Retrieval-augmented generation (RAG)

Last updated: Saturday, 29 June 2024

Hitching or coupling information retrieval and generative models. Enhancing the quality and relevance of generated content by drawing on external knowledge.

Feeds the retrieved context along with the original prompt into a language model, allowing the model to draw upon the retrieved information when generating a response.

Unlikely things that could qualify as RAG: collage, oral storytelling, DJing, the legal system, the immune system, improv comedy riffing on audience suggestions.

• [⎈] Investigate work on the interpretability of RAG systems; how can users understand what retrieved information the model is relying on and how it is being used?
• [?] What are the computational challenges in deploying RAG systems at scale? How do they compare in terms of resource use?
• [?] What happens when access to external sources is limited or unreliable?

Backlinks

: Gopnikism : Prompt engineering