prompt-api-polyfill 1.3.0 → 1.4.0

package/README.md

@@ -140,6 +140,17 @@ npm install prompt-api-polyfill
     apiKey: 'dummy', // Required for now by the loader
     device: 'webgpu', // 'webgpu' or 'cpu'
     dtype: 'q4f16', // Quantization level
+    env: {
+      // Optional: Pass low-level Transformers.js environment overrides
+      allowRemoteModels: true,
+      backends: {
+        onnx: {
+          wasm: {
+            wasmPaths: 'https://cdn.example.com/wasm-assets/',
+          },
+        },
+      },
+    },
   };
 
   if (!('LanguageModel' in window)) {
@@ -169,7 +180,7 @@ including:
 - `prompt()` and `promptStreaming()`
 - Multimodal inputs (text, image, audio)
 - `append()` and `measureContextUsage()`
-- Quota handling via `oncontextwindowoverflow`
+- Quota handling via `oncontextoverflow`
 - `clone()` and `destroy()`
 
 A simplified version of how it is wired up:
@@ -225,6 +236,17 @@ This repo ships with a template file:
   // For Transformers.js:
   "device": "webgpu",
   "dtype": "q4f16",
+  // Optional library-level overrides:
+  "env": {
+    "allowRemoteModels": true,
+    "backends": {
+      "onnx": {
+        "wasm": {
+          "wasmPaths": "https://cdn.example.com/wasm-assets/",
+        },
+      },
+    },
+  },
 }
 ```
 
@@ -281,7 +303,17 @@ Then open `.env.json` and fill in the values.
   "apiKey": "dummy",
   "modelName": "onnx-community/gemma-3-1b-it-ONNX-GQA",
   "device": "webgpu",
-  "dtype": "q4f16"
+  "dtype": "q4f16",
+  "env": {
+    "allowRemoteModels": false,
+    "backends": {
+      "onnx": {
+        "wasm": {
+          "wasmPaths": "https://cdn.example.com/wasm-assets/"
+        }
+      }
+    }
+  }
 }
 ```
 
@@ -304,6 +336,9 @@ Then open `.env.json` and fill in the values.
 
 - `device`: **Transformers.js only**. Either `"webgpu"` or `"cpu"`.
 - `dtype`: **Transformers.js only**. Quantization level (e.g., `"q4f16"`).
+- `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.
 
 - `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

@@ -1,6 +1,6 @@
 import { initializeApp as Fe } from "firebase/app";
 import { VertexAIBackend as ze, GoogleAIBackend as Ue, getAI as je, getGenerativeModel as Ve, InferenceMode as We } from "firebase/ai";
-import { P as Ke, D as Ge } from "../chunks/defaults-CNQngzSd.js";
+import { P as Ke, D as Ge } from "../chunks/defaults-_qJIFiOb.js";
 const qe = () => {
 };
 const be = function(t) {

package/dist/backends/gemini.js

@@ -1,5 +1,5 @@
 import { GoogleGenAI as s } from "@google/genai";
-import { P as a, D as i } from "../chunks/defaults-CNQngzSd.js";
+import { P as a, D as i } from "../chunks/defaults-_qJIFiOb.js";
 class u extends a {
   #n;
   #e;

package/dist/backends/openai.js

@@ -1,5 +1,5 @@
 import u from "openai";
-import { P as m, D as l } from "../chunks/defaults-CNQngzSd.js";
+import { P as m, D as l } from "../chunks/defaults-_qJIFiOb.js";
 class y extends m {
   #e;
   constructor(t) {

