RAGStem

The problem

Accountants sit on exactly the kind of archive retrieval was built for: years of tax manuals, client financials, filings, regulatory guidance. And exactly the kind nobody will upload to a chatbot. Client financial records go through a cloud service only if someone can explain, to a partner and eventually to a regulator, precisely what was sent and where it went. For most firms that conversation ends the project — so the documents stay in folders, and finding anything means remembering where it is.

The requirement was uncomfortable: give them semantic search and grounded answers over their own archive, without their documents leaving their computers.

A desktop application that keeps the entire knowledge base on the user’s machine.

Indexing runs entirely on-device — documents are parsed, chunked, and embedded locally through a quantized model running in-process, with an embedded vector store. No server, no external search engine, no upload. Nothing about indexing touches the network.

Only the answer step reaches the cloud, and only barely. When a question is asked, retrieval happens locally against the local vectors. Just the handful of passages that actually answer the question, plus the question itself, are sent to the model — no identity, no metadata, no document, no archive. Every answer comes back with inline citations pointing at the exact chunks it used, so the user can see precisely what was shared and check the answer against the source.

The user supplies their own API key, encrypted at rest through the operating system’s keychain — the firm’s key, the firm’s account, the firm’s control. The desktop app itself is locked down: sandboxed renderer, isolated context, and a single typed, allowlisted bridge between the interface and the system rather than a browser with filesystem access.