Jina Embeddings v3 Task LoRA Adapters: Separate retrieval.query and retrieval.passage Encoders for Asymmetric RAG

arXiv·medium signal

Jina-embeddings-v3 ships five task-specific LoRA adapters (retrieval.query, retrieval.passage, separation, classification, text-matching) injected into all 24 transformer layers, adding less than 3% parameter overhead. Passing task='retrieval.query' for the user question and task='retrieval.passage' for document chunks encodes them through different low-rank projection paths, closing the asymmetry gap that generic single-encoder models miss. The model supports 8192 token context and 100+ languages, making it production-ready for multilingual RAG.