package/dist/backends/transformers.js

@@ -1,68 +1,74 @@
-import { pipeline as z, TextStreamer as C } from "@huggingface/transformers";
-import { P as j, D as y } from "../chunks/defaults-CNQngzSd.js";
-class D extends j {
+import { env as j, pipeline as z, TextStreamer as C } from "@huggingface/transformers";
+import { P as $, D as y } from "../chunks/defaults-_qJIFiOb.js";
+class I extends $ {
   #e;
   #t;
-  #a;
-  #n;
+  #i;
   #o;
-  constructor(n = {}) {
-    super(n.modelName || y.transformers.modelName), this.#a = n.device || y.transformers.device, this.#n = n.dtype || y.transformers.dtype;
+  #n;
+  constructor(o = {}) {
+    if (super(o.modelName || y.transformers.modelName), this.#i = o.device || y.transformers.device, this.#o = o.dtype || y.transformers.dtype, o.env) {
+      const t = (n, a) => {
+        for (const [l, e] of Object.entries(a))
+          e && typeof e == "object" && !Array.isArray(e) && n[l] && typeof n[l] == "object" ? t(n[l], e) : n[l] = e;
+      };
+      t(j, o.env);
+    }
   }
   /**
    * Loaded models can be large, so we initialize them lazily.
    * @param {EventTarget} [monitorTarget] - The event target to dispatch download progress events to.
    * @returns {Promise<Object>} The generator.
    */
-  async #s(n) {
+  async #s(o) {
     if (!this.#e) {
-      const t = /* @__PURE__ */ new Map(), a = await $(this.modelName, {
-        dtype: this.#n
+      const t = /* @__PURE__ */ new Map(), n = await M(this.modelName, {
+        dtype: this.#o
       });
-      for (const { path: e, size: o } of a)
-        t.set(e, { loaded: 0, total: o });
-      const i = (e) => {
-        if (!n)
+      for (const { path: e, size: r } of n)
+        t.set(e, { loaded: 0, total: r });
+      const a = (e) => {
+        if (!o)
           return;
-        const o = 1 / 65536, c = Math.floor(e / o) * o;
-        c <= n.__lastProgressLoaded || (n.dispatchEvent(
+        const r = 1 / 65536, c = Math.floor(e / r) * r;
+        c <= o.__lastProgressLoaded || (o.dispatchEvent(
           new ProgressEvent("downloadprogress", {
             loaded: c,
             total: 1,
             lengthComputable: !0
           })
-        ), n.__lastProgressLoaded = c);
+        ), o.__lastProgressLoaded = c);
       }, l = (e) => {
         if (e.status === "initiate")
           if (t.has(e.file)) {
-            const o = t.get(e.file);
-            e.total && (o.total = e.total);
+            const r = t.get(e.file);
+            e.total && (r.total = e.total);
           } else
             t.set(e.file, { loaded: 0, total: e.total || 0 });
         else if (e.status === "progress")
           t.has(e.file) && (t.get(e.file).loaded = e.loaded);
         else if (e.status === "done") {
           if (t.has(e.file)) {
-            const o = t.get(e.file);
-            o.loaded = o.total;
+            const r = t.get(e.file);
+            r.loaded = r.total;
           }
         } else if (e.status === "ready") {
-          i(1);
+          a(1);
           return;
         }
         if (e.status === "progress" || e.status === "done") {
-          let o = 0, c = 0;
+          let r = 0, c = 0;
           for (const { loaded: f, total: d } of t.values())
-            o += f, c += d;
+            r += f, c += d;
           if (c > 0) {
-            const f = o / c;
-            i(Math.min(f, 0.9999));
+            const f = r / c;
+            a(Math.min(f, 0.9999));
           }
         }
       };
-      i(0), this.#e = await z("text-generation", this.modelName, {
-        device: this.#a,
-        dtype: this.#n,
+      a(0), this.#e = await z("text-generation", this.modelName, {
+        device: this.#i,
+        dtype: this.#o,
         progress_callback: l
       }), this.#t = this.#e.tokenizer;
     }
@@ -73,9 +79,9 @@ class D extends j {
    * @param {Object} options - LanguageModel options.
    * @returns {string} 'available' or 'unavailable'.
    */
-  static availability(n) {
-    if (n?.expectedInputs && Array.isArray(n.expectedInputs)) {
-      for (const t of n.expectedInputs)
+  static availability(o) {
+    if (o?.expectedInputs && Array.isArray(o.expectedInputs)) {
+      for (const t of o.expectedInputs)
         if (t.type === "audio" || t.type === "image")
           return "unavailable";
     }
@@ -88,55 +94,55 @@ class D extends j {
    * @param {EventTarget} [monitorTarget] - The event target to dispatch download progress events to.
