Conventional item response data analysis typically relies on several assumptions, such as local item independence, respondent independence, and homogeneity. However, these assumptions are often violated in practice and difficult to verify. To weaken the reliance on these assumptions, I propose a novel latent space approach to item response data which assumes both items and respondents are embedded in an unobserved metric space. In this approach, the probability of a correct response decreases as a function of the distance between the respondent’s and the item’s position in the latent space where the respondent-item distances represent their interactions. The resulting latent space approach provides a low-dimensional geometric space referred to as an interaction map. With empirical examples, I will illustrate the utilities of the proposed method, focusing on how the interaction map can help derive insightful diagnostic information on items and respondents.