BGE Large EN v1.5

bge-large-en-v1.5 from the Beijing Academy of Artificial Intelligence is a BERT-large-based English embedding model producing 1024-dim vectors with a 512-token limit. The v1.5 release recalibrated the cosine-similarity distribution so scores spread more usefully across the range, and it no longer strictly requires the retrieval instruction prefix for symmetric similarity. It ranked at the top of the MTEB English leaderboard when released and remains a common open-weight default for RAG and semantic search. Served through Hugging Face with the feature-extraction pipeline; for asymmetric retrieval prepend the recommended query instruction to questions.