    * @returns {Promise<Object>} The generator.
    */
-  async createSession(n, t, a) {
-    return n.responseConstraint && console.warn(
+  async createSession(o, t, n) {
+    return o.responseConstraint && console.warn(
       "The `responseConstraint` flag isn't supported by the Transformers.js backend and was ignored."
-    ), await this.#s(a), this.generationConfig = {
+    ), await this.#s(n), this.generationConfig = {
       max_new_tokens: 512,
       // Default limit
       temperature: t.generationConfig?.temperature ?? 1,
       top_p: 1,
       do_sample: t.generationConfig?.temperature !== 0,
       return_full_text: !1
-    }, this.#o = t.systemInstruction, this.#e;
+    }, this.#n = t.systemInstruction, this.#e;
   }
-  async generateContent(n) {
-    const t = await this.#s(), a = this.#r(n), i = this.#t.apply_chat_template(a, {
+  async generateContent(o) {
+    const t = await this.#s(), n = this.#r(o), a = this.#t.apply_chat_template(n, {
       tokenize: !1,
       add_generation_prompt: !0
-    }), e = (await t(i, {
+    }), e = (await t(a, {
       ...this.generationConfig,
       add_special_tokens: !1
-    }))[0].generated_text, o = await this.countTokens(n);
-    return { text: e, usage: o };
+    }))[0].generated_text, r = await this.countTokens(o);
+    return { text: e, usage: r };
   }
-  async generateContentStream(n) {
-    const t = await this.#s(), a = this.#r(n), i = this.#t.apply_chat_template(a, {
+  async generateContentStream(o) {
+    const t = await this.#s(), n = this.#r(o), a = this.#t.apply_chat_template(n, {
       tokenize: !1,
       add_generation_prompt: !0
     }), l = [];
-    let e, o = new Promise((r) => e = r), c = !1;
-    const f = (r) => {
-      l.push(r), e && (e(), e = null);
+    let e, r = new Promise((i) => e = i), c = !1;
+    const f = (i) => {
+      l.push(i), e && (e(), e = null);
     }, d = new C(this.#t, {
       skip_prompt: !0,
       skip_special_tokens: !0,
       callback_function: f
     });
-    return t(i, {
+    return t(a, {
       ...this.generationConfig,
       add_special_tokens: !1,
       streamer: d
     }).then(() => {
       c = !0, e && (e(), e = null);
-    }).catch((r) => {
-      console.error("[Transformers.js] Generation error:", r), c = !0, e && (e(), e = null);
+    }).catch((i) => {
+      console.error("[Transformers.js] Generation error:", i), c = !0, e && (e(), e = null);
     }), (async function* () {
       for (; ; ) {
-        for (l.length === 0 && !c && (e || (o = new Promise((r) => e = r)), await o); l.length > 0; ) {
-          const r = l.shift();
+        for (l.length === 0 && !c && (e || (r = new Promise((i) => e = i)), await r); l.length > 0; ) {
+          const i = l.shift();
           yield {
-            text: () => r,
+            text: () => i,
             usageMetadata: { totalTokenCount: 0 }
           };
         }
@@ -145,67 +151,67 @@ class D extends j {
       }
     })();
   }
-  async countTokens(n) {
+  async countTokens(o) {
     await this.#s();
-    const t = this.#r(n);
+    const t = this.#r(o);
     return this.#t.apply_chat_template(t, {
       tokenize: !0,
       add_generation_prompt: !1,
       return_tensor: !1
     }).length;
   }
-  #r(n) {
-    const t = n.map((a) => {
-      let i = a.role === "model" ? "assistant" : a.role === "system" ? "system" : "user";
-      const l = a.parts.map((e) => e.text).join("");
-      return { role: i, content: l };
+  #r(o) {
+    const t = o.map((n) => {
+      let a = n.role === "model" ? "assistant" : n.role === "system" ? "system" : "user";
+      const l = n.parts.map((e) => e.text).join("");
+      return { role: a, content: l };
     });
-    if (this.#o && !t.some((a) => a.role === "system") && t.unshift({ role: "system", content: this.#o }), this.modelName.toLowerCase().includes("gemma")) {
-      const a = t.findIndex((i) => i.role === "system");
-      if (a !== -1) {
-        const i = t[a], l = t.findIndex(
-          (e, o) => e.role === "user" && o > a
+    if (this.#n && !t.some((n) => n.role === "system") && t.unshift({ role: "system", content: this.#n }), this.modelName.toLowerCase().includes("gemma")) {
+      const n = t.findIndex((a) => a.role === "system");
+      if (n !== -1) {
+        const a = t[n], l = t.findIndex(
+          (e, r) => e.role === "user" && r > n
         );
-        l !== -1 ? (t[l].content = i.content + `
+        l !== -1 ? (t[l].content = a.content + `
 
