.#r(n), this.generationConfig = { max_tokens: 512 }, this.#t = t.systemInstruction, this.responseSchema = t.generationConfig?.responseSchema, this.#e;
+	}
+	async generateContent(e) {
+		let t = await this.#r(), n = {
+			messages: this.#i(e),
+			...this.generationConfig
+		};
+		this.responseSchema && (n.response_format = {
+			type: "json_object",
+			schema: JSON.stringify(this.responseSchema)
+		});
+		let r = await t.chat.completions.create(n), i = r.choices[0].message.content;
+		return this.#n += (r.usage?.prompt_tokens ?? 0) + (r.usage?.completion_tokens ?? 0), {
+			text: i,
+			usage: this.#n
+		};
+	}
+	async generateContentStream(e) {
+		let t = await this.#r(), n = {
+			messages: this.#i(e),
+			...this.generationConfig,
+			stream: !0,
+			stream_options: { include_usage: !0 }
+		};
+		this.responseSchema && (n.response_format = {
+			type: "json_object",
+			schema: JSON.stringify(this.responseSchema)
+		});
+		let r = await t.chat.completions.create(n), i = this;
+		return (async function* () {
+			let e = null;
+			for await (let t of r) {
+				let n = t.choices[0]?.delta?.content ?? "";
+				n && (yield {
+					text: () => n,
+					usageMetadata: { totalTokenCount: 0 }
+				}), t.usage && (e = t.usage);
+			}
+			i.#n += (e?.prompt_tokens ?? 0) + (e?.completion_tokens ?? 0), yield {
+				text: () => "",
+				usageMetadata: { totalTokenCount: i.#n }
+			};
+		})();
+	}
+	async countTokens(e) {
+		let t = "";
+		for (let n of e ?? []) for (let e of n?.parts ?? []) t += e.text ?? "";
+		return Math.ceil(t.length / 4);
+	}
+	#i(e) {
+		let t = e.map((e) => ({
+			role: e.role === "model" ? "assistant" : e.role === "system" ? "system" : "user",
+			content: e.parts.map((e) => e.text).join("")
+		}));
+		return this.#t && !t.some((e) => e.role === "system") && t.unshift({
+			role: "system",
+			content: this.#t
+		}), this.responseSchema && (t.length > 0 && t[0].role === "system" ? t[0].content += "\n\nRespond with valid JSON." : t.unshift({
+			role: "system",
+			content: "Respond with valid JSON."
+		})), t;
+	}
+};
+//#endregion
+export { i as default };

package/dist/chunks/{defaults-B5W7MP9T.js → defaults-DMD-8IKq.js}

@@ -29,7 +29,8 @@ var e = class {
 		modelName: "onnx-community/gemma-3-1b-it-ONNX-GQA",
 		device: "webgpu",
 		dtype: "q4f16"
-	}
+	},
+	webllm: { modelName: "Llama-3.2-3B-Instruct-q4f32_1-MLC" }
 };
 //#endregion
 export { e as n, t };

package/dist/prompt-api-polyfill.js

@@ -158,6 +158,10 @@ var e = class {
 	{
 		config: "TRANSFORMERS_CONFIG",
 		path: "./backends/transformers.js"
+	},
+	{
+		config: "WEBLLM_CONFIG",
+		path: "./backends/webllm.js"
 	}
 ];
 async function n(e) {
@@ -165,6 +169,7 @@ async function n(e) {
 	if (e === "./backends/gemini.js") return (await import("./backends/gemini.js")).default;
 	if (e === "./backends/openai.js") return (await import("./backends/openai.js")).default;
 	if (e === "./backends/transformers.js") return (await import("./backends/transformers.js")).default;
+	if (e === "./backends/webllm.js") return (await import("./backends/webllm.js")).default;
 	throw Error(`Unknown backend path "${e}"`);
 }
 //#endregion

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "prompt-api-polyfill",
-  "version": "1.17.0",
+  "version": "1.18.0",
   "description": "Polyfill for the Prompt API (`LanguageModel`) backed by Firebase AI Logic, Gemini API, OpenAI API, or Transformers.js.",
   "type": "module",
   "main": "./dist/prompt-api-polyfill.js",
@@ -47,19 +47,20 @@
   },
   "devDependencies": {
     "@eslint/js": "^10.0.1",
-    "eslint": "^10.3.0",
+    "eslint": "^10.4.0",
     "eslint-config-prettier": "^10.1.8",
     "globals": "^17.6.0",
     "node-addon-api": "^8.7.0",
     "node-gyp": "^12.3.0",
     "prettier": "^3.8.3",
     "prettier-plugin-curly": "^0.4.1",
-    "vite": "^8.0.11"
+    "vite": "^8.0.13"
   },
   "dependencies": {
-    "@google/genai": "^1.52.0",
+    "@google/genai": "^2.3.0",
     "@huggingface/transformers": "^4.2.0",
-    "firebase": "^12.12.1",
-    "openai": "^6.36.0"
+    "@mlc-ai/web-llm": "^0.2.83",
+    "firebase": "^12.13.0",
+    "openai": "^6.37.0"
   }
 }