Northfield Pumps' on-call problem
Northfield Pumps (made up) runs plants on three continents. Its IT operations team has fifteen years of knowledge spread across ServiceNow tickets, Confluence runbooks, Jira changes, Slack threads and Git repositories. At 2 a.m., the on-call engineer's real question is simple: has this happened before, and what fixed it? All names here are fictional.
The set-up
Insight Ask connects to the team's tools (it lists SharePoint, Teams, OneDrive, Jira, Confluence, Slack, Git, ServiceNow, databases and object stores) and gives everyone one search box that answers in plain language, with the documents behind the answer. Underneath, Insight RAG does the ingestion: connectors, extractors for each format, chunking and embedding into a vector database, with metadata captured at ingestion and enforced at query time.
Permissions are the hard part
The first question from security is always: can an engineer find a document through search that they couldn't open directly? The design answer is to carry access rules and classification metadata with every chunk and filter at query time, then audit each retrieval. Sensitive values can be redacted or tokenised before they reach the index. Northfield starts with the sources where access is simplest (runbooks and resolved tickets) and adds restricted spaces only after testing what each role can see.
What a search looks like
The on-call engineer types: "Line 3 historian stopped collecting after the Windows patch, has this happened before?" Ask returns an answer that cites a resolved ticket from last spring, the runbook step that fixed it, and the Jira change that introduced the patch, each linked. Entity relationships connect the people and systems involved, and long threads are summarised. Ask also offers Chrome and Edge extensions, so engineers can search from wherever they're working.
Knowing it works
Every answer is traced in Insight Lense: which sources were retrieved, which model answered, how long it took and what it cost. When an answer is wrong, the trace shows whether retrieval found the wrong document or the model misread the right one, which are two different fixes. Ask's own analytics show what people search for and what they don't find, which is a to-do list for the knowledge base.
Rolling it out
- Two sources, one team. Confluence runbooks and resolved ServiceNow tickets, for on-call engineers only.
- A test set. Fifty real past incidents and the documents that solved them, to check retrieval before launch.
- Permissions review with security, by role.
- Widen: Jira and Slack next, then other teams, adding a source only when it has an owner.
Search reflects the knowledge base. If the runbook is out of date, the answer will be too. Treat "no good answer" results as a signal to fix documentation, not just the search.
Beyond IT operations
The same pattern fits other teams once the first one works, each with its own sources and access rules:
- Engineering: code, commits and pull requests in Git alongside design docs, to answer "why was this built this way?"
- HR and people teams: policies and procedures, so employees find answers without opening a ticket.
- Customer-facing teams: customer communications and account history in one search.
Each new team is a new set of sources, permissions and test questions, not a new system. That's where the investment in query-time access control pays off.
Related: the reference architecture · a phased adoption plan.