🧠 DocuMind AI: Evidence-First RAG for Document Intelligence

🚀 View Live Demo | 📂 View Codebase

Modern AI Orchestration | Next.js 15 | Verifiable Citations

DocuMind AI is a professional Retrieval-Augmented Generation (RAG) platform designed to transform static PDFs into interactive, grounded intelligence. Built in March 2026, this project focuses on AI reliability and source transparency, ensuring every response is backed by specific evidence from the uploaded documentation.

---

✨ Key Features

• Evidence-First Chat: Implemented a metadata-driven retrieval system that provides "Source Pills" for every response, allowing users to verify AI claims against raw PDF text.
• Streaming Serverless Architecture: Optimized for Vercel/Serverless environments using the Vercel AI SDK to bypass 10-second execution timeouts via real-time token streaming.
• Zero-Hallucination Guardrails: Engineered strict system prompts and semantic similarity thresholds to ensure the model only answers based on provided context.
• Session-Based Memory: Utilizes a secure, in-memory vector store that isolates document context to the current browser session, maintaining data privacy without persistent database costs.

---

🛠️ Tech Stack

• Frontend: Next.js 15 (App Router), Tailwind CSS, Shadcn/UI (Nova Glassmorphic Theme)
• AI Orchestration: LangChain.js & Vercel AI SDK
• LLM & Embeddings: OpenAI gpt-4o-mini & text-embedding-3-small
• PDF Processing: pdf2json (Server-side text extraction)
• Vector Storage: MemoryVectorStore (Cosine Similarity Search)

---

🏗️ Technical Challenges & Solutions

1. The "DOMMatrix" Server-Side Conflict

Challenge: Standard PDF libraries (like pdf-parse) often rely on browser-only Canvas APIs, causing DOMMatrix is not defined crashes in Next.js 15 server environments. Solution: Refactored the ingestion pipeline to use pdf2json in a strict text-only mode with globalThis polyfills. This ensured stable server-side parsing without the overhead of browser-dependent renderers or Canvas dependencies.

2. Stream Protocol Alignment

Challenge: Mismatches between raw text streams and the AI SDK's expected data protocol often result in "empty UI" responses despite successful network calls. Solution: Aligned the server-side toTextStreamResponse with custom headers (x-vercel-ai-data-stream: v1) and configured the frontend hooks with streamProtocol: "text" to bridge the protocol gap and enable real-time UI rendering.

3. The "Subway Slicer" Logic (Chunking)

Challenge: Passing an entire PDF to an LLM is costly and leads to "lost-in-the-middle" accuracy issues. Solution: Implemented a RecursiveCharacterTextSplitter with a 1000-character chunk size and 200-character overlap. This preserves semantic context across chunk boundaries, significantly improving retrieval accuracy for complex queries.

---

⚠️ Infrastructure Limitation

Warning

DocuMind AI utilizes an Ephemeral In-Memory Store (MemoryVectorStore) for its RAG engine.

Due to the stateless nature of Vercel Serverless Functions, the application "forgets" the document context immediately after ingestion in the deployed environment. This results in the AI responding with "I don't know" and missing source pills during live chat.

To verify this test locally (where the memory persists), follow the Getting Started instructions below.

For the full persistent experience (Cloud Storage, Multi-Session Retrieval, and Hardened Monitoring), please visit 🛡️ Sentinel Docs — a production-grade evolution of this project utilizing Upstash Vector for permanent cloud memory.

---

🧪 Verification: The "Neon Pink Moon" Test

To ensure the RAG engine is 100% grounded and ignoring its own pre-trained biases, I utilize a custom verification suite:

1. Ingestion: A PDF containing "nonsense" facts (e.g., "The moon is Neon Pink" or "Apple's CEO is a Golden Retriever named Sparky") is uploaded.
2. Retrieval: The system is queried: "What color is the moon?"
3. Validation: The app is verified only if it retrieves the "Neon Pink" chunk and ignores its general training data, proving the Semantic Search and Prompt Grounding are functional.

The RAG engine successfully identifies the "Neon Pink" moon color, bypassing general LLM training.

---

📋 Sample Test Data

To verify the grounding of the RAG engine, I utilized the following "nonsense" data points in a test PDF. This ensures the model is retrieving specific context rather than relying on its pre-trained general knowledge:

Document Content: "The official color of the moon is Neon Pink. The CEO of Apple is a Golden Retriever named Sparky. To reset your password, you must dance for 30 seconds."

Test Queries to run:

1. "What is the official color of the moon?" (Expect: Neon Pink)
2. "Who is the CEO of Apple?" (Expect: Sparky the Golden Retriever)
3. "How do I reset my password?" (Expect: Dance for 30 seconds)

---

🚦 Getting Started

1. Clone & Install:

   npm install

2. Environment Setup: Create a .env.local in the root and add your OpenAI key:

   OPENAI_API_KEY=sk-proj-xxxx...

3. Run Development:

   npm run dev

---

🧭 Engineering Philosophy

As an engineer with nearly 5 years of experience in security environments (Trend Micro), I believe AI should be a "glass box," not a "black box." DocuMind AI demonstrates my ability to build verifiable, cost-efficient, and secure AI systems.