The Perl Toolchain Summit needs more sponsors. If your company depends on Perl, please support this very important event.

NAME

KinoSearch::Docs::Cookbook::CustomQueryParser - Sample subclass of QueryParser.

ABSTRACT

Implement a custom search query language using KinoSearch::QueryParser and Parse::RecDescent.

Grammar-based vs. hand-rolled

There are two classic strategies for writing a text parser.

  1. Create a grammar-based parser using Perl modules like Parse::RecDescent or Parse::YAPP, C utilities like lex and yacc, etc.

  2. Hand-roll your own parser.

We'll start off with hand-rolling, but we'll ultimately move to the grammar-based parsing technique because of its superior flexibility.

The language

At first, our query language will support only simple term queries and phrases delimited by double quotes. For simplicity's sake, it will not support parenthetical groupings, boolean operators, or prepended plus/minus. The results for all subqueries will be unioned together -- i.e. joined using an OR -- which is usually the best approach for small-to-medium-sized document collections.

Later, we'll add support for trailing wildcards.

Single-field regex-based parser

Hand-rolling a parser can be labor-intensive, but our proposed query language is simple enough that chewing up the query string with some simple regular expressions will do the trick.

We'll use a fixed field name of "content", and a fixed choice of English PolyAnalyzer.

    package FlatQueryParser;
    use KinoSearch::Search::TermQuery;
    use KinoSearch::Search::PhraseQuery;
    use KinoSearch::Search::ORQuery;
    use Carp;
    
    sub new { 
        my $analyzer = KinoSearch::Analysis::PolyAnalyzer->new(
            language => 'en',
        );
        return bless { 
            field    => 'content',
            analyzer => $analyzer,
        }, __PACKAGE__;
    }

Some private helper subs for creating TermQuery and PhraseQuery objects will help keep the size of our main parse() subroutine down:

    sub _make_term_query {
        my ( $self, $term ) = @_;
        return KinoSearch::Search::TermQuery->new(
            field => $self->{field},
            term  => $term,
        );
    }
    
    sub _make_phrase_query {
        my ( $self, $terms ) = @_;
        return KinoSearch::Search::PhraseQuery->new(
            field => $self->{field},
            terms => $terms,
        );
    }

Our private _tokenize() method treats double-quote delimited material as a single token and splits on whitespace everywhere else.

    sub _tokenize {
        my ( $self, $query_string ) = @_;
        my @tokens;
        while ( length $query_string ) {
            if ( $query_string =~ s/^\s*// ) {
                next;    # skip whitespace
            }
            elsif ( $query_string =~ s/^("[^"]*(?:"|$))// ) {
                push @tokens, $1;    # double-quoted phrase
            }
            else {
                $query_string =~ s/(\S+)//;
                push @tokens, $1;    # single word
            }
        }
        return \@tokens;
    }

The main parsing routine creates an array of tokens by calling _tokenize(), runs the tokens through through the PolyAnalyzer, creates TermQuery or PhraseQuery objects according to how many tokens emerge from the PolyAnalyzer's split() method, and adds each of the sub-queries to the primary ORQuery.

    sub parse {
        my ( $self, $query_string ) = @_;
        my $tokens   = $self->_tokenize($query_string);
        my $analyzer = $self->{analyzer};
        my $or_query = KinoSearch::Search::ORQuery->new;
    
        for my $token (@$tokens) {
            if ( $token =~ s/^"// ) {
                $token =~ s/"$//;
                my $terms = $analyzer->split($token);
                my $query = $self->_make_phrase_query($terms);
                $or_query->add_child($phrase_query);
            }
            else {
                my $terms = $analyzer->split($token);
                if ( @$terms == 1 ) {
                    my $query = $self->_make_term_query( $terms->[0] );
                    $or_query->add_child($query);
                }
                elsif ( @$terms > 1 ) {
                    my $query = $self->_make_phrase_query($terms);
                    $or_query->add_child($query);
                }
            }
        }
    
        return $or_query;
    }

Single-field Parse::RecDescent-based parser

Instead of using regular expressions to tokenize the string, we can use Parse::RecDescent.

    my $grammar = <<'END_GRAMMAR';
    
    leaf_queries:
        leaf_query(s?)
        { $item{'leaf_query(s?)'} }
    
    leaf_query:
          phrase_query
        | term_query
    
    term_query:
        /(\S+)/
        { $1 }
    
    phrase_query:
        /("[^"]*(?:"|$))/   # terminated by either quote or end of string
        { $1 }
    
    END_GRAMMAR
    
    sub new { 
        my $analyzer = KinoSearch::Analysis::PolyAnalyzer->new(
            language => 'en',
        );
        my $rd_parser = Parse::RecDescent->new($grammar);
        return bless { 
            field     => 'content',
            analyzer  => $analyzer,
            rd_parser => $rd_parser,
        }, __PACKAGE__;
    }

The behavior of a Parse::RecDescent parser based on the grammar above is exactly the same as that of our regex-based tokenization routine from before, so we can leave parse() intact and simply change _tokenize():

    sub _tokenize {
        my ( $self, $query_string ) = @_;
        return $self->{rd_parser}->leaf_queries($query_string);
    }

Multi-field Parse::RecDescent-based parser

Most often, the end user will want their search query to match not only a single 'content' field, but also 'title' and so on. To make that happen, we have to turn queries such as this...

    foo AND NOT bar

... into the logical equivalent of this:

    (title:foo OR content:foo) AND NOT (title:bar OR content:bar)

Rather than continue with our own from-scratch parser class and write the routines to accomplish that expansion, we're now going to subclass KinoSearch::QueryParser and take advantage of some of its existing methods.

