Key points
- Current-page context is metadata only: route, page area, active tab, team and scope, selected resource IDs, and allow-listed filters.
- The composer renders a removable context chip so you can see and drop the context before sending.
- The Assistant does not scrape page text, logs, secrets, credentials, or arbitrary query parameters.
- Assistant selects from the same models pipelines use, filtered by what the caller may access.
- Conversations are per-user and can be deleted; memory can be compacted with a summarize action.
- Assistant availability and individual features are configuration-controlled.
How it works
AI Evaluation on analysis surfaces is a separate mechanism. It uses the analysis evaluation endpoint rather than creating an Assistant conversation, and it does not ask hosted MCP to inspect resources.
If Assistant reports that no model is available, configure provider, model, credential reference, and scope access in Models — the same place pipeline goals get their profiles.
Examples
# feature families enabled for the install
curl -s -H "Authorization: Bearer $NOPSAI_TOKEN" "$NOPSAI_URL/v1/assistant/config" | jq
# models the caller may drive the assistant with
curl -s -H "Authorization: Bearer $NOPSAI_TOKEN" "$NOPSAI_URL/v1/assistant/models" | jq
# conversations belong to the caller
curl -s -H "Authorization: Bearer $NOPSAI_TOKEN" "$NOPSAI_URL/v1/assistant/conversations" | jqImplementation evidence
doc/assistant-capabilities.mdUser-facing capabilities and example prompts per feature area.
services/ui/src/features/assistant/pageContext.tsExactly which page metadata is collected.

