AI Can Read Your Documentation. But Can It Understand the Relationships?
I've been thinking about a problem that becomes obvious once a project gets large enough.
The problem isn't always finding information.
Sometimes the information is already there.
The problem is understanding how everything is connected.
Imagine an AI coding agent working on an authentication service.
It finds the authentication documentation.
But does it know that:
The service depends on a specific API contract?
That API is consumed by three other services?
A security policy requires MFA?
An architecture decision prohibits a particular dependency?
A database schema is shared with another domain?
A deployment workflow requires additional validation?
A previous incident changed the recommended implementation?
All of those pieces may exist in different documents.
A human engineer can connect the dots.
An AI agent needs those relationships to be explicit.
That's why I've started thinking about knowledge as a graph, not just a collection of documents.
For example:
Architecture | +---- defines ----> Service | +---- exposes ----> API | +---- depends ----> Database | +---- governed by -> Security Policy | +---- tested by ---> Test Suite | +---- deployed by -> Workflow
Now the AI isn't just searching for "authentication."
It can traverse the engineering knowledge connected to authentication.
Architecture.
Dependencies.
Policies.
APIs.
Tests.
Incidents.
Deployments.
Decisions.
This changes the retrieval problem.
Instead of asking:
"Which document is relevant?"
We can start asking:
"Which knowledge and relationships are relevant to this engineering decision?"
That's a much more interesting problem.
And I think it becomes increasingly important as AI agents move from generating individual functions to modifying entire systems.
This is another area I'm exploring while building NAEOS.
The goal isn't to build a smarter search engine.
It's to create an engineering knowledge layer where relationships are first-class.
Because good engineering decisions rarely depend on one piece of information.
They depend on how many pieces of information fit together.
Do you think AI coding agents need a knowledge graph to work effectively on large software systems?
