Whether figuring out where to eat in an unfamiliar city or deciding which college to attend, large collections of consumer generated data (i.e. reviews and forum posts) are an important influence in online decision making. To make sense of these rich repositories of diverse opinions, users need to sift through a large number of items to learn common attributes, what the important features are, and determine how those match with their personal preferences. We introduce a novel system, SearchLens, which allows the user to build up “lenses” that enable them to evolve representations of their interests in a way that matches how their interests are instantiated in the data. SearchLens goes beyond previous approaches such as faceted browsing, user intent models, and personalized search by introducing the idea that each lens can match part of a users’ latent interests and concepts, and that these lenses can be composed together to match a user’s particular context (e.g., searching for restaurants with outdoor spaces and that serve alcohol and aren’t too crowded). Across a lab and field study we find that users find benefits in the SearchLens approach, including being able to transfer and reuse lenses across contexts, find and capture “honest signals” of their concepts in the data, transparency and explainability, and working at multiple levels of specificity and hierarchy.