AI settings

AI settings

Several Keystone features can use a Large Language Model (LLM) to suggest, summarise, or classify:

• Triage classification — when a requester types in the search bar, the LLM picks the best matching template (better than keyword fallback)
• KB suggestions — when an agent opens a ticket, the LLM surfaces 3 candidate KB articles based on the conversation
• Ticket summarisation — for long conversations, an “auto-summary” tab shows what’s happened so far
• Trend detection — for the problem rules engine, LLM-aided pattern detection in the ticket stream

These are all optional. With no LLM provider configured, the features degrade gracefully — keyword-based triage instead of LLM, no KB suggestions panel, no auto-summary. Nothing breaks.

Supported providers

• Azure OpenAI — most common in education due to the data-residency story
• OpenAI direct
• Anthropic Claude — via Anthropic’s API
• Self-hosted (vLLM, Ollama, LM Studio) — for trusts running their own inference

Each provider has its own auth scheme and model list. The settings page surfaces what each requires.

Per-feature toggles

Each AI feature can be enabled or disabled independently. A trust might enable KB suggestions but disable summarisation if they want to limit cost.

The toggles live next to the provider config. After enabling a feature, the relevant UI surfaces appear in the agent shell.

Cost monitoring

Each feature reports approximate token usage to the AI usage dashboard. Trusts can set monthly budgets per feature; when 80% of the budget is consumed, an alert fires. When 100% is consumed, the feature pauses for the rest of the month.

This is a coarse safeguard — actual provider billing is what counts.

Data sent to the provider

The platform is conservative about what it ships:

• Triage classification — only the requester’s input + the catalogue of template names/summaries
• KB suggestions — the ticket subject + body of the most recent 3 messages; explicitly excludes attachments and pupil names (PII filter applied first)
• Summarisation — the ticket conversation; PII redacted

The PII filter is on by default for any LLM call that includes user-generated content. Configurable thresholds per data class.

Self-hosted recommendations

If your trust’s data-protection posture rules out third-party LLMs, the self-hosted path:

• A 7B-parameter model (like Mistral 7B Instruct) runs on a single GPU; good enough for triage + KB
• A 13B model gives better summarisation; needs more VRAM
• The platform speaks the OpenAI-compatible API; vLLM and most self-hosted runtimes support this

Self-hosted runs without sending data to anyone. Slower than hosted but reliably private.

Disabling

Turn the provider off and every feature falls back to its non-AI path. Existing AI-generated content (summaries, suggestions) stays where it was generated; nothing’s deleted.