Assistants
Settings → Assistants (/admin/settings/ai/assistants) configures the two LLM-backed features Capacitor runs on your behalf: the conversational analytics agent and session title/summary generation. Both reuse vendor credentials from AI providers — this page has no API key fields of its own.
Analytics agent
Section titled “Analytics agent”Conversational analytics on the Analytics tab. Uses the Eval Runner API key for the selected vendor. On the tenant-wide view the agent also runs live, event-level queries over KurrentDB Flight SQL (requires a licensed KurrentDB ≥26.1).
| Key | Default | Description |
|---|---|---|
Analytics:Agent:Enabled | — | Master switch for the agent. |
Analytics:Agent:Vendor | — | Anthropic, OpenAI, KurrentAI (on plans that allow it), or Cohere. |
Analytics:Agent:Model | — | The chat model to use for the selected vendor. |
Analytics:Agent:MaxRows | 5000 | Row cap on read-model query results returned to the agent. |
Analytics:Agent:TimeoutSeconds | 30 | Per-query timeout. |
See Analytics for the user-facing feature this configures.
Session titles & summaries
Section titled “Session titles & summaries”Lets the server (re)generate session titles via the refresh button next to a title. API keys are reused from the Eval Runner blocks on AI providers (Evals:ServerRunner:*:ApiKey) — no separate key is needed here.
With one vendor keyed, it’s used automatically. With both keyed, pick a provider with SessionTitle:Provider (anthropic, openai, kurrentai, or cohere) — the regenerate button stays hidden until one is selected. Saved changes take effect immediately; no restart required.
Models
Section titled “Models”Per-vendor model overrides, each with a built-in default:
| Key | Default |
|---|---|
SessionTitle:Anthropic:Model | claude-haiku-4-5 |
SessionTitle:OpenAI:Model | gpt-5.5-mini |
SessionTitle:KurrentAI:Model | @cf/meta/llama-3.3-70b-instruct-fp8-fast |
SessionTitle:Cohere:Model | command-a-03-2025 |
SessionTitle:RunTimeoutSeconds bounds the per-attempt wall-clock budget: 5–120, default 30.
Auto-backfill
Section titled “Auto-backfill”When enabled, the server periodically generates titles for ended sessions that only have a raw first-message title (e.g. Claude Desktop sessions whose watcher couldn’t reach the title model).
| Key | Range | Default | Description |
|---|---|---|---|
SessionTitle:AutoBackfill:Enabled | — | off | Master switch for the sweep. |
SessionTitle:AutoBackfill:IntervalSeconds | 60–3600 | 300 | Seconds between sweep ticks. |
SessionTitle:AutoBackfill:MinEndedAgeMinutes | 1–60 | 2 | Only sessions ended at least this long ago are eligible, giving in-flight watcher title work a grace window. |
SessionTitle:AutoBackfill:MaxAttempts | 1–10 | 3 | Max LLM attempts per session before giving up. |
SessionTitle:AutoBackfill:BatchSize | 1–100 | 20 | Max sessions processed per sweep tick. |
Related
Section titled “Related”- AI providers — the vendor credentials both features reuse.
- Analytics — the user-facing Analytics tab the analytics agent powers.