✳flâneur — a map of the web's best reading

Better RAG 2: Single-shot is not good enough

olickel.com · 1,832 words · saved by 1 readers

This is part 2 of a series on improving retrieval-augmented-generation systems. Part 1 covers the basics of such systems, key concerns such as question complexity, and the need for new solutions. In this part, we'll cover the basics of multi-turn retrieval - what it is, why it's needed, and how to implement it. If you were interested in how WalkingRAG works, this article should leave you with a working, implementable understanding of how to build a similar system. Almost all RAG systems today work 'single-shot' - for a given question, they retrieve information, trim and modify, and use an LLM to generate an answer. Prima facie this seems okay - until you consider how humans answer questions today. Let's presume that you're an above GPT-4 level intelligence. How often have you been able to solve problems with a single round of retrieval? In most cases, your first round of Google searches, combined with any residual information you have, gets you closer to finding the answer - rarely do

Better RAG 2: Single-shot is not good enough Hrishi Olickel Better RAG 2: Single-shot is not good enough How to 'walk' to better answers This is part 2 of a series on improving retrieval-augmented-generation systems. Part 1 covers the basics of such systems, key concerns such as question complexity, and the need for new solutions. In this part, we'll cover the basics of multi-turn retrieval - what it is, why it's needed, and how to implement it. If you were interested in how WalkingRAG works, this article should leave you with a working, implementable understanding of how to build a similar sy

Explore this link on the map →

related reading