Skip to content
Open Net Worth Docs

Performance

Most screens open in well under a second, even on a large book. When something feels slow, it’s usually one of three things: the way the app is being run, the AI model Lucius is using, or one particularly heavy report. Here’s how to tell which, and what to do.

A screen is slow

  1. Check how you’re running it. The desktop app, onw start, Docker and make start all serve the finished, optimized web app — nothing to do. If you run from a source checkout with the developer setup (npm run dev and uvicorn --reload), every page loads hundreds of unoptimized files, which is genuinely slow, especially on a phone. For everyday use, run make setup start (or npm run build once, then make start) instead.
  2. Expect a few things to take longer. A full financial report, returns over a long history, or the first load after the app starts do more work than an ordinary screen. Prices show the last known value straight away and refresh in the background, so a slow price provider never holds up a page.
  3. Still slow? Note which screen. “The Commitments tab takes 4 seconds” is something that can be fixed; “the app is slow” usually isn’t. Send it with Report a problem and tick the box to include recent log lines — they carry the timings described below.

Seeing the timings yourself

Every request the app answers writes one line to the log, like:

PERF-REQ GET /api/commitments total=850ms db=120ms(14q) compute=730ms

total is how long the request took, db is time spent in the database (and how many queries ran), and compute is everything else. A request that’s mostly db with a very large query count points at the data side; one that’s mostly compute points at the calculation.

Lucius is slow

Lucius’s speed is almost entirely the speed of the AI model behind him.

  • See how your models compare: Settings → Cost & Performance → Model performance shows, per model, the time to the first word (TTFT p50), how long a reply takes (Duration p50) and output speed (Tok/sec p50), over the last 50 recorded model calls. It fills in after Lucius’s first reply.
  • Pick a faster model for everyday questions in Settings → AI — see Models and cost.
  • Replies stopped mid-task? A turn (one reply, or one Automation run) is stopped after 20 minutes by default. If a slow model keeps hitting that, raise it under Settings → AI → Turn time limit → Stop a turn after — up to 2 hours, or No limit. With no limit, a turn runs until it finishes or you press Stop.