Troubleshooting¶

Rate limit errors¶

Increase the per-hour cap:

# .env
RATE_LIMIT_SUMMARIZE=60 per hour

Or disable limits entirely for a single run:

DEBUG=True

See Rate limiting for the full syntax.

You haven't set a provider API key that matches any enabled entry in config/providers.yaml. Set at least ANTHROPIC_API_KEY in .env, or edit the provider registry to enable a different provider.

The container logs this on startup when no providers are enabled:

[bakwas] No enabled providers. Set at least one provider API key in .env ...

Permission errors on the database¶

chmod 755 ./data

The container runs as a non-root user (appuser, UID 1000). The host ./data directory must be writable by that UID.

Template not found¶

Templates must live at templates/ in the project root, not inside src/templates/. This only matters if you're building your own image or running locally without Docker.

Container uses more memory than expected¶

Idle memory is ~60 MB. After the first summary, it jumps to ~200 MB because LiteLLM and its provider SDKs are loaded lazily on demand and kept resident for subsequent requests. This is expected and steady-state.

If you see memory growing continuously over time (past ~250 MB with no activity), that's abnormal — check docker stats, container logs, and open an issue.

Long LLM calls time out behind Cloudflare¶

Cloudflare's proxy enforces a 100-second response limit. A long transcript through a slow model can exceed it. Two fixes:

Use a faster model for that request (the streaming endpoint keeps the connection alive via incremental data, which typically stays within Cloudflare's buffer budget).
Set the DNS record for the Bakwas subdomain to DNS-only (grey cloud) to bypass the proxy.

Streaming doesn't work¶

Streaming requires the reverse proxy to disable response buffering. The Bakwas container sends X-Accel-Buffering: no on stream responses, which Nginx and NPM respect. If your proxy still buffers, check its proxy_buffering setting (Nginx) or equivalent.

Cost shows as $0.0000 or "unknown"¶

The LiteLLM pricing database ships with most major models but occasionally lags brand-new releases. If your cost is missing:

Confirm the model name matches an entry in LiteLLM's model_prices_and_context_window.json.
Update LiteLLM to the latest version in requirements.txt and rebuild.
For custom proxies where LiteLLM doesn't know the underlying model, cost is reported as "unknown" — that's expected.