Onthology guided RAG augmentation - evaluation framework for experimentation with different augmentation strategies , extration of corpus and evaluation

This repository contains the implementation code, ontology builder, and evaluation scripts for our paper on improving Retrieval-Augmented Generation (RAG) systems in acronym-heavy educational environments. Overview Educational RAG systems struggle with acronyms due to vocabulary mismatch between queries and documents. This repository implements a scalable solution that: Builds corpus-scale acronym ontologies through multi-phase extraction and disambiguation Augments retrieval at two leverage points: Query-Time Augmentation (QTA) and Ingestion-Time Enrichment (ITE) Fine-tunes embedders on acronym↔expansion pairs for improved semantic matching Achieves substantial gains: Up to 102% improvement in Recall@10 for acronym queries Key Results QTA alone: +55.3% nDCG@10, +67.2% Recall@10 on average QTA+ITE combined: +64.5% overall improvement for open-source embedders Small augmented models match or exceed larger proprietary models while eliminating API costs