Stop using naive RAG

Eduardo Borges — Sat, 18 Apr 2026 13:45:37 +0000

Most RAG setups look good in demos — until things get slightly complex.

You ask a question, it retrieves “relevant” chunks, and everything seems fine.

But as soon as your system spans multiple documents — APIs, billing, infra, workflows — things start breaking down.

Not because the information isn’t there.
But because the relationships between them are lost.

The problem with RAG

RAG works by retrieving chunks based on similarity.

That means:

It finds text that looks relevant
But doesn’t understand how pieces connect
And can’t reconstruct system behavior

So you end up with answers that are:

technically correct
but incomplete
and often misleading

Real systems aren’t flat

In real systems:

a deploy triggers a pipeline
the pipeline applies changes to Kubernetes
monitoring evaluates the rollout
failures trigger rollback logic

None of this lives in a single document.

And RAG doesn’t connect these dots.

What I built instead

I built Mindex:
https://usemindex.dev/

Instead of just retrieving chunks, it builds a knowledge graph on top of your documents.

So your AI can:

connect documents
follow relationships
reconstruct flows

Not just match text.

RAG vs Graph-based context

Here’s a simplified comparison:

❌ Naive RAG

Returns a flat list of documents
No relationships
No ordering
No system understanding

✅ Mindex (GraphRAG)

Connects documents
Traverses relationships
Infers flows (cause → effect)
Provides structured context

Why this matters

The difference is subtle at first.

But when you're working with:

internal documentation
APIs
distributed systems

It becomes critical.

You don’t just need relevant text.

You need to understand how things work together.

How it works

Mindex combines:

semantic search
a knowledge graph layer
relationship traversal

It’s available via:

CLI
MCP (works with tools like Claude Code, Cursor, etc.)
REST API

Try it

You can try it here:

https://usemindex.dev/

Feedback welcome

I’m especially interested in feedback from people:

building with RAG
working with internal knowledge bases
building AI dev tools

Curious to hear how you're handling this today.

Forem: Eduardo Borges