DocuMind: AI Document Assistant
A RAG assistant that lets support teams answer from 10,000 pages of docs with cited, trustworthy answers.
- Year
- 2024
- Timeline
- 8 weeks
- Category
- AI Application
- −74%
- Time to first answer
- 92%
- Answers rated helpful
- 100%
- Answers cite a source
The Problem
A support team spent hours hunting docs to answer customers, and generic chatbots hallucinated. They needed answers grounded in their own documents, with sources.
The Solution
DocuMind: a RAG assistant. Docs are chunked, embedded, and retrieved to ground an OpenAI answer that always cites sources; low-confidence answers flag for review.
My Role
AI engineer and architect. Built the ingestion and retrieval pipeline, tuned chunking and prompts, and shipped the streaming Next.js chat UI with citations.
Tech Stack
- Next.js
- Node.js
- OpenAI
- PostgreSQL
- Prisma ORM
- Vector Search
- Docker
- AWS
How it fits together.
The high-level architecture: services, data stores, and the external systems they talk to.
Approach
RAG fails in production on retrieval quality, not the model. I focused on chunking and metadata filtering so the right passages surface, and forced the model to answer only from retrieved context, or refuse. That refusal behavior earned trust.
Highlights
- Every answer links to its exact source passage.
- Low-confidence answers flag for human review.
- Incremental ingestion: new docs embed without reprocessing the corpus.
Screens
