This thesis investigates the performance characteristics of the Learned Metric Index (LMI) for efficient vector retrieval, focusing on improving its construction speed and search performance. The student will conduct a literature review to understand what performance optimizations are used in other existing indexes for this purpose. The student will systematically analyze the existing LMI implementation and determine its primary performance limitations. The student will consider using numba, a just-in-time compiler, to speed up code execution, implementing multithreading for parallel query processing, using an automatic hyperparameter selection framework called Optuna, and training the model only on a sample instead of the full dataset. The refined implementation will be evaluated against the original version on a subset of embeddings from the LAION-5B image collection.