Retrieval uses all- MiniLM-L6-v2 embeddings with a FAISS index, a vector-search index, while Qwen2.5-7B-Instruct handles the main decomposition setup. In the benchmark comparison, context-window ...