AI Features

This page is also in the Groundskeeper wiki: https://github.com/rsjsoftwareltd/groundskeeper/wiki/AI-Features

Groundskeeper can use AI to turn raw monitoring data into plain-English summaries and answer questions about your IT estate. AI is entirely optional -- the dashboard, alerts, SLT view and every connector work fully without it. You choose whether to enable it, and which provider to use.

The short version: for most schools, run the local Ollama option. It runs on your own server, costs nothing, and no data leaves your network. Only choose a cloud provider if you specifically want the extra capability and are comfortable with monitoring summaries being sent to that provider.

Choosing a provider #

AI is pluggable -- you pick one provider and can switch later from Settings > AI (the AI Provider card). Every provider is included in every edition, Community too -- you pay the cloud provider directly, not us.

Ollama (local) -- recommended #

Ollama runs an open-weight language model directly on your Groundskeeper server. Nothing is sent to any third party -- summaries are generated on-box, which keeps you fully aligned with the "data stays on your network" model.

Cost: free.
Privacy: the strongest option -- no data leaves the server.
Hardware: needs more RAM than the rest of Groundskeeper. ~8 GB RAM is a sensible floor; more is smoother. It is CPU-only viable -- a GPU helps but isn't required for the summary workload.
What affects response speed: mostly the CPU, not the RAM. Once the model fits in memory, extra RAM gives you headroom rather than speed -- answer times are driven by how many cores you have and how modern they are. A measured reference point, so you can compare against something real rather than guess: on an 8-core / 16 GB VM with an older-generation CPU, answers come back in about 10 seconds (and use under 8 GB of RAM). A newer CPU may match that on fewer cores. We deliberately don't publish a minimum core count, because core count alone predicts this badly across CPU generations.
If summaries feel slow: that's usually CPU, not a fault. Summaries are generated in the background on a schedule, so speed matters less than it might seem -- the dashboard reads the cached result rather than waiting on the model. Switching to a smaller model in Settings is the quickest improvement if you want it faster.
Setup: Groundskeeper can detect Ollama and pull a model for you (the default is a small, capable general model). See setup below.

Running Ollama on another machine #

Ollama doesn't have to run on the Groundskeeper server. If you have a machine with a good graphics card, run Ollama there and point Groundskeeper at it. Answers come back much faster, and the Groundskeeper server can stay a small virtual machine that's easy to back up and restore. Your data still stays on your network.

Install Ollama on that machine and let it listen on the network. Set the environment variable OLLAMA_HOST to 0.0.0.0 there, then restart Ollama.
Firewall it. Ollama has no password: anyone who can reach it can use it. On that machine, allow port 11434 from your Groundskeeper server's address only.
Point Groundskeeper at it. In Settings > AI (the AI Provider card), choose Ollama, enter its address (for example http://gpu-server:11434), click Test connection, then save.
Manage models from Groundskeeper as usual. The Model Library lists, downloads and removes models on that machine. Installing or updating Ollama itself is done on that machine, not from Groundskeeper.

Cloud providers #

Settings offers five: Azure OpenAI, OpenAI, Anthropic Claude, Google Gemini (suits Google Workspace schools) and a custom OpenAI-compatible endpoint. Each sends scrubbed monitoring context to that provider (see the PII scrubber below). The two most schools ask about:

Azure OpenAI (cloud) #

A cloud provider with strong capability. Monitoring context is sent to your Azure OpenAI resource for processing.

Cost: pay-as-you-go via your Azure subscription.
Privacy: data is sent to Azure. Prefer a UK or EU region for data residency. Azure is the cloud option most schools will find easiest to justify, since many already have an Azure tenant and a data-processing relationship with Microsoft.

Anthropic Claude (cloud) #

Another capable cloud provider. Monitoring context is sent to Anthropic's API for processing.

Cost: pay-as-you-go via an Anthropic API key.
Privacy: data is sent to Anthropic. Consider this where your data-handling policy permits a US-based processor.

LM Studio (local, GPU) #

An alternative local engine for a server with a GPU, under Settings > AI Provider's GPU tier. Like Ollama, nothing leaves the server.

Turning AI off #

Switch off Enable AI features at the top of Settings > AI. Your provider settings are kept, so you can switch it back on later. Everything else in Groundskeeper works exactly the same -- you simply won't see the AI-generated narrative or briefing. The Ask Groundskeeper boxes on the Dashboard and the Intelligence page stay where they are; if you use one, it answers that AI features are turned off and points you to Settings > AI.

Privacy and the PII scrubber #

When a cloud provider is selected, Groundskeeper runs the monitoring context through a PII scrubber before sending it, which removes things like IP addresses, DOMAIN\user names, email addresses, UPNs and MAC addresses from the text. This is a best-effort safeguard, not a guarantee -- it reduces what is sent but cannot promise that every identifier in free-form data is caught.

Because of that, the privacy trade-off is simple:

Local (Ollama): nothing leaves the server. No scrubbing needed.
Cloud (any of the cloud providers): scrubbed monitoring context is sent to the provider. Choose a region appropriate to your data policy (UK/EU via Azure for the tightest residency), and treat it like any other cloud processor in your records.

Groundskeeper logs a one-time warning when a cloud provider is selected, as a reminder that context will leave the network.

Where AI appears in Groundskeeper #

Once enabled, AI shows up in a few places -- all of which simply go quiet if AI is off:

Dashboard summary -- a short narrative of the current estate state, refreshed by a background task. The SLT view does not depend on AI, so it always loads instantly.
Intelligence page -- a fuller workspace: an AI Briefing, a Recurring Events panel highlighting issues that keep coming back, and an Ask Groundskeeper box where you can ask free-form questions about your monitoring data ("what should I look at first?", "why does Veeam keep failing?").
Why? and Draft Comms on the dashboard feed -- Why? asks the AI to explain an event in plain English. Draft Comms, on critical events, drafts a short note for senior leaders, with a Copy button so you can paste it into an email and adjust it there.
Morning summary -- a fresh summary is made each morning (07:30 by default), ready for the start of the day.
AI Diagnostics page -- provider health, a live call log, latency, and a test-prompt box so you can confirm the AI is working and see exactly what it returns. Open AI Diagnostics at the top of Settings > AI takes you there.

The questions people ask (Ask Groundskeeper, Why?, Draft Comms, the Teams bot and the Network page's explanation) count towards a daily limit; when it is reached, Groundskeeper says so, and the count resets at midnight UTC. The automatic summaries are not counted.

Notes for the AI about your school (optional, on the AI Provider card) gives the model background it can rely on, such as "the server room is in Block B". Keep it short and factual. With a cloud provider these notes leave your network along with the events, so write nothing there you would not send to that provider.

Setting it up #

Go to Settings > AI (the AI Provider card).
Choose a provider:
- Ollama: if it isn't installed yet, click Try automatic install,
- A cloud provider: enter its API key -- and, for Azure OpenAI, your
Save. The background summary task picks up the new provider on its next run (every few minutes), or you can trigger a refresh from the Intelligence page.
Check the AI Diagnostics page to confirm the provider is healthy and send a test prompt.

You can change provider or turn AI off again at any time -- it never becomes a hard dependency.