Embedding models, rerankers and evaluation datasets I use for retrieval pipelines in production. Selection criteria: latency under load, multilingual