Give an AI the Right Amount of Information, Not the Most.

Give an AI too little and it guesses; too much and the fact you needed gets buried in the middle.

Main Takeaway

An AI answer is only as good as the information it has access to. Give it too little and it guesses; give it too much and the one fact you needed gets buried in the middle, where models reliably lose track of that fact.

Who is this for

Anyone using AI to answer questions based on a bundle of trusted materials. Think company documentation, legal, medical, or regulated industries with strict rules about what counts as true.

Background

GINGER (Grounded Information Nugget-Based Generation of Responses) is a 2025 research paper by Weronika Łajewska and Krisztian Balog at the University of Stavanger. It tackles Retrieval-Augmented Generation (RAG), the now-standard setup where an AI first retrieves documents and then provides an answer based on them. The system was evaluated on TREC RAG’24, a standardized test from the long-running Text REtrieval Conference, where it reached the best results anyone had achieved at the time.