The problem
The research assistants worth using are hosted, metered, and opaque about what they searched. The self-hosted options mostly wrap a single model call and call it research.
Constraints
- Runs on my own hardware, in the lab, alongside everything else.
- Has to show its sources, because a synthesis you cannot check is worth nothing.
What I built
A deep research platform built on LangGraph, using a STORM-style approach to plan a topic, gather sources across several passes, and write up what it found. It runs self-hosted.
An earlier attempt at the same problem lives at ai-researchwizard, which runs as one of the services in the lab.
Outcome
Early, and marked rough for that reason. It runs self-hosted. Whether it beats the hosted tools is not a claim I can make yet.
ai-researchwizard is the one currently serving people in the lab. It is public and free to use, which was the point of building it.