-` + t[l].content, t.splice(a, 1)) : (i.content += `
+` + t[l].content, t.splice(n, 1)) : (a.content += `
 
-`, i.role = "user");
+`, a.role = "user");
       }
     }
     return t;
   }
 }
-async function $(g, n = {}) {
-  const { dtype: t = "q8", branch: a = "main" } = n;
-  let i = null;
-  const l = `transformers_model_files_${g}_${t}_${a}`;
+async function M(g, o = {}) {
+  const { dtype: t = "q8", branch: n = "main" } = o;
+  let a = null;
+  const l = `transformers_model_files_${g}_${t}_${n}`;
   try {
     const s = localStorage.getItem(l);
     if (s) {
-      i = JSON.parse(s);
-      const { timestamp: p, files: u } = i, v = 1440 * 60 * 1e3;
+      a = JSON.parse(s);
+      const { timestamp: p, files: u } = a, v = 1440 * 60 * 1e3;
       if (Date.now() - p < v)
         return u;
     }
   } catch (s) {
     console.warn("Failed to read from localStorage cache:", s);
   }
-  const e = `https://huggingface.co/api/models/${g}/tree/${a}?recursive=true`;
-  let o;
+  const e = `https://huggingface.co/api/models/${g}/tree/${n}?recursive=true`;
+  let r;
   try {
-    if (o = await fetch(e), !o.ok)
-      throw new Error(`Manifest fetch failed: ${o.status}`);
+    if (r = await fetch(e), !r.ok)
+      throw new Error(`Manifest fetch failed: ${r.status}`);
   } catch (s) {
-    if (i)
+    if (a)
       return console.warn(
         "Failed to fetch manifest from network, falling back to cached data (expired):",
         s
-      ), i.files;
+      ), a.files;
     throw s;
   }
-  const c = await o.json(), f = new Map(c.map((s) => [s.path, s.size])), d = [], h = (s) => f.has(s), r = (s) => h(s) ? (d.push({ path: s, size: f.get(s) }), !0) : !1;
-  r("config.json"), r("generation_config.json"), r("preprocessor_config.json"), h("tokenizer.json") ? (r("tokenizer.json"), r("tokenizer_config.json")) : (r("tokenizer_config.json"), r("special_tokens_map.json"), r("vocab.json"), r("merges.txt"), r("vocab.txt"));
+  const c = await r.json(), f = new Map(c.map((s) => [s.path, s.size])), d = [], h = (s) => f.has(s), i = (s) => h(s) ? (d.push({ path: s, size: f.get(s) }), !0) : !1;
+  i("config.json"), i("generation_config.json"), i("preprocessor_config.json"), h("tokenizer.json") ? (i("tokenizer.json"), i("tokenizer_config.json")) : (i("tokenizer_config.json"), i("special_tokens_map.json"), i("vocab.json"), i("merges.txt"), i("vocab.txt"));
   const w = "onnx";
   let m = [];
   t === "fp32" ? m = [""] : t === "quantized" ? m = ["_quantized"] : (m = [`_${t}`], t === "q8" && m.push(""));
@@ -228,11 +234,11 @@ async function $(g, n = {}) {
     (s) => s.includes("decoder_model_merged")
   ), b = _.filter((s) => !(x && s.includes("decoder_model") && !s.includes("merged")));
   for (const s of b) {
-    r(s);
+    i(s);
     const p = `${s}_data`;
-    if (r(p)) {
+    if (i(p)) {
       let u = 1;
-      for (; r(`${p}_${u}`); )
+      for (; i(`${p}_${u}`); )
         u++;
     }
   }
@@ -250,5 +256,5 @@ async function $(g, n = {}) {
   return d;
 }
 export {
-  D as default
+  I as default
 };

