Answers are grounded in this paper and cite their sources.
Multilingual Retrieval-Augmented Generation (RAG) systems serving low-resource languages can suffer from unintended retrieval bias when language preferences are applied without accounting for the score scale of the underlying fusion method. This paper investigates a failure mode in a trilingual agricultural advisory system, Bhunaga, supporting English, Hindi, and Marathi. The system uses BGE-M3 dense and sparse retrieval combined through Reciprocal Rank Fusion (RRF) with a rank constant of k = 60. We demonstrate that an additive same-language bonus of +0.10 exceeds the maximum attainable fused score range of 0.0217 by approximately 4.6 times, effectively turning a soft reranking preference into a hard language filter. We derive the partition condition mathematically and evaluate three configurations: additive bonus, multiplicative bonus, and no language preference. Experiments on a corpus of 114 chunks across 24 documents show that the additive bonus reduces keyword recall by 45% for Hindi and 39% for Marathi compared with no bonus. English retrieval remains unchanged, acting as an internal control. We also identify a misleading monitoring metric: same-language retrieval rate reaches 1.00 precisely when retrieval quality is at its worst. The study proposes a scale-aware approach to score adjustment, recommends multiplicative adjustments where appropriate, and demonstrates why multilingual retrieval systems should evaluate cross-lingual retrieval quality rather than relying solely on same-language retrieval rates.
No takes yet. Share an insight, caveat, or question.
Agre et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: