Maximum inner product search (MIPS) is a crucial component of recommendation systems, search, and retrieval-augmented generation (RAG), involving the search for a key vector in a large database that best matches a query vector. As the number of keys and vector dimensions increase, the computational cost of exhaustive comparison grows significantly, leading to the development of approximate methods that trade off accuracy for speed. These methods, such as clustering and quantization, focus on the properties of the key set to build indexing structures, but often assume the query set will resemble the key set or neglect query considerations, highlighting the need for more efficient and effective MIPS approaches to support better RAG performance.