Known Limitations

• Large WebGPU models can fail on low-memory GPUs or mobile browsers.
• First model download can be slow because model weights are large.
• Deep Search depends on third-party search/extraction providers and can degrade when providers rate limit or fail.
• RAG currently extracts text, not full visual understanding of scanned images or diagrams.
• In-memory serverless rate limits are best-effort and reset per deployment instance.
• Generated code and HTML previews should be reviewed before trusting their behavior.
• Opt-in telemetry records funnel events only; it is not a full analytics warehouse.