Doc Research OS

What it is

A local document research tool: you load papers and course material into it, and query them in a chat interface backed by retrieval over your own documents rather than the model’s memory. Built because studying under load means re-reading the same material repeatedly to find one paragraph.

What runs today

Ingestion routes PDFs, Office files, and plain text through a single endpoint into a ChromaDB vector store, chunked at 512 tokens with 100-token overlap. Queries run with retrieval on or off; when on, the retrieved chunks are prepended to the message as a text block rather than injected as a system message, which keeps the model’s instruction context and the document context separate.

The stack is a FastAPI backend and a React frontend under Docker Compose, with embeddings served through OpenRouter. It runs locally on macOS under Colima.

Where it’s going

The intended direction is a study tool rather than a chat wrapper: material organized by course or topic, retrieval that knows which source a claim came from, and eventually generated questions and review material from documents already loaded. The longer-term version of this is an educational platform rather than a personal tool.

None of that is built. What exists is the retrieval layer it would sit on.

Status

Active but not prioritized. Anchor comes first; Doc gets attention when it stops being good enough for what I need from it.

The open question I’d want to answer before building anything on top is whether retrieval quality is actually good enough — chunk size, overlap, and the retrieval threshold are currently reasonable defaults rather than measured choices, and I have no evaluation set to tell me otherwise.