The architecture. On its own it predicts text; it knows nothing about your data.
Technology path
From the Transformer to a RAG application
The path the PRD names: how the architecture underneath becomes an application that answers from your own documents.
4 steps · curated, not generated
-
-
A Transformer trained at scale to follow instructions.
-
Retrieve relevant passages first, then let the model answer from them. This is what makes an answer traceable to a source.
-
The vector database that makes the retrieval step fast enough to run on every question.