Soft Compression

A RAG context compression technique that encodes retrieved text passages into compact continuous embedding vectors, referred to as memory embeddings, which are passed as soft prompts into the language model's attention layers. Compared with hard compression, it achieves higher information density.