Featuredpersonal2025

LumiDoc - AI-Powered Document Q&A

lumidoc.vercel.app

Overview

Project Overview

LumiDoc is an AI-powered document intelligence platform that lets users "chat" with their files. Upload a PDF, Word doc, spreadsheet, slide deck, or scanned image, and ask questions in plain English — LumiDoc returns accurate answers with citations pointing back to the exact source in your documents. Built end-to-end as a solo project using a production-grade RAG (Retrieval-Augmented Generation) pipeline. Documents are processed asynchronously, split into semantic chunks, embedded as vectors, and retrieved using hybrid search (semantic + keyword). Answers stream in real time from Anthropic's Claude models. Tech stack: Next.js 15, TypeScript, Tailwind CSS, Supabase (PostgreSQL + pgvector), Anthropic Claude, and Voyage AI embeddings deployed on Vercel.

Case Study

Problem → Solution

Problem

Professionals and teams sit on large volumes of documents — contracts, reports, research papers, financial spreadsheets — but finding a specific answer means manually scrolling through dozens of pages. Traditional keyword search returns matching files, not actual answers, and generic AI chatbots can't read private documents or cite where information came from. The result: hours lost to manual searching and no way to trust an AI's answer without verifying the source.

My Role

Want to build something like this?

Book a free discovery call and let's scope your project — no commitment required.

Available for workTypical response: <24h

Book a Free Call Browse All Projects

Solution

Built a full-stack RAG application where users upload documents that are processed through an asynchronous pipeline: text extraction per file type, semantic chunking (512 tokens), and vector embedding via Voyage AI stored in Supabase pgvector. At query time, a hybrid search engine combines 70% semantic vector similarity with 30% full-text search — enhanced by HyDE query expansion to retrieve the most relevant context. Answers are generated by Anthropic's Claude and streamed to the UI with inline source citations, so every response is verifiable. Auth, storage, and row-level security are handled through Supabase, with the whole system deployed serverlessly on Vercel.

Outcome

Document search time: Minutes of manual scrolling → seconds with cited answers · Supported file formats: 6+ formats (PDF, DOCX, XLSX, PPTX, images, text) · Answer accuracy: Source-cited responses — every answer traceable to the document · Retrieval method: Hybrid search (vector + full-text) for higher relevance than keyword-only

Results

Impact Metrics

Minutes of manual scrolling → seconds with cited answers

Document search time

6+ formats (PDF, DOCX, XLSX, PPTX, images, text)

Supported file formats

Source-cited responses — every answer traceable to the document

Answer accuracy

Hybrid search (vector + full-text) for higher relevance than keyword-only

Retrieval method