package/dist/chunks/{defaults-CNQngzSd.js → defaults-_qJIFiOb.js}

@@ -1,4 +1,4 @@
-class r {
+class a {
   #e;
   /**
    * @param {string} modelName - The name of the model.
@@ -49,9 +49,9 @@ class r {
     throw new Error("Not implemented");
   }
 }
-const a = {
+const r = {
   firebase: { modelName: "gemini-2.5-flash-lite" },
-  gemini: { modelName: "gemini-2.0-flash-lite-preview-02-05" },
+  gemini: { modelName: "gemini-2.5-flash-lite" },
   openai: { modelName: "gpt-4o" },
   transformers: {
     modelName: "onnx-community/gemma-3-1b-it-ONNX-GQA",
@@ -60,6 +60,6 @@ const a = {
   }
 };
 export {
-  a as D,
-  r as P
+  r as D,
+  a as P
 };

package/dist/prompt-api-polyfill.js

@@ -283,7 +283,7 @@ async function C(s, t = globalThis) {
   }
   return e;
 }
-class h extends EventTarget {
+class d extends EventTarget {
   #o;
   #f;
   #r;
@@ -302,16 +302,16 @@ class h extends EventTarget {
   get contextWindow() {
     return 1e6;
   }
-  get oncontextwindowoverflow() {
+  get oncontextoverflow() {
     return this.#c;
   }
-  set oncontextwindowoverflow(t) {
+  set oncontextoverflow(t) {
     this.#c && this.removeEventListener(
-      "contextwindowoverflow",
+      "contextoverflow",
       this.#c
-    ), this.#c = t, typeof t == "function" && this.addEventListener("contextwindowoverflow", t);
+    ), this.#c = t, typeof t == "function" && this.addEventListener("contextoverflow", t);
   }
-  static #d(t) {
+  static #h(t) {
     try {
       if (!t || !t.document || t.document.defaultView !== t)
         throw new Error();
@@ -326,13 +326,13 @@ class h extends EventTarget {
     }
   }
   #a() {
-    h.#d(this.#t);
+    d.#h(this.#t);
   }
   static async availability(t = {}) {
     const e = this.__window || globalThis;
-    h.#d(e);
+    d.#h(e);
     try {
-      await h.#w(t, e);
+      await d.#w(t, e);
     } catch (o) {
       if (o instanceof RangeError) {
         if (o.message.includes("language tag"))
@@ -348,23 +348,23 @@ class h extends EventTarget {
       }
       return "unavailable";
     }
-    return (await h.#p(e)).availability(t);
+    return (await d.#p(e)).availability(t);
   }
   static #x = A;
-  static #h(t = globalThis) {
-    for (const r of h.#x) {
+  static #d(t = globalThis) {
+    for (const r of d.#x) {
       const o = t[r.config] || globalThis[r.config];
       if (o && o.apiKey)
         return { ...r, configValue: o };
     }
-    const e = h.#x.map((r) => `window.${r.config}`).join(", ");
+    const e = d.#x.map((r) => `window.${r.config}`).join(", ");
     throw new (t.DOMException || globalThis.DOMException)(
       `Prompt API Polyfill: No backend configuration found. Please set one of: ${e}.`,
       "NotSupportedError"
     );
   }
   static async #p(t = globalThis) {
-    const e = h.#h(t);
+    const e = d.#d(t);
     return I(e.path);
   }
   static async #w(t = {}, e = globalThis) {
@@ -372,13 +372,13 @@ class h extends EventTarget {
       for (const o of t.expectedInputs) {
         if (o.type !== "text" && o.type !== "image" && o.type !== "audio")
           throw new TypeError(`Invalid input type: ${o.type}`);
-        o.languages && h.#y(o.languages);
+        o.languages && d.#y(o.languages);
       }
     if (t.expectedOutputs)
       for (const o of t.expectedOutputs) {
         if (o.type !== "text")
           throw new RangeError(`Unsupported output type: ${o.type}`);
-        o.languages && h.#y(o.languages);
+        o.languages && d.#y(o.languages);
       }
     const r = t.expectedInputs ? ["text", ...t.expectedInputs.map((o) => o.type)] : ["text"];
     if (t.initialPrompts && Array.isArray(t.initialPrompts)) {
@@ -434,7 +434,7 @@ class h extends EventTarget {
   }
   static async create(t = {}) {
