- gunicorn gthread 8 -> 32 threads per worker: each assistant stream holds a thread for
the whole answer; with 3x8, about 20 concurrent asks already queued ordinary page loads
for seconds (up to 23 s at 50 asks locally, 7 s at 20 asks on the live server). With 3x32,
75 concurrent asks kept page loads under 110 ms (laptop, simulated model and Qdrant)
- deploy.sh: new step closes the project dir to other accounts and sets the database and
backups to 0600/0700 on every release; user `nobody` on the shared server could read
db.sqlite3 and every backup
- requirements: python-dotenv 1.2.2, urllib3>=2.8.0 (known CVEs in the installed versions),
gunicorn pinned to the server's version (was only installed by hand)
- security-before-production.md: items 19-23 (file permissions, DEBUG=1 still in the systemd
unit, dependencies, frontend headers, guest AI cap per shared IP) and the reproduced
X-Forwarded-For spoof through the origin IP
- model-prices.md: 0.183 cents per answer from 50 answers in the live load test
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
What GapGPT actually billed per model, with dates and evidence, so a
later price rise can be compared with what it was. gpt-4.1-mini and
gpt-4o-mini were re-checked from the 852 recorded calls in
docs/benchmarks/2026-09-26-tutor-model: every call equals the official
OpenAI per-token price or less (input cache), none more.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
