Self-Hosted RAG Document API

The problem

Asking questions about private documents means pasting them into a hosted chat tool, which sends them to someone else's server. Small local models fix that but cannot take a whole document, so something has to decide which passages they see.

What I built

• Documents are split into overlapping chunks and embedded locally with Ollama.
• Retrieval is filtered by owner, so one person's chunks never leak into another's answer.
• Answers stream back as newline-delimited JSON, so cited chunks appear before the first token.
• Everything except the two model calls is deterministic Python.
• Includes a browser UI and a guide to running Ollama on your own machine.