¶ Known Limitations and Current Issues — L&S AI Inference
This page documents known limitations, current issues, and expected behaviors of the L&S AI Inference platform during early access. Check back as the service evolves.
Last updated: [date]
| Issue |
Status |
Workaround |
| Occasional high latency during peak usage |
Ongoing |
Try off-peak hours; use gemma-4-31b for faster responses |
| Web search may return no results for niche queries |
Ongoing |
Disable web search and rely on model knowledge; rephrase query |
| Very long responses may be cut off near token limits |
Ongoing |
Ask for shorter responses or request output in sections |
Issues will be added and resolved here as the service evolves.
All models support 128k tokens, but longer contexts come with tradeoffs:
- Response quality degrades near the context limit. Models may lose track of details mentioned early in a long conversation.
- Response latency increases with context size. Very long conversations will be noticeably slower.
- Context is not automatically managed. The model does not summarize or truncate older turns — the full conversation history is sent every time until the limit is reached.
See the Context Windows Guide for strategies.
- Search queries leave UCSB infrastructure. Queries are sent to an external search engine when web search is enabled. See data and privacy.
- Search results may be incomplete or outdated. External search engines may not index all relevant sources. For authoritative sources, retrieve them directly and upload or paste the content.
- Web search adds latency. Each search-enabled request makes an external network call before the model responds.
¶ Data and Privacy Notes
- All inference runs on UCSB hardware
- Conversations may be accessible to platform administrators for operational purposes
- Web search sends query text to an external engine — disable for sensitive topics
- Do not submit FERPA, HIPAA, CUI, or P3+ data
If you encounter a problem not listed here, please contact help@cit.ucsb.edu with:
- What you were trying to do
- The model you were using
- The error or unexpected behavior you observed
- Approximate time it occurred