@nevescloud/pip 3.5.1 → 3.7.0

package/README.md

@@ -48,6 +48,18 @@ const pip = createPip({ onSubmit: rt.onSubmit, onSlash: rt.onSlash, slashSource:
 // the turn loop entirely).
 ```
 
+```js
+// Chrome — on-device Gemini Nano via the Prompt API (zero download for users on Chrome 138+ that already has weights; ~2B-effective-param quality)
+import { createRuntime }  from 'https://cdn.jsdelivr.net/npm/@nevescloud/pip@latest/runtime.esm.js';
+import { createPip }      from 'https://cdn.jsdelivr.net/npm/@nevescloud/pip@latest/pip-core.esm.js';
+import { chrome }         from 'https://cdn.jsdelivr.net/npm/@nevescloud/pip@latest/providers/chrome.esm.js';
+
+const rt = createRuntime({ provider: chrome({ temperature: 0.1 }) });
+const pip = createPip({ onSubmit: rt.onSubmit, onSlash: rt.onSlash, slashSource: rt.slashSource });
+// No bundle — Chrome doesn't need its own re-export of createPip + createRuntime
+// (`bundle/anthropic` already brings those, and chrome() composes alongside).
+```
+
 On jsdelivr the `.esm.js` suffix is required — jsdelivr serves files by raw path, not via `package.json` exports. npm-installed consumers can use the shorter `@nevescloud/pip/bundle/anthropic` (Node ESM resolver honors the exports map). `pip/bundle.esm.js` (or `pip/bundle` via npm) is an alias for `bundle/anthropic` — the default when you haven't picked a brain. Bundles are sugar over the layered files; hosts with a different brain shape (UI only, custom provider, in-browser model) import the granular files directly. See [CONSUMERS.md](../../CONSUMERS.md) for the full entry-point list.
 
 ## Options

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nevescloud/pip",
-  "version": "3.5.1",
+  "version": "3.7.0",
   "description": "Floating assistant bubble + panel + chat runtime. ESM, no build.",
   "type": "module",
   "main": "pip-core.esm.js",
@@ -19,7 +19,8 @@
     "./bundle/local.esm.js": "./bundle/local.esm.js",
     "./providers/anthropic.esm.js": "./providers/anthropic.esm.js",
     "./providers/openai.esm.js": "./providers/openai.esm.js",
-    "./providers/local.esm.js": "./providers/local.esm.js"
+    "./providers/local.esm.js": "./providers/local.esm.js",
+    "./providers/chrome.esm.js": "./providers/chrome.esm.js"
   },
   "files": [
     "pip-core.esm.js",

package/providers/chrome.esm.js

@@ -0,0 +1,111 @@
+// Chrome's built-in Prompt API (on-device Gemini Nano / Gemma-derived).
+// Wraps `LanguageModel.create()` + `session.promptStreaming()` into a
+// runtime-compatible provider — zero-download for users on Chrome that
+// already has the weights, reply quality in the ~2B-effective-param
+// range (well above what transformers.js practically pulls in-browser).
+//
+// Usage:
+//   import { createRuntime } from '@nevescloud/pip/runtime.esm.js';
+//   import { chrome }        from '@nevescloud/pip/providers/chrome.esm.js';
+//
+//   const rt = createRuntime({ provider: chrome({ temperature: 0.1 }) });
+//
+// Surface shifted across Chrome versions: the API moved from
+// `window.ai.languageModel` (earlier flag-gated builds) to the top-level
+// `LanguageModel` constructor as the Prompt API spec settled. We try
+// the newer surface first and fall back. Chrome 138+ ships the origin
+// trial; ~Chrome 148+ runs without a flag for many origins. Non-Chrome
+// browsers throw a friendly error on first invocation.
+//
+// Limitations: no tool-use, no images. Tools registered on the runtime
+// won't be exposed — pip's turn loop still works for slash commands and
+// chat, but tool dispatch is a no-op with this provider.
+
+const UNAVAILABLE =
+  "Chrome's built-in AI isn't available here. Use Chrome 138+ or enable " +
+  "chrome://flags#prompt-api-for-gemini-nano on earlier versions.";
+
+function getApi() {
+  // Newer spec: top-level constructor. Older: nested under window.ai.
