Chunks retrieved–
Closest topic–
Farthest kept–
Drag to rotate, scroll or pinch to zoom

A toy space with real math

Real embeddings have hundreds or thousands of dimensions, which no screen can draw. This model squeezes a small library of example farm passages into three dimensions so you can see the idea: passages about similar things sit near each other, and retrieval is literally measuring distance and keeping the nearest ones.

Why RAG matters for business

A language model on its own only knows what it saw in training, which goes stale and never included your private documents. RAG lets a company keep its knowledge in a store it controls and update it any time, while the model handles the language. It also makes answers auditable, because each one can point back to the passages it used.

Tuning tip: slide top k from 1 to 8 while asking the same question. Watch when the extra chunks start coming from the wrong topic. Production systems add a reranking step to catch exactly that.

Where Avid uses this

AvidRAG, our nine level agentic RAG architecture on Red Hat OpenShift, is the engine behind GovPulse.io and VerdantaIQ.

Keep exploring