Reciprocal Rank Fusion (RRF) - How to Stop Worrying about Boosting - Lecture
OpenSource Connections via YouTube
Overview
Explore the concept of Reciprocal Rank Fusion (RRF) in this 39-minute conference talk from Haystack EU 2023. Dive into the algorithm that combines multiple result sets with different relevance indicators into a single result set without requiring tuning. Learn how RRF can be an effective alternative to boosting when combining different types of searches. Discover the algorithm's inner workings, implementation techniques, and potential surprises when transitioning from BM25 search. Gain insights from Philipp Krenn, a developer advocate and team lead at Elastic, as he demonstrates the practical applications of RRF in various systems, including Elasticsearch. Understand the RRF function, ranking constant, and window size, and explore a minimal example to grasp its efficiency and advantages over traditional boosting methods.
Syllabus
Introduction
What is RRF
Scores as percentages
WhyScores as percentages
What is the next solution
Maxmeanscale
Generating Ranking
Reciprocal Rank
Optional Extension
RRF Function
Ranking Constant
Window Size
RRF Example
RRF Blended
RRF Efficiency
RF in Various Systems
Minimal Example
Why is RF great
Renquest
Elastic Search
Questions
Taught by
OpenSource Connections