Abstract Basically there are two approaches for dealing with incomplete or imprecise information in the framework of (relational) databases. Buckles and Petry introduced fuzzy similarity relations in order to estimate to what extent possible values of an attribute can be regarded as interchangeable. In the approach of Dubois. Prade and Testemale possibility distributions were used to represent all possible kinds of incompletely or fuzzily known values. This paper describes the application of a combination of both techniques. We focus on the very specific domain of criminal investigation, especially on criminal identification by means of a personal description. However, the developed method can be applied to a lot of other domains, where a similar sort of fuzziness and uncertainty shows up. For most attributes the fuzzy values are considered as primitive notions. Each domain is provided with a sup- W-transitive likeness relation in order to represent an existing kind of overlapping between the possible fuzzy values. Moreover, in order to restrict the number of attributes we allow multivalued ones: their values are represented as possibility distributions on the power class of the domain. The pattern-matching process uses a Prade-Testemale-like technique based on possibility distributions.