Suggested Request: Swap MediaPipe GenAI with LiteRT LM web

BLOG.md

@@ -8,7 +8,7 @@ By leveraging the raw power of the **Snapdragon X Elite** and its high-performan
 
 ## 🚀 The Vision: AI at the Edge
 
-Most AI-integrated IDEs rely on heavy cloud APIs. This introduces latency, subscription costs, and—most importantly—privacy concerns. **Nova IDE** flips the script. It uses **MediaPipe GenAI** and **WebGPU** to run Large Language Models (LLMs) locally.
+Most AI-integrated IDEs rely on heavy cloud APIs. This introduces latency, subscription costs, and—most importantly—privacy concerns. **Nova IDE** flips the script. It uses **LiteRT LM** and **WebGPU** to run Large Language Models (LLMs) locally.
 
 When you run Nova IDE on a machine powered by the **Snapdragon X Elite**, you're not just running a web app; you're utilizing one of the most efficient NPU/GPU architectures ever designed for portable computing.
 
@@ -29,7 +29,7 @@ graph TD
     end
 
     subgraph AI_Runtime ["AI Execution Layer"]
-        MP["MediaPipe GenAI (WASM)"]
+        MP["LiteRT LM (WASM)"]
         WG["WebGPU API"]
     end
 
@@ -47,8 +47,8 @@ graph TD
     MP -->|Streaming Tokens| UI
 ```
 
-### 1. The Inference Engine (MediaPipe + WebGPU)
-At the heart of Nova IDE is the MediaPipe GenAI runtime. Unlike traditional JavaScript which runs on the CPU, Nova IDE uses **WebGPU** to talk directly to the **Qualcomm Adreno GPU**. 
+### 1. The Inference Engine (LiteRT LM + WebGPU)
+At the heart of Nova IDE is the LiteRT LM runtime. Unlike traditional JavaScript which runs on the CPU, Nova IDE uses **WebGPU** to talk directly to the **Qualcomm Adreno GPU**. 
 - **WebGPU** allows for massively parallel tensor operations required by transformers.
 - On the Snapdragon X Elite, the Adreno GPU provides the floating-point performance needed to generate tokens at lightning speed, rivaling cloud-based solutions.
 
@@ -74,9 +74,9 @@ The Snapdragon X Elite is a breakthrough for web-based AI. While Intel and AMD h
 
 When you click **"⚡ Load Local AI Model"** in Nova IDE, the following sequence occurs:
 
-1. **WASM Initialization**: The IDE loads the MediaPipe GenAI WebAssembly runtime.
+1. **WASM Initialization**: The IDE loads the LiteRT LM WebAssembly runtime.
 2. **GPU Adapter Request**: The browser requests a WebGPU adapter. On an X Elite machine, this identifies the **Qualcomm Adreno GPU**.
-3. **Model Loading**: The model (in `.task` format) is fetched into a `SharedArrayBuffer`.
+3. **Model Loading**: The model (in `.litertlm` format) is fetched into a `SharedArrayBuffer`.
 4. **GPU Compilation**: The model's computation graph is compiled into GPU-specific kernels.
 5. **Streaming Inference**: When you type a prompt, the tokens are generated on the Adreno GPU and streamed back to the UI in real-time.
 

README.md

@@ -1,6 +1,6 @@
 # ⚡ Nova IDE
 
-Nova IDE is a lightweight, browser-based coding environment designed for the future of **on-device AI**. It leverages WebGPU and MediaPipe to run high-performance Large Language Models (LLMs) entirely within your browser—no cloud, no APIs, and total privacy.
+Nova IDE is a lightweight, browser-based coding environment designed for the future of **on-device AI**. It leverages WebGPU and LiteRT LM to run high-performance Large Language Models (LLMs) entirely within your browser—no cloud, no APIs, and total privacy.
 
 ![Nova IDE Screenshot](src/assets/hero.png)
 
@@ -60,7 +60,7 @@ Nova IDE is a lightweight, browser-based coding environment designed for the fut
 
 - **Core**: JavaScript (ESM), HTML5, CSS3
 - **Editor**: [CodeMirror 6](https://codemirror.net/)
-- **AI Engine**: [MediaPipe GenAI](https://ai.google.dev/edge/mediapipe/solutions/genai/llm_inference)
+- **AI Engine**: [LiteRT LM](https://ai.google.dev/edge/litert-lm/js)
 - **Bundler**: [Vite](https://vitejs.dev/)
 - **Storage**: IndexedDB (Virtual File System)
 

package.json

@@ -20,7 +20,7 @@
     "@codemirror/lang-python": "^6.2.1",
     "@codemirror/lint": "^6.9.6",
     "@codemirror/theme-one-dark": "^6.1.3",
-    "@mediapipe/tasks-genai": "^0.10.27",
+    "@litert-lm/core": "^0.12.1",
     "codemirror": "^6.0.2"
   }
 }

packages/web-agent-core/README.md

@@ -2,7 +2,7 @@
 
 A purely in-browser, WebGPU-accelerated, zero-dependency autonomous ReAct (Reason+Act) Agent Framework.
 
-Built for local-first, privacy-respecting AI applications using MediaPipe GenAI and lightweight models like Gemma.
+Built for local-first, privacy-respecting AI applications using LiteRT LM and lightweight models like Gemma.
 
 ## Features
 - **Zero Backend**: Runs entirely in the browser using WebGPU.
@@ -41,7 +41,7 @@ const myTools = [
 // Your LLM object just needs a \`generateRaw(history, onToken)\` method
 const aiEngine = {
   generateRaw: async (prompt, onToken) => {
-    // Call MediaPipe or Transformers.js here
+    // Call LiteRT LM, Transformers.js, or another local model runtime here
     // stream tokens to onToken(token)
   }
 };

packages/web-agent-core/package.json

@@ -15,7 +15,7 @@
     "agent",
     "llm",
     "webgpu",
-    "mediapipe",
+    "litert-lm",
     "react",
     "autonomous",
     "browser",
@@ -26,4 +26,4 @@
   "publishConfig": {
     "registry": "https://npm.pkg.github.com"
   }
-}
+}

packages/web-agent-core/website/docs/intro.md

@@ -6,7 +6,7 @@ sidebar_position: 1
 
 Welcome to **Nova Web Agent Core**!
 
-This package provides a purely in-browser, WebGPU-accelerated, zero-dependency autonomous ReAct (Reason+Act) Agent Framework. It is built for local-first, privacy-respecting AI applications using MediaPipe GenAI and lightweight models like Gemma.
+This package provides a purely in-browser, WebGPU-accelerated, zero-dependency autonomous ReAct (Reason+Act) Agent Framework. It is built for local-first, privacy-respecting AI applications using LiteRT LM and lightweight models like Gemma.
 
 ## Why this framework?
 
@@ -80,7 +80,7 @@ Your LLM wrapper only needs one method: `generateRaw(prompt, onTokenCallback)`.
 ```javascript
 const aiEngine = {
   generateRaw: async (prompt, onToken) => {
-    // Call MediaPipe, Transformers.js, or even a cloud API here
+    // Call LiteRT LM, Transformers.js, or even a cloud API here
     // stream tokens to onToken(token)
   }
 };

packages/web-agent-core/website/src/components/HomepageFeatures/index.js

@@ -17,7 +17,7 @@ const FeatureList = [
     icon: '🧠',
     description: (
       <>
-        Powered by the cutting-edge WebGPU backend and MediaPipe, execute your agent's reasoning loops completely client-side without API keys.
+        Powered by the cutting-edge WebGPU backend and LiteRT LM, execute your agent's reasoning loops completely client-side without API keys.
       </>
     ),
   },