AI Document Management

Overview

An AI-powered document management system built for a hackathon — and the project that won it. The idea was to take the familiar shape of Google Drive (folders, files, sharing) and layer retrieval-augmented generation (RAG) on top, so a team can not just store documents but actually ask questions of them and get answers grounded in the source files rather than in a model’s general knowledge.

How it works

Uploaded documents are chunked, embedded, and stored in a Chroma vector database. At query time the system runs a similarity search to pull the most relevant chunks and passes them, as context, to a locally hosted language model through LangChain — retrieval-augmented generation instead of answering from the model’s own weights. Running the model locally means documents never leave the environment, which matters for the kind of internal, sensitive files a department keeps in a shared drive.

Access control

Access is governed by role-based access control (RBAC) scoped to both departments and individual users, so a single drive can hold each team’s documents while keeping each person’s view bounded to what they’re allowed to see.

A drive with AI built in

The whole thing is wrapped in a Google-Drive-style interface — browse, upload, and organize files — with the AI features (semantic search and document Q&A) built into the same workspace rather than bolted on as a separate tool.

Tech stack

• App — React frontend with a NestJS backend, TypeScript end to end.
• Retrieval — LangChain orchestrating a RAG pipeline over a Chroma vector store with similarity search.
• Model — a locally hosted LLM, so both inference and documents stay on-prem.
• Access — department- and user-level RBAC over storage and retrieval.