Sgd + Number Of Doc In Ss

using gradient descent to find the optimal number of similarity search such that:

1. we guarantee a 100% hit rate for ground truth datasets (needs human or llm curation)
2. we guarantee a quality lower bound for ground