Contents
In brief
Retrieval-augmented generation only helps if the right passage arrives and the model is willing to admit when it did not.
Original CloudExchangers explainer. It describes established public practice and labels the desk’s judgment as analysis. It is not a report of a newly announced event.
Lewis, Perez, Piktus and colleagues described retrieval-augmented generation as a way to condition a generator on passages fetched for the question. The idea is still the right default for private knowledge. The implementation is where assistants quietly fail.
What you are testing
Separate the pipeline into stages. Retrieval chooses passages. The reader, usually a language model, writes an answer from those passages and the question. A knowledge graph, if you have one, is another index: entities and relationships instead of only chunks of text. Each stage can be wrong while the final paragraph still sounds smooth.