facebookresearch/faiss ? reverse-engineered prompt

Build me a fast vector search library for finding similar items from dense numeric embeddings. I want to add millions or even billions of vectors, give each one an integer id, then query with a new vector and get back the closest matches.

Support Euclidean distance, dot product, and cosine similarity when the vectors are normalized. Include both exact search for smaller data and approximate search options that trade a little accuracy for speed and lower memory use. It should also include clustering, evaluation helpers, and simple tuning tools so I can compare speed, quality, memory use, and training time.

Make the main implementation in C++ with clean Python bindings so I can use it from Python with array data. Start with CPU support, but design it so GPU acceleration can be enabled when CUDA or ROCm is available. Include practical demos, tests, and clear install instructions. Look up current docs online if you need to.