We'll probably need to use a cache (lowercase) to avoid invoking ANTLR on each request.
Ideally we could think of caching parameterized templates, or recognize constants: in case of constants the output could degenerate in caching a Lucene Filter.

Priority do be defined after performance tests.