     const e = this.__window || globalThis;
-    if (h.#d(e), await h.#w(t, e), t.signal?.aborted)
+    if (d.#h(e), await d.#w(t, e), t.signal?.aborted)
       throw t.signal.reason || new (e.DOMException || globalThis.DOMException)(
         "Aborted",
         "AbortError"
@@ -455,8 +455,8 @@ class h extends EventTarget {
         "Aborted",
         "AbortError"
       );
-    const o = h.#h(e), n = await h.#p(e), i = new n(o.configValue), c = { ...t };
-    h.#g(
+    const o = d.#d(e), n = await d.#p(e), i = new n(o.configValue), c = { ...t };
+    d.#g(
       c.responseConstraint,
       e
     );
@@ -471,16 +471,16 @@ class h extends EventTarget {
       ), p = c.initialPrompts.filter(
         (f) => f.role !== "system"
       );
-      w.length > 0 && (a.systemInstruction = w.map((f) => typeof f.content == "string" ? f.content : Array.isArray(f.content) ? f.content.filter((d) => d.type === "text").map((d) => d.value || d.text || "").join(`
+      w.length > 0 && (a.systemInstruction = w.map((f) => typeof f.content == "string" ? f.content : Array.isArray(f.content) ? f.content.filter((h) => h.type === "text").map((h) => h.value || h.text || "").join(`
 `) : "").join(`
 `)), l = await C(p, e);
       for (const f of c.initialPrompts) {
         if (typeof f.content != "string")
           continue;
-        const d = h.#E([
+        const h = d.#E([
           { text: f.content }
         ]);
-        if (d === "QuotaExceededError" || d === "contextwindowoverflow") {
+        if (h === "QuotaExceededError" || h === "contextoverflow") {
           const E = e.QuotaExceededError || e.DOMException || globalThis.QuotaExceededError || globalThis.DOMException, u = new E(
             "The initial prompts are too large, they exceed the quota.",
             "QuotaExceededError"
@@ -489,7 +489,7 @@ class h extends EventTarget {
             value: 22,
             configurable: !0
           });
-          const g = d === "QuotaExceededError" ? 1e7 : 5e5;
+          const g = h === "QuotaExceededError" ? 1e7 : 5e5;
           throw u.requested = g, u.quota = 1e6, u;
         }
       }
@@ -518,10 +518,10 @@ class h extends EventTarget {
             lengthComputable: !0
           })
         ), x.__lastProgressLoaded = f;
-      } catch (d) {
-        console.error("Error dispatching downloadprogress events:", d);
+      } catch (h) {
+        console.error("Error dispatching downloadprogress events:", h);
       }
-      return await new Promise((d) => setTimeout(d, 0)), !t.signal?.aborted;
+      return await new Promise((h) => setTimeout(h, 0)), !t.signal?.aborted;
     };
     if (!await b(0))
       throw t.signal.reason || new (e.DOMException || globalThis.DOMException)(
@@ -573,7 +573,7 @@ class h extends EventTarget {
         "Aborted",
         "AbortError"
       );
-    const e = JSON.parse(JSON.stringify(this.#r)), r = { ...this.#s, ...t }, o = await h.#p(this.#t), n = h.#h(this.#t), i = new o(n.configValue), c = await i.createSession(
+    const e = JSON.parse(JSON.stringify(this.#r)), r = { ...this.#s, ...t }, o = await d.#p(this.#t), n = d.#d(this.#t), i = new o(n.configValue), c = await i.createSession(
       r,
       this.#e
     );
@@ -609,7 +609,7 @@ class h extends EventTarget {
     if (typeof t == "object" && t !== null && !Array.isArray(t) && Object.keys(t).length === 0)
       return "[object Object]";
     if (e.responseConstraint) {
-      h.#g(
+      d.#g(
         e.responseConstraint,
         this.#t
       );
@@ -652,15 +652,15 @@ class h extends EventTarget {
     }), c = (async () => {
       const a = this.#u(o);
       if (a === "QuotaExceededError") {
