As image datasets become ubiquitous, the problem of ad-hoc searches over image data is increasingly important. Many high-level data tasks in machine learning, such as constructing datasets for training and testing object detectors, imply finding ad-hoc objects or scenes within large image datasets as a key sub-problem. New foundational visual-semantic embeddings trained on massive web datasets such as Contrastive Language-Image Pre-Training (CLIP) can help users start searches on their own data, but we find there is a long tail of queries where these models fall short in practice. SeeSaw is a system for interactive ad-hoc searches on image datasets that integrates state-of-the-art embeddings like CLIP with user feedback in the form of box annotations to help users quickly locate images of interest in their data even in the long tail of harder queries. One key challenge for SeeSaw is that, in practice, many sensible approaches to incorporating feedback into future results, including state-of-the-art active-learning algorithms, can worsen results compared to introducing no feedback, partly due to CLIP's high-average performance. Therefore, SeeSaw includes several algorithms that empirically result in larger and also more consistent improvements. We compare SeeSaw's accuracy to both using CLIP alone and to a state-of-the-art active-learning baseline and find SeeSaw consistently helps improve results for users across four datasets and more than a thousand queries. SeeSaw increases Average Precision (AP) on search tasks by an average of .08 on a wide benchmark (from a base of .72), and by a .27 on a subset of more difficult queries where CLIP alone performs poorly.