SHIFT is a method for improving multilingual retrieval-augmented generation by correcting language bias in embedding models. It models each language as a fixed offset in vector space and removes these offsets to align texts by meaning rather than language, offering a low-cost solution compared to translation pipelines or larger models.