No matter how efficient indexing-based Internet search engines could be, indexing or inverse representations of data, is not in the mainstream of pattern recognition. One reason for a lack of interest in indexing methods on the part of pattern recognition community is instability of results due to a use of noise-prone measurements as features, rather than key words. The paper suggests a multidimensional numerical data indexing method that opens a path to accurate indexing-based pattern recognition systems that inherit from their search engines predecessors the ability to efficiently deal with large amounts of data.