I think we have same problem and I found this

embedding_function = OllamaEmbeddings(model="llama3.2:3b")
evaluator = load_evaluator("pairwise_embedding_distance", embeddings=embedding_function)

instead asking for "llm" parameter, it should provide "embeddings" param

source: https://python.langchain.ac.cn/docs/guides/productionization/evaluation/comparison/pairwise_embedding_distance/