Our first parser implementation had the "content" field name and the choice of English PolyAnalyzer hard-coded for simplicity, but we don't need to do that this time -- KinoSearch::QueryParser's constructor requires a Schema which conveys field and Analyzer information, so we can just defer to that.

    package FlatQueryParser;
    use base qw( KinoSearch::QueryParser );
    use KinoSearch::Search::TermQuery;
    use KinoSearch::Search::PhraseQuery;
    use KinoSearch::Search::ORQuery;
    use KinoSearch::Search::NoMatchQuery;
    use PrefixQuery;
    use Parse::RecDescent;
    use Carp;
    
    our %rd_parser;
    
    sub new { 
        my $class = shift;
        my $self = $class->SUPER::new(@_);
        $rd_parser{$$self} = Parse::RecDescent->new($grammar);
        return $self;
    }
    
    sub DESTROY {
        my $self = shift;
        delete $rd_parser{$$self};
        $self->SUPER::DESTROY;
    }

If we modify our Parse::RecDescent grammar slightly, we can eliminate the _tokenize(), _make_term_query(), and _make_phrase_query() helper subs, and our parse() subroutine can be chopped way down. We'll have the term_query and phrase_query productions generate LeafQuery objects, and add a tree production which joins the leaves together with an ORQuery.

    my $grammar = <<'END_GRAMMAR';
    
    tree:
        leaf_queries
        { 
            $return = KinoSearch::Search::ORQuery->new;
            $return->add_child($_) for @{ $item[1] };
        }
    
    leaf_queries:
        leaf_query(s?)
        { $item{'leaf_query(s)'} }
    
    leaf_query:
          phrase_query
        | term_query
    
    term_query:
        /(\S+)/
        { KinoSearch::Search::LeafQuery->new( text => $1 ) }
    
    phrase_query:
        /("[^"]*(?:"|$))/   # terminated by either quote or end of string
        { KinoSearch::Search::LeafQuery->new( text => $1 ) }
    
    END_GRAMMAR
    
    ...
    
    sub parse {
        my ( $self, $query_string ) = @_; 
        my $tree = $self->tree($query_string);
        return $tree ? $self->expand($tree) :
        KinoSearch::Search::NoMatchQuery->new;
    }

    sub tree {
        my ( $self, $query_string ) = @_; 
        return $rd_parser{$$self}->tree($query_string);
    }

The magic happens in KinoSearch::QueryParser's expand() method, which walks the ORQuery object we supply to it looking for LeafQuery objects, and calls expand_leaf() for each one it finds. expand_leaf() performs field-specific analysis, decides whether each query should be a TermQuery or a PhraseQuery, and if multiple fields are required, creates an ORQuery which mults out e.g. foo into (title:foo OR content:foo).

Extending the query language

To add support for trailing wildcards to our query language, first we need to modify our grammar, adding a prefix_query production and tweaking the leaf_query production to accommodate it.

    leaf_query:
          phrase_query
        | prefix_query
        | term_query
    
    prefix_query:
        /(\w+\*)/
        { KinoSearch::Search::LeafQuery->new( text => $1 ) }

Second, we need to override expand_leaf() to accommodate PrefixQuery, while deferring to its original implementation on TermQuery and PhraseQuery.

    sub expand_leaf {
        my ( $self, $leaf_query ) = @_;
        my $text = $leaf_query->get_text;
        if ( $text =~ /\*$/ ) {
            my $or_query = KinoSearch::Search::ORQuery->new;
            for my $field ( @{ $self->get_fields } ) {
                my $prefix_query = PrefixQuery->new(
                    field        => $field,
                    query_string => $text,
                );
                $or_query->add_child($prefix_query);
            }
            return $or_query;
        }
        else {
            return $self->SUPER::expand_leaf($leaf_query);
        }
    }

Usage

Insert any of our custom parsers into the search.cgi sample app to get a feel for how they behave:

    my $parser = FlatQueryParser->new( schema => $searcher->get_schema );
    my $query  = $parser->parse( $cgi->param('q') || '' );
    my $hits   = $searcher->hits(
        query      => $query,
        offset     => $offset,
        num_wanted => $hits_per_page,
    );
    ...

COPYRIGHT

Copyright 2008-2009 Marvin Humphrey

LICENSE, DISCLAIMER, BUGS, etc.

See KinoSearch version 0.30.