-        const f = this.#t && this.#t.QuotaExceededError || this.#t && this.#t.DOMException || globalThis.QuotaExceededError || globalThis.DOMException, d = new f(
+        const f = this.#t && this.#t.QuotaExceededError || this.#t && this.#t.DOMException || globalThis.QuotaExceededError || globalThis.DOMException, h = new f(
           "The prompt is too large, it exceeds the quota.",
           "QuotaExceededError"
         );
-        Object.defineProperty(d, "code", { value: 22, configurable: !0 });
+        Object.defineProperty(h, "code", { value: 22, configurable: !0 });
         const E = 1e7;
-        throw d.requested = E, d.quota = this.contextWindow, d;
-      } else if (a === "contextwindowoverflow")
-        return this.dispatchEvent(new Event("contextwindowoverflow")), "Mock response for quota overflow test.";
+        throw h.requested = E, h.quota = this.contextWindow, h;
+      } else if (a === "contextoverflow")
+        return this.dispatchEvent(new Event("contextoverflow")), "Mock response for quota overflow test.";
       const l = [...this.#r, n];
       this.#e.systemInstruction && l.unshift({
         role: "system",
@@ -670,13 +670,13 @@ class h extends EventTarget {
         l
       );
       if (y > this.contextWindow) {
-        const f = this.#t && this.#t.QuotaExceededError || this.#t && this.#t.DOMException || globalThis.QuotaExceededError || globalThis.DOMException, d = new f(
+        const f = this.#t && this.#t.QuotaExceededError || this.#t && this.#t.DOMException || globalThis.QuotaExceededError || globalThis.DOMException, h = new f(
           `The prompt is too large (${y} tokens), it exceeds the quota of ${this.contextWindow} tokens.`,
           "QuotaExceededError"
         );
-        throw Object.defineProperty(d, "code", { value: 22, configurable: !0 }), d.requested = y, d.quota = this.contextWindow, d;
+        throw Object.defineProperty(h, "code", { value: 22, configurable: !0 }), h.requested = y, h.quota = this.contextWindow, h;
       }
-      y > this.contextWindow && this.dispatchEvent(new Event("contextwindowoverflow"));
+      y > this.contextWindow && this.dispatchEvent(new Event("contextoverflow"));
       const x = [...this.#r, n];
       let b;
       try {
@@ -737,7 +737,7 @@ class h extends EventTarget {
         o && o.addEventListener("abort", c);
         try {
           if (e.responseConstraint) {
-            h.#g(
+            d.#g(
               e.responseConstraint,
               r.#t
             );
@@ -767,8 +767,8 @@ class h extends EventTarget {
             });
             const O = 1e7;
             throw g.requested = O, g.quota = r.contextWindow, g;
-          } else if (x === "contextwindowoverflow") {
-            r.dispatchEvent(new Event("contextwindowoverflow")), n.enqueue("Mock response for quota overflow test."), n.close();
+          } else if (x === "contextoverflow") {
+            r.dispatchEvent(new Event("contextoverflow")), n.enqueue("Mock response for quota overflow test."), n.close();
             return;
           }
           const b = [...r.#r, y];
@@ -789,7 +789,7 @@ class h extends EventTarget {
               configurable: !0
             }), g.requested = m, g.quota = r.contextWindow, g;
           }
-          m > r.contextWindow && r.dispatchEvent(new Event("contextwindowoverflow"));
+          m > r.contextWindow && r.dispatchEvent(new Event("contextoverflow"));
           const w = [...r.#r, y];
           let p;
           try {
@@ -797,20 +797,20 @@ class h extends EventTarget {
           } catch (u) {
             throw r.#m(u, l), u;
           }
-          let f = "", d = !1, E = "";
+          let f = "", h = !1, E = "";
           for await (const u of p) {
             if (i) {
               typeof p.return == "function" && await p.return();
