"Upload Documents. Search Smarter. Get Source-Based Answers — Powered by Client-Side Wasm Retrievals."
A production-grade, serverless-friendly Retrieval-Augmented Generation (RAG) application. By vectorizing files and executing similarity searches entirely inside the browser, it completely eliminates backend database overhead, saves 100% of embedding API costs, and resolves the Gemini API quota rate limits.
DocuMind AI is an advanced document assistant built on Next.js 15, React 19, Tailwind CSS, and TypeScript. Rather than using typical backend vector stores that lead to API connection leaks, cold starts, and database bills, DocuMind AI implements a hybrid Stateless Local RAG Architecture.
The client extracts text from PDFs locally, splits it into overlapping chunks, generates 384-dimensional embeddings directly in the browser via WebAssembly (ONNX), and runs cosine similarity matching in-memory. The API route only receives the pre-retrieved context chunks to answer questions using Google Gemini 2.5 Flash, keeping the serverless backend completely stateless and fast.
🔗 https://docu-mind-ai-olive-six.vercel.app/
- 📂 Multi-PDF Parsing & Processing: Upload several PDF documents simultaneously. Parsing is performed locally via a fast Mozilla
pdf.jsworker. - 🧠 Local ONNX Embeddings (Strategy 1): Runs the
all-MiniLM-L6-v2transformer model directly inside the browser using WebAssembly. Model files cache in the user's browser IndexedDB on the first run, allowing subsequent uploads to run instantly with zero network load. - 📐 Adaptive Page-Aware Chunking (Strategy 3): Splitting is performed strictly within page boundaries, keeping citations 100% accurate. Chunk sizes scale dynamically depending on document page count (from
800up to2500characters) to keep the browser memory footprint low. - 🔎 Offline Semantic Search Engine: Toggle from chat to a dedicated vector search UI. Query the document and retrieve matched chunks in real-time, displaying cosine similarity percentages, page numbers, and filenames.
- 💬 5 Tailored AI Mode Personas:
- Universal AI 🌐: General document intelligence, summaries, and Q&A.
- StudyMate AI 📚: Academic revisions, explanations, revision notes, and exam questions.
- ResearchMind AI 🔬: Extracts methodology, findings, paper limits, and future works.
- ResumeVault AI 📄: Evaluates candidate resumes, extracts skills, and analyzes role fitness.
- BusinessInsight AI 💼: Corporate report parser extracting KPIs, risk factors, and summaries.
- 🏷️ Interactive Click-to-Pulse Citations: Responses contain tiny circular numbered badges linked to sources. Hovering shows context snippets in tooltips; clicking scrolls the window to the reference details card and highlights it with a pulsing animation.
- 📊 Live Document Statistics: Floating dashboard tracks uploaded document count, total page count, chunk counts, generated embeddings, and user queries.
- 🌌 Ambient Aurora Glow Backdrop: Slow-drifting blurred gradient background blobs styled behind premium glassmorphic cards.
- 👥 Developer Footer: Beautiful profile footer section linking to developer Jay Jadhav's GitHub, LinkedIn, and Portfolio.
sequenceDiagram
autonumber
actor User as User Browser
participant Client as Client-side State
participant ONNX as Local ONNX Engine
participant API as Next.js API Route
participant LLM as Google Gemini API
User->>Client: Drag & Drop PDF files
Client->>Client: Extract text & perform adaptive chunking
Note over Client, ONNX: Strategy 1: Check browser IndexedDB for cached model
ONNX-->>Client: Load model (cached) or download (~30MB)
Client->>ONNX: Send chunks for vector extraction
ONNX-->>Client: Return 384-dimension embedding vectors
User->>Client: Type query and send
Client->>ONNX: Vectorize query string
Client->>Client: Perform offline cosine similarity match against chunk vectors
Client->>API: POST /api/chat (Send top 5 context text chunks + citations + query)
API->>LLM: Send constructed system instruction prompt + context
LLM-->>API: Return markdown response containing citations [1], [2]
API-->>Client: Send response payload
Client->>User: Render Chat UI + Scroll-to-pulse citations
A premium glassmorphic workspace interface utilizing slow-moving ambient aurora blobs, displaying upload cards, statistics tracker, and AI Mode buttons.
High-contrast bubble text for dark mode readers. Shows compact source pills that pulse-highlight upon citation badge click.
Modern links containing Lucide icons with smooth transitions pointing to GitHub, LinkedIn, and Portfolio.
- Node.js: v18.0.0 or higher
- Package Manager:
npm(included with Node) - Google Gemini API Key: Grab one from Google AI Studio
- Clone the Repository:
git clone https://github.com/your-username/documind-ai.git cd documind-ai - Install Project Dependencies:
npm install
- Environment Variables Configuration:
Create a
.env.localfile in the root folder:# Add your Google Gemini API Key below GOOGLE_API_KEY=AIzaSy...your_gemini_api_key_here
- Run Development Server:
npm run dev
- Open in Browser: Navigate to http://localhost:3000 inside your web browser.
DocuMind AI is engineered to run 100% database-free, making it extremely cost-effective and easy to host on Vercel serverless pipelines.
- Install the Vercel CLI or import the repository on the Vercel Dashboard:
npm install -g vercel vercel
- Add Project Environment Variables:
Under your Vercel Project Settings, add:
GOOGLE_API_KEY: Your Google Gemini API Key.
- Build & Launch: Vercel will build the Next.js production package. The client-side ONNX indexing architecture ensures your app avoids Vercel's payload restrictions and the 10-second serverless execution limits.
- Framework: Next.js 15 (App Router)
- Runtime Logic: React 19, TypeScript
- Styling: Tailwind CSS v3, Glassmorphism, CSS Blob Keyframe Animations
- Icons: Lucide React
- Local Embedding Engine:
@xenova/transformers(Wasm-compiledall-MiniLM-L6-v2) - Local PDF Parser: Mozilla
pdf.jsworker CDN - Large Language Model: Google Gemini 2.5 Flash SDK
- Deployment Target: Vercel
Add these high-impact, ATS-optimized bullet points directly to your software engineering resume:
- Stateless Browser RAG Architecture: Engineered a production-ready, database-free RAG application using Next.js 15, React 19, and TypeScript, generating embeddings and similarity indexes entirely on the client side via WebAssembly (ONNX).
- API Cost Optimization & Rate-Limit Prevention: Eliminated 100% of server-side embedding API fees and prevented Gemini 429 quota limits by implementing client-side vector search, caching the model (~30MB) locally in browser IndexedDB.
- Adaptive Page-Aware Chunking Engine: Coded an adaptive chunking utility in TypeScript that automatically scales chunk boundaries (from 800 to 2500 characters) depending on file page counts to optimize client memory limits and ensure 100% citation accuracy.
- Real-time Cosine Similarity search: Authored a custom in-memory vector search library to execute cosine similarity calculations locally, reducing payload sizes sent to serverless Next.js functions from megabytes to kilobytes.
- Interactive Source Citation Navigation: Designed a high-fidelity UI using Tailwind CSS featuring dynamic dark mode contrast syncing, floating ambient blur animations, and custom scroll-to-pulse inline citation navigation.
Designed & Developed by Jay Jadhav.
- GitHub: github.com/jayjadhav
- LinkedIn: linkedin.com/in/jayjadhav
- Portfolio: jayjadhav.dev