People search “what is RAG” and “wat is rag” after a vendor slide used the letters as a product. The letters mean retrieval-augmented generation. The operational test is uglier and cheaper than the slide.
What is RAG? The model must fetch named passages before it answers. Retrieval is the product. Generation is the last mile. If you cannot inspect which chunk was used, you do not have RAG. You have extra context in a prompt.
Not “the model read our docs”
Stuffing a PDF into the window is not retrieval. A vector store is a component. Embeddings are a component. Naming the passage is the system.
Walk one bad answer.
- Which passage was gathered?
- Why did it rank for this question?
- What was assembled around it?
- Did anyone inspect a miss, or only a demo?
- What would you navigate to next?
If those lines are blank, stop arguing about chunk size. You shipped generation.
RAG and GRAIN
GRAIN is the inspectable loop: Gather, Rank, Assemble, Inspect, Navigate. RAG names the architecture. GRAIN names the honesty check. They are not the same word.
Evaluate the reasoning, not the fluency is the adjacent test when the sentence is tidy and the source is not. Fluency is cheap. A named chunk is not.
The book, the page
The RAG Engineer owns the longer practice. It is in production. It is not for sale. You do not need the hardcover to demand the chunk.
If the team cannot open the passage behind the sentence, do not call it RAG. Call it a prompt with extras.