Buddhi AI: Private, On-Device Academic Intelligence & Research Platform

Buddhi AI is an open-source, client-side academic research platform and intelligent comprehension environment running 100% privately in the browser. Powered by Google LiteRT, client-side language and embedding models, in-browser PGlite vector storage, and Recursive Language Models (RLM), Buddhi AI eliminates server computation costs while guaranteeing complete data confidentiality.

---

Key Highlights

• 100% Client-Side Privacy: Unpublished manuscripts, proprietary datasets, and personal reading notes never leave your device. All inference and vector queries execute locally.
• On-Device LiteRT Runtime: Powered by @litert-lm/core and @litertjs/core running on WebGPU with WebAssembly fallbacks.
• Local Language & Embedding Models: Native support for Gemma 4 E2B (litert-community/gemma-4-E2B-it-litert-lm) and EmbeddingGemma 300M (litert-community/embeddinggemma-300m).
• In-Browser PGlite Vector Store: Embedded PostgreSQL running in WebAssembly via @electric-sql/pglite and @electric-sql/pglite-pgvector for instant semantic vector search and paper chunk indexing.
• Recursive Language Models (RLM): Intelligent recursive document decomposition, literature matrix synthesis, and hierarchical paper summarization.
• Client-Side Execution Sandbox: Direct integration with @buddhilive/sandbox and @buddhilive/sandbox-sw for zero-backend, client-side Next.js code execution and live interactive previews.
• Academic Research Tools: Integrated PDF extraction (pdfjs-dist), citation generation (APA, IEEE, ACM, BibTeX), and literature review drafting.

---

System Architecture

graph TD
    User(("Researcher")) -->|Upload Paper / Prompt| UI["Buddhi AI Research Workspace"]

    subgraph Browser_Client_Side ["Browser Environment - 100% Client-Side"]
        direction TB

        subgraph Ingestion_Layer ["Ingestion & Processing"]
            PDF["PDF / Document Parser"] --> CHUNK["Semantic Text Splitter"]
            CHUNK --> EMB["LiteRT Embedding Worker<br/>EmbeddingGemma-300M"]
        end

        subgraph Storage_Layer ["In-Browser Storage"]
            EMB -->|Vector Embeddings| PGLITE[("PGlite WASM<br/>pgvector Store")]
            PDF -->|Document Cache| IDB[("IndexedDB Storage")]
        end

        subgraph Reasoning_Layer ["Cognitive & Reasoning Pipeline"]
            PGLITE -->|Semantic Retrieval| RAG["RAG Engine"]
            RAG --> RLM["Recursive Language Model<br/>RLM Service"]
            RLM --> CTX["Context Assembler & Prompts"]
        end

        subgraph Inference_Layer ["On-Device Model Execution"]
            CTX --> LITERT["Google LiteRT WebGPU Runtime"]
            LITERT --> GEMMA["Gemma 4 E2B<br/>Chat Template v4"]
        end

        subgraph Sandbox_Layer ["Client Execution Sandbox"]
            GEMMA -->|Generated Code / Previews| SB["@buddhilive/sandbox<br/>Service Worker Runtime"]
        end
    end

    GEMMA -->|Streaming Token Responses| UI
    SB -->|Live In-Browser Preview| UI

Loading

---

Core Capabilities

1. On-Device LiteRT AI Engine

Buddhi AI utilizes Google's LiteRT runtime (@litert-lm/core, @litertjs/core) for hardware-accelerated local inference:

• Gemma 4 E2B: Fast edge-optimized language model running in WebGPU mode, utilizing official gemma4 channel chat templates.
• EmbeddingGemma 300M: On-device vector embeddings for academic paper chunking, literature similarity scoring, and dense retrieval.
• Zero-Cloud Dependency: Downloaded models are cached in the browser's origin storage; no external API keys or server quotas needed.

2. In-Browser PGlite Vector Store

Rather than sending paper text to remote vector databases, Buddhi AI runs a full SQL database in WebAssembly:

• Powered by @electric-sql/pglite and @electric-sql/pglite-pgvector.
• Generates Euclidean and cosine similarity vector queries directly within browser memory.
• Fast, transactional storage of paper sections, abstracts, methodology notes, and citation references.

3. Recursive Language Models (RLM)

Handling long scientific papers requires hierarchical digestion:

• rlm-service.ts coordinates recursive summarization and literature matrix generation.
• Breaks down complex 50+ page manuscripts into interconnected conceptual sections.
• Synthesizes findings across multiple publications to isolate methodology biases and research gaps.

4. In-Browser Sandbox Execution

Buddhi AI embeds @buddhilive/sandbox and @buddhilive/sandbox-sw:

• Compiles, bundles, and previews generated Next.js and React code entirely client-side.
• Service Worker routes requests dynamically inside the browser without spinning up any Node.js container or remote cloud environment.

---

Supported Local Models

Model	Type	Architecture	File	Provider
Gemma 4 E2B	Language	WebGPU / WASM	gemma-4-E2B-it-web.litertlm	litert-community/gemma-4-E2B-it-litert-lm
EmbeddingGemma 300M	Embedding	WebGPU / WASM	embeddinggemma-300M_seq2048_mixed-precision.tflite	litert-community/embeddinggemma-300m

Note: EmbeddingGemma requires a Hugging Face User Access Token due to repository gating.

---

Tech Stack

• Framework: Next.js 16 (App Router)
• Frontend Core: React 19, TypeScript 5
• On-Device AI Engine: Google LiteRT (@litert-lm/core, @litertjs/core)
• Vector Database: PGlite (@electric-sql/pglite, @electric-sql/pglite-pgvector)
• Sandbox Environment: @buddhilive/sandbox, @buddhilive/sandbox-sw
• Styling: Tailwind CSS v4, Motion (Framer Motion v12)
• Document Processing: pdfjs-dist, @langchain/textsplitters
• Markdown & Streaming: streamdown, @streamdown/code, @streamdown/math, @streamdown/mermaid
• State Management: Zustand v5

---

Developer Quickstart

Prerequisites

• Node.js: >= 20.x
• Package Manager: pnpm >= 9.x
• Browser: Modern Chromium browser with WebGPU enabled (Chrome 113+, Edge 113+, or Brave).

1. Clone the Repository

git clone https://github.com/Buddhilive/buddhi-ai.git
cd buddhi-ai

2. Install Dependencies

During install, postinstall scripts automatically copy required Service Worker and LiteRT WASM binaries to public/:

pnpm install

(Under the hood, this executes node scripts/copy-sw.js && node scripts/copy-litert-wasm.js)

3. Launch Development Server

pnpm dev

Open http://localhost:3000 in your browser.

4. Build for Production

pnpm build

---

Development Scripts

• pnpm dev: Start Next.js development server with Turbopack.
• pnpm build: Create optimized production build.
• pnpm start: Start production server.
• pnpm lint: Run ESLint checks across all TypeScript files.

---

License

This project is licensed under the MIT License - see the LICENSE file for details.