Anders Öhrn

Thanks for sharing. I am also building applications that deal with technical text, though mostly English. One thing that helped my information retrieval pipeline was to add a reranker (e.g https://huggingface.co/BAAI/bge-reranker-v2-m3 ) after the dense/hybrid search. If needed, these are lightweight enough that they can be fine-tuned to very specific domains of text (see https://huggingface.co/blog/train-reranker), though that hasn’t been needed yet for what I am building.