From "New flexible query parser" thread by Micheal Busch
in my team at IBM we have used a different query parser than Lucene's in
our products for quite a while. Recently we spent a significant amount
of time in refactoring the code and designing a very generic
architecture, so that this query parser can be easily used for different
products with varying query syntaxes.
This work was originally driven by Andreas Neumann (who, however, left
our team); most of the code was written by Luis Alves, who has been a
bit active in Lucene in the past, and Adriano Campos, who joined our
team at IBM half a year ago. Adriano is Apache committer and PMC member
on the Tuscany project and getting familiar with Lucene now too.
We think this code is much more flexible and extensible than the current
Lucene query parser, and would therefore like to contribute it to
Lucene. I'd like to give a very brief architecture overview here,
Adriano and Luis can then answer more detailed questions as they're much
more familiar with the code than I am.
The goal was it to separate syntax and semantics of a query. E.g. 'a AND
b', '+a +b', 'AND(a,b)' could be different syntaxes for the same query.
We distinguish the semantics of the different query components, e.g.
whether and how to tokenize/lemmatize/normalize the different terms or
which Query objects to create for the terms. We wanted to be able to
write a parser with a new syntax, while reusing the underlying
semantics, as quickly as possible.
In fact, Adriano is currently working on a 100% Lucene-syntax compatible
implementation to make it easy for people who are using Lucene's query
parser to switch.
The query parser has three layers and its core is what we call the
QueryNodeTree. It is a tree that initially represents the syntax of the
original query, e.g. for 'a AND b':
The three layers are:
1. The upper layer is the parsing layer which simply transforms the
query text string into a QueryNodeTree. Currently our implementations of
this layer use javacc.
2. The query node processors do most of the work. It is in fact a
configurable chain of processors. Each processors can walk the tree and
modify nodes or even the tree's structure. That makes it possible to
e.g. do query optimization before the query is executed or to tokenize
3. The third layer is also a configurable chain of builders, which
transform the QueryNodeTree into Lucene Query objects.
Furthermore the query parser uses flexible configuration objects, which
are based on AttributeSource/Attribute. It also uses message classes that
allow to attach resource bundles. This makes it possible to translate
messages, which is an important feature of a query parser.
This design allows us to develop different query syntaxes very quickly.
Adriano wrote the Lucene-compatible syntax in a matter of hours, and the
underlying processors and builders in a few days. We now have a 100%
compatible Lucene query parser, which means the syntax is identical and
all query parser test cases pass on the new one too using a wrapper.
Recent posts show that there is demand for query syntax improvements,
e.g improved range query syntax or operator precedence. There are
already different QP implementations in Lucene+contrib, however I think
we did not keep them all up to date and in sync. This is not too
surprising, because usually when fixes and changes are made to the main
query parser, people don't make the corresponding changes in the contrib
parsers. (I'm guilty here too)
With this new architecture it will be much easier to maintain different
query syntaxes, as the actual code for the first layer is not very much.
All syntaxes would benefit from patches and improvements we make to the
underlying layers, which will make supporting different syntaxes much