How do you improve the embedding quality of encoder models in your RAG pipeline? By making them aware of similar documents and the context of the corpus.