A week after Ship London I finally stopped explaining our knowledge agent as "the KT bot" in every meeting.

Katie Knows is the product name. Knowledge Transfer shortened to KT, Katie roughly phonetic, and then we kept the branding because it stuck. What matters is not the nickname. What matters is that a VMO2 knowledge specialist can live on Eve with a sandbox, a durable ephemeral filesystem, and two front doors: Slack and a custom web UI.

People use it collaboratively in a channel. People also use it alone when they need an answer before standup. Both modes are intentional.

Enterprise estates are not a single wiki. They are thousands of repos, tribal architecture decisions, partner boundaries, and diagrams that only make sense if you were in the room.

We needed an agent that could:

Eve's filesystem-first agent model (directory in, production path included) was the substrate. Katie Knows is the specialist we put on top.

The heart of the quality jump is Google's Open Knowledge Format (OKF).

Treat knowledge as a graph:

Better RAG. Not dump-PDFs-and-pray RAG. Structured retrieval over concepts and links, then answers that can point at evidence.

At our current scale across VMO2 and partners:

That is the honest boast: the estate becomes knowable as a graph in a way no single person can hold. Katie Knows is how humans query that graph without pretending one architect memorised the lot.

When a relationship is the answer, Katie Knows does not only write a paragraph. It can emit Mermaid: flowcharts for system shape, sequence diagrams for "what calls what, in what order".

That sounds cosmetic until you watch a platform conversation derail for twenty minutes because everyone imagined a different arrow. A diagram generated from the graph cuts the theology short.

We did not start here.

Early days: databases plus embeddings through Gemini Enterprise. Classic RAG. Fine for demos. Painful at estate scale.

We hit vector and embedding limits. Worse, Gemini would often steer off mid-answer: fluent, confident, gently wrong by the third paragraph. Familiar failure mode if you have ever watched an LLM "helpfully" abandon the retrieved context.

Then: same underlying knowledge, remodelled as an OKF graph, served through RAG on an Eve filesystem (sandbox + durable ephemeral storage for the working set).

Quality changed in a way evals could see. Answers stayed on the rails. Evidence showed up. Runs that used to wander started passing in our evals almost every time, and the path was fast enough and efficient enough to use in the flow of real work rather than as a weekend experiment.

That is the lesson I will keep repeating in rooms that want another embedding vendor bake-off: format and graph structure beat "more vectors" once you are past toy corpora.

Katie Knows started as a knowledge-transfer helper. It is growing into a knowledge worker / knowledge agent: research, explain, diagram, and increasingly act through tools we add carefully.

Example of the trajectory: read GCP system config as a tool, so answers about how something is actually configured do not depend on a stale Confluence page. More tools will follow. Each one is a privilege boundary, not a party trick.

Slack remains the collaborative surface. The custom web UI is for deeper sessions, artefacts, and people who want a proper workspace rather than a thread.

Not our branding. The stack shape:

Katie Knows works because OKF turned our estate into a graph we can trust, and Eve gave that graph an agent worth talking to.