               return;
             }
             let g = u.text();
-            if (a && !d) {
+            if (a && !h) {
               E += g;
               const O = E.match(/^\s*{\s*"Rating"\s*:\s*/);
               if (O)
-                g = E.slice(O[0].length), d = !0, E = "";
+                g = E.slice(O[0].length), h = !0, E = "";
               else if (E.length > 50)
-                g = E, d = !0, E = "";
+                g = E, h = !0, E = "";
               else
                 continue;
             }
@@ -857,7 +857,7 @@ class h extends EventTarget {
       this.#i = i || 0;
     } catch {
     }
-    this.#i > this.contextWindow && this.dispatchEvent(new Event("contextwindowoverflow"));
+    this.#i > this.contextWindow && this.dispatchEvent(new Event("contextoverflow"));
   }
   async measureContextUsage(t) {
     if (this.#a(), this.#n)
@@ -873,7 +873,7 @@ class h extends EventTarget {
           "InvalidStateError"
         );
       const r = this.#u(e);
-      return r === "QuotaExceededError" ? 1e7 : r === "contextwindowoverflow" ? 5e5 : await this.#o.countTokens([
+      return r === "QuotaExceededError" ? 1e7 : r === "contextoverflow" ? 5e5 : await this.#o.countTokens([
         { role: "user", parts: e }
       ]) || 0;
     } catch {
@@ -884,13 +884,13 @@ class h extends EventTarget {
   }
   // Volkswagen mode detection to avoid cloud costs for WPT tests.
   #u(t) {
-    return h.#E(t);
+    return d.#E(t);
   }
   static #E(t) {
     if (t.length !== 1 || !t[0].text)
       return null;
     const e = t[0].text;
-    return typeof e != "string" || !e.startsWith("Please write a sentence in English.") ? null : e.length > 1e7 ? "QuotaExceededError" : e.length > 5e4 ? "contextwindowoverflow" : null;
+    return typeof e != "string" || !e.startsWith("Please write a sentence in English.") ? null : e.length > 1e7 ? "QuotaExceededError" : e.length > 5e4 ? "contextoverflow" : null;
   }
   static #g(t, e) {
     if (t)
@@ -1003,7 +1003,7 @@ const D = (s) => {
   try {
     if (!s || s.LanguageModel?.__isPolyfill)
       return;
-    const t = class extends h {
+    const t = class extends d {
     };
     t.__window = s, t.__isPolyfill = !0, s.LanguageModel = t, s.DOMException && (s.QuotaExceededError = s.DOMException);
   } catch {
@@ -1037,9 +1037,9 @@ globalThis.document?.documentElement && (P.observe(globalThis.document.documentE
 }), globalThis.document.querySelectorAll("iframe").forEach((s) => {
   D(s.contentWindow);
 }));
-(!("LanguageModel" in globalThis) || globalThis.__FORCE_PROMPT_API_POLYFILL__) && (globalThis.LanguageModel = h, h.__isPolyfill = !0, console.log(
+(!("LanguageModel" in globalThis) || globalThis.__FORCE_PROMPT_API_POLYFILL__) && (globalThis.LanguageModel = d, d.__isPolyfill = !0, console.log(
   "Polyfill: window.LanguageModel is now backed by the Prompt API polyfill."
 ));
 export {
-  h as LanguageModel
+  d as LanguageModel
 };

package/dot_env.json

@@ -7,5 +7,15 @@
   "reCaptchaSiteKey": "",
   "useLimitedUseAppCheckTokens": true,
   "device": "webgpu",
-  "dtype": "q4f16"
+  "dtype": "q4f16",
+  "env": {
+    "allowRemoteModels": true,
+    "backends": {
+      "onnx": {
+        "wasm": {
+          "wasmPaths": ""
+        }
+      }
+    }
+  }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "prompt-api-polyfill",
-  "version": "1.3.0",
+  "version": "1.4.0",
   "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",
@@ -47,7 +47,9 @@
   "devDependencies": {
     "prettier": "^3.8.1",
     "prettier-plugin-curly": "^0.4.1",
-    "vite": "^7.3.1"
+    "vite": "^7.3.1",
+    "node-gyp": "^12.2.0",
+    "node-addon-api": "^8.5.0"
   },
   "dependencies": {
     "@google/genai": "^1.42.0",