DocuMind

What it does

DocuMind turns a static PDF into a searchable, queryable knowledge source:

• Secure document management — every document record and chunk is owned by a user and every read/write is filtered by the authenticated user id.
• PDF ingestion — uploads are validated (PDF only, non-empty, ≤ 10 MB) and parsed server-side.
• Text extraction — text is extracted per page so page boundaries are preserved.
• Document chunking — page text is normalized and split into overlapping, page-aware chunks.
• Semantic search / retrieval — each chunk is embedded into a dense vector; a query is embedded the same way and ranked against the document's chunks using cosine similarity.
• Evidence-backed document Q&A — only the retrieved excerpts are supplied to the answer step, and the response cites the supporting chunks.
• Source / page references — citations returned to the client carry the chunk id and the page range the evidence came from.

How it works

1. PDF upload — POST /api/documents accepts a multipart/form-data upload. The file is rejected unless it is a non-empty PDF no larger than 10 MB.
2. Text extraction — the PDF buffer is parsed page by page, and the text items of each page are concatenated. Pages without extractable text produce an explicit ingestion error.
3. Page-aware chunking — each page's text is whitespace-normalized and split into chunks of up to 1200 characters with a 200-character overlap, so a chunk never spans pages and page numbers remain accurate.
4. Semantic embeddings — every chunk is embedded into a dense float vector. Embeddings are generated with bounded concurrency so ingestion stays within provider rate limits.
5. MongoDB storage — chunks (with their vectors) and the document metadata are persisted, each tied to the owning userId.

Features

• Area: Capability
• Authentication: Email + password accounts, bcrypt password hashing, signed JWT session in an HttpOnly cookie, server-side session lookup, logout
• Document management: Upload, list, open, rename, and delete documents; per-user chunk counts
• PDF pipeline: Real PDF text extraction with page awareness, overlapping chunking, batch embedding generation
• Retrieval: Query embedding, cosine similarity ranking, configurable topK, minimum similarity threshold
• Grounded answers: Responses constrained to retrieved excerpts, citation validation, explicit "not enough evidence" answer
• Chunk browsing: Inspect stored chunks per document without exposing embedding vectors
• Ownership enforcement: All document, chunk, and RAG queries are scoped by userId
• UI: Responsive dashboard, workspace, and settings screens built on the existing Stitch design system

Technology

• Layer: Technology
• Framework: Next.js 15 (App Router, Route Handlers, Server Actions-less API routes)
• Language: TypeScript 5.9 (strict mode)
• UI: React 19, Tailwind CSS 4, motion, lucide-react, class-variance-authority
• Database: MongoDB with Mongoose 9 (cached global connection pool)
• Auth: jsonwebtoken (HS256 JWT) + bcryptjs password hashing
• Document processing: pdf-parse for page-level text extraction
• Embeddings / generation: Server-side calls to a managed embedding + text generation API (keyed by GEMINI_API_KEY)
• Testing: Vitest
• Linting: ESLint with eslint-config-next
• Hosting: Vercel (serverless) with MongoDB Atlas