DARTH
DARTH is a software framework designed to optimize Approximate Nearest Neighbor Search by solving the inherent tradeoff between performance and recall quality. It introduces a declarative recall mechanism that allows users to specify a precise target recall level rather than manually tuning complex, algorithm-dependent parameters. The system employs a novel adaptive early termination strategy integrated directly into the search process, ensuring that search operations stop as soon as the user-defined recall target is met. This approach eliminates the need for extensive parameter experimentation and prevents inconsistent quality across different query difficulties. Integrated as a component of the FAISS library, DARTH supports popular index structures such as HNSW and IVF. Experimental evaluations demonstrate that it successfully meets user-defined recall targets while delivering substantial speedups, ranging from an average of 6.8x faster for HNSW to 13.6x faster for IVF compared to standard search without ea