Scientific applications usually handle large amount of data that have a proprietary and complex representation. This characteristics represent a great challenge for sharing data between scientific applications. The main goal of this work is to provide an architecture of software services that allows a flexible and efficient access to large amount of data served by such applications. Case estudies are presented to show the flexibility that we can achieve with this architecture. These experiments are strongly based in actual data used in scientific applications developed by Tecgraf/PUCRio. We also present an evaluation of different techniques of data encoding based on experiments conducted to measure the performance achieved by an implementation of the proposed architecture.