A privacy-first RAG system for an entrepreneur who needed to ask questions of their own notes, PDFs, and research — without handing the archive to a vendor stack they couldn’t explain or own.
Notes in. Answers out. Infrastructure you can name.
Knowledge was already there — journals, project notes, PDFs, clipped research — but it wasn’t queryable as a whole. Cloud “AI over your docs” tools were either too opaque, too expensive at scale, or too casual with private material. The need wasn’t a chatbot demo. It was a system that stayed on hardware they controlled, kept improving as new notes arrived, and stayed cheap enough to run every day.
We treated the knowledge base like any other asset worth improving: understand what already existed, improve the pipeline, document the restore path, and share the pattern.
Notes and documents sync into a private git-backed store. Incremental indexers embed changed files into a local vector database. Workflow automation handles capture, clipping, scheduled research digests, and daily briefings. A chat interface queries the archive in natural language. Local models handle embeddings; generation routes to chosen APIs when speed and quality matter — with a clear path to keep more of that local later.
A working knowledge system: sub-second answers over private notes, automated research digests, and a documented upgrade path. The entrepreneur didn’t get a black-box subscription — they got infrastructure they can name, maintain, and improve.
Local embeddings plus careful routing beat “run everything on a GPU” for most small-business knowledge work. The hard part isn’t the model — it’s the pipeline: sync, index, capture, failure modes, and documentation so the system survives contact with Tuesday morning.