+  if (globalThis.LanguageModel) return globalThis.LanguageModel;
+  if (globalThis.ai?.languageModel) return globalThis.ai.languageModel;
+  return null;
+}
+
+async function ensureAvailable(LM) {
+  // Newer: availability() → 'available' | 'downloadable' | 'downloading' | 'unavailable'
+  // Older: capabilities() → { available: 'readily' | 'after-download' | 'no' }
+  if (typeof LM.availability === 'function') {
+    const v = await LM.availability();
+    if (v === 'unavailable') throw new Error(UNAVAILABLE);
+    return;
+  }
+  if (typeof LM.capabilities === 'function') {
+    const c = await LM.capabilities();
+    if (c?.available === 'no') throw new Error(UNAVAILABLE);
+  }
+}
+
+export function chrome({ systemPrompt, temperature, topK } = {}) {
+  return ({ messages, signal, system }) => (async function* () {
+    const LM = getApi();
+    if (!LM) throw new Error(UNAVAILABLE);
+    await ensureAvailable(LM);
+
+    // Runtime's per-call `system` wins over the factory default — same
+    // precedence anthropic/openai providers use.
+    const sys = system || systemPrompt;
+    const initialPrompts = [];
+    if (sys) initialPrompts.push({ role: 'system', content: sys });
+    // Replay prior turns. The Prompt API doesn't model tool dispatch, so
+    // skip non-string content (tool_use / tool_result turns) — they'd
+    // serialize to "[object Object]" and confuse the model.
+    for (const m of messages.slice(0, -1)) {
+      if (typeof m.content === 'string') {
+        initialPrompts.push({ role: m.role, content: m.content });
+      }
+    }
+    const tail = messages[messages.length - 1];
+    const userText = typeof tail?.content === 'string' ? tail.content : '';
+
+    const opts = {};
+    if (initialPrompts.length) opts.initialPrompts = initialPrompts;
+
+    // The Prompt API requires both topK and temperature to be set, or
+    // neither — passing one alone throws "Initializing a new session
+    // must either specify both topK and temperature, or neither". If
+    // the caller specified one, fetch the other's default from
+    // LanguageModel.params() so a half-spec doesn't reject the session.
+    const hasT = temperature != null;
+    const hasK = topK != null;
+    if (hasT || hasK) {
+      let t = temperature, k = topK;
+      if (hasT !== hasK && typeof LM.params === 'function') {
+        try {
+          const p = await LM.params();
+          if (!hasT) t = p?.defaultTemperature;
+          if (!hasK) k = p?.defaultTopK;
+        } catch {}
+      }
+      // Last-resort defaults if params() isn't available.
+      if (t == null) t = 1.0;
+      if (k == null) k = 40;
+      opts.temperature = t;
+      opts.topK = k;
+    }
+
+    const session = await LM.create(opts);
+
+    try {
+      for await (const chunk of session.promptStreaming(userText)) {
+        if (signal?.aborted) throw new DOMException('Aborted', 'AbortError');
+        if (typeof chunk === 'string' && chunk) {
+          yield { type: 'text_delta', text: chunk };
+        }
+      }
+      yield { type: 'turn_end', stopReason: 'end_turn' };
+    } finally {
+      try { session.destroy?.(); } catch {}
+    }
+  })();
+}

package/providers/local.esm.js

@@ -1,7 +1,7 @@
 // In-browser model via transformers.js + WebGPU. Two shapes ship:
 //
-// 1. `local({ model, dtype, maxTokens, genParams })` — runtime-compatible
-//    provider, slots into `createRuntime({ models: [{ provider: local(...) }] })`
+// 1. `local({ model, dtype, maxTokens, genParams, chatTemplate })` — runtime-
+//    compatible provider, slots into `createRuntime({ models: [{ provider: local(...) }] })`
 //    next to anthropic() and openai(). Wraps the renderer below and adapts
 //    its setReplyText-callback paint into the runtime's async-generator
 //    event protocol. Use this when local should participate in `/model`.
@@ -194,9 +194,13 @@ export function createTransformersRenderer() {
     await ensureLoaded(turnEl);
 
     const tf = await loadTransformers();
+    // chatTemplate spread last so consumers can override defaults (e.g.
+    // Gemma 4 needs { enable_thinking: false } to suppress channeled
+    // thought-token leaks; Qwen has its own templating knobs).
     const inputs = tokenizer.apply_chat_template(messages, {
       add_generation_prompt: true,
       return_tensors: 'pt',
+      ...(config.chatTemplate || {}),
     });
 
     const start = performance.now();
@@ -269,9 +273,9 @@ export function createTransformersRenderer() {
 // buffer (no new tokens emitted between calls), so the diff guards
 // against zero-length deltas. AbortSignal flows through naturally —
 // the underlying TextStreamer throws AbortError, which we surface.
-export function local({ model, dtype = 'q4', maxTokens = 256, genParams } = {}) {
+export function local({ model, dtype = 'q4', maxTokens = 256, genParams, chatTemplate } = {}) {
   const renderer = createTransformersRenderer();
-  if (model) renderer.setModel({ id: model, dtype, maxTokens, genParams });
+  if (model) renderer.setModel({ id: model, dtype, maxTokens, genParams, chatTemplate });
 
   return ({ messages, signal, turnEl, setReplyText }) => (async function* () {
     let lastFull = '';