Please enable JavaScript.
Coggle requires JavaScript to display documents.
IR Projects - Coggle Diagram
IR Projects
Main idea
-
-
neural re-ranking
A retriever (like BM25, a dense retriever, or hybrid search) pulls back an initial set of candidate documents (usually top-100 or top-1000).
A neural re-ranker (typically a deep learning model) re-evaluates these candidates, scoring them more accurately by deeply modeling the query-document interaction.
Ranking
BM25
is a ranking function used in information retrieval — especially in search engines — to estimate the relevance of documents to a given search query.
It scores documents by matching query terms to document terms, giving higher weight to documents where the terms appear often — but with diminishing returns (so just repeating the same word over and over won’t trick the system).
Literature Review
Are Re-Ranking in Retrieval-Augmented Generation Methods Impactful for Small Agriculture QA Datasets? A Small Experiment
-
-
-
-