// the find
pashpashpash/vault-ai
OP Vault ChatGPT: Give ChatGPT long-term memory using the OP Stack (OpenAI + Pinecone Vector Database). Upload your own custom knowledge base files (PDF, txt, epub, etc) using a simple React frontend.
A self-hosted RAG system where you upload documents (PDF, epub, docx, txt) and ask questions against them. Go backend handles chunking and embedding via OpenAI, Pinecone or Qdrant stores the vectors, React frontend wraps it. Built in 2023 when this pattern was novel; now it's table stakes.
Go backend is a reasonable choice here — low overhead, easy to deploy as a single binary alongside the React build. The vectordb abstraction layer (pinecone/ and qdrant/ subdirs behind a common interface) means you can swap backends without touching the query logic. Source citation is included in answers — you see the filename and character offsets of the chunks that informed the response, which is the most important feature for a document Q&A tool. The chunking and upsert pipeline is straightforward to follow and modify.
Secrets are stored in flat files under a `secret/` directory, which is gitignored but a bad habit — no support for env vars or a secrets manager out of the box. The 3MB per-file default is laughably small for the 'upload an entire library' pitch; the README tells you to edit a Go constant to raise it, which means recompiling. No auth layer — whoever can reach port 8100 can upload documents and query your OpenAI key. The project has been essentially frozen since 2023: dependencies are unpinned, node v19 is EOL, and the toolchain (Go 1.18, webpack without a version lock) will require archaeology to reproduce a working build.