NAME DBIx::TextIndex - Perl extension for full-text searching in SQL databases SYNOPSIS use DBIx::TextIndex; my $index = DBIx::TextIndex->new({ doc_dbh => $doc_dbh, doc_table => 'doc_table', doc_fields => ['column_1', 'column_2'], doc_id_field => 'primary_key', index_dbh => $index_dbh, collection => 'collection_1', proximity_index => 1, errors => { empty_query => "your query was empty", quote_count => "phrases must be quoted correctly", no_results => "your seach did not produce any results", no_results_stop => "no results, these words were stoplisted: " }, language => 'en', # cz or en stoplist => [ 'en' ], max_word_length => 12, result_threshold => 5000, phrase_threshold => 1000, min_wildcard_length => 4, decode_html_entities => 1, print_activity => 0 }); $index->initialize; $index->add_doc(\@doc_ids); my $results = $index->search({ column_1 => '"a phrase" +and -not or', column_2 => 'more words', }); foreach my $doc_id (sort {$$results{$b} <=> $$results{$a}} keys %$results ) { print "DocID: $doc_id Score: $$results{$doc_id} \n"; } $index->delete; DESCRIPTION DBIx::TextIndex was developed for doing full-text searches on BLOB columns stored in a database. Almost any database with BLOB and DBI support should work with minor adjustments to SQL statements in the module. Implements a crude parser for tokenizing a user input string into phrases, can-include words, must-include words, and must-not-include words. Operates in case insensitive manner. The following methods are available: $index = DBIx::TextIndex->new(\%args) Constructor method. The first time an index is created, the following arguments must be passed to new(): my $index = DBIx::TextIndex->new({ doc_dbh => $doc_dbh, doc_table => 'doc_table', doc_fields => ['column_1', 'column_2'], doc_id_field => 'primary_key', index_dbh => $index_dbh, collection => 'collection_1' }); Other arguments are optional. doc_dbh DBI connection handle to database containing text documents doc_table Name of database table containing text documents doc_fields Reference to a list of column names to be indexed from doc_table doc_id_field Name of a unique integer key column in doc_table index_dbh DBI connection handle to database containing TextIndex tables. Using a separate database for your TextIndex is recommended, because the module creates and drops tables without warning. collection A name for the index. Should contain only alpha-numeric characters or underscores [A-Za-z0-9_] proximity_index Newer compressed proximity index is turned on by default as of version 0.22. "some phrase"~2 => matches "some nice phrase" "some phrase"~1 => matches only exact "some phrase" "some phrase"~10 => matches "some [1..9 words] phrase" Defaults to ~1 when omitted. The proximity matches work only forwards, not backwards, that means: "some phrase"~3 does not match "phrase nice some" or "phrase some" db In past versions, the database driver name was passed in this argument. As of version 0.22, the driver name is read from the database handle passed to index_dbh. Both MySQL (DBD::mysql) and PostgreSQL (DBD::Pg) are supported. errors This hash reference can be used to override default error messages. Please refer to the SYNOPSIS for meaning of the particular keys and values. charset Default is 'iso-8859-1'. Accented characters are converted to ASCII equivalents based on the charset. Pass 'iso-8859-2' for Czech or other Slavic languages. stoplist Activates stoplisting of very common words that are present in almost every document. Default is not to use stoplisting. Value of the parameter is a reference to array of two-letter language codes in lower case. Currently only two stoplists exist: en => English cz => Czech max_word_length Specifies maximum word length resolution. Defaults to 20 characters. result_threshold Defaults to 5000 documents. phrase_threshold Defaults to 1000 documents. decode_html_entities Decode html entities before indexing documents (e.g. & -> &). Default is 1. print_activity Activates STDOUT debugging. Higher value increases verbosity. After creating a new TextIndex for the first time, and after calling initialize(), only the index_dbh, doc_dbh, and collection arguments are needed to create subsequent instances of a TextIndex. $index->initialize This method creates all the inverted tables for the TextIndex in the database specified by doc_dbh. This method should be called only once when creating a new index! It drops all the inverted tables before creating new ones. initialize() also stores the doc_table, doc_fields, doc_id_field, language, stoplist, error attributes, proximity_index, max_word_length, result_threshold, phrase_threshold and min_wildcard_length preferences in a special table called "collection," so subsequent calls to new() for a given collection do not need those arguments. Calling initialize() will upgrade the collection table created by earlier versions of DBIx::TextIndex if necessary. $index->upgrade_collection_table Upgrades the collection table to the latest format. Usually does not need to be called by the programmer, because initialize() handles upgrades automatically. $index->add_doc(\@doc_ids) Add all the @docs_ids from doc_id_field to the TextIndex. All further calls to add_doc() must use @doc_ids higher than those previously added to the index. Reindexing previously-indexed documents will yield unpredictable results! $index->remove_doc(\@doc_ids) This method accepts a reference to an array of doc ids as its parameter. The specified documents will be removed from the index, but not from the actual documents table that is being indexed. The documents itself must be accessible when you remove them from the index. The ids should be sorted from lowest to highest. It's actually not possible to completely recover the space taken by the documents that are removed, therefore it's recommended to rebuild the index when you remove a significant amount of documents. $index->search(\%search_args) search() returns $results, a hash reference. The keys of the hash are doc ids, and the values are the relative scores of the documents. If an error occured while searching, search will throw a DBIx::TextIndex::Exception::Query object. eval { $results = $index->search({ first_field => '+andword -notword orword "phrase words"', second_field => ... ... }); }; if ($@) { if ($@->isa('DBIx::TextIndex::Exception::Query') { print "No results: " . $@->error . "\n"; } else { # Something more drastic happened $@->rethrow; } } else { print "The score for $doc_id is $results->{$doc_id}\n"; } $index->unscored_search(\%search_args) unscored_search() returns $doc_ids, a reference to an array. Since the scoring algorithm is skipped, this method is much faster than search(). A DBIx::TextIndex::Exception::Query object will be thrown if the query is bad or no results are found. eval { $doc_ids = $index->unscored_search({ first_field => '+andword -notword orword "phrase words"', second_field => ... }); }; if ($@) { if ($@->isa('DBIx::TextIndex::Exception::Query') { print "No results: " . $@->error . "\n"; } else { # Something more drastic happened $@->rethrow; } } else { print "Here's all the doc ids:\n"; map { print "$_\n" } @$doc_ids; } @doc_ids = $index->all_docs_ids all_doc_ids() return a list of all doc_ids currently in the index. $index->stat Allows you to obtain some meta information about the index. Accepts one parameter that specifies what you want to obtain. $index->stat('total_words') Returns a total count of words in the index. This number may differ from the total count of words in the documents itself. $index->delete delete() removes the tables associated with a TextIndex from index_dbh. SUPPORT FOR SEARCH MASKS DBIx::TextIndex can apply boolean operations on arbitrary lists of doc ids to search results. Take this table: doc_id category doc_full_text 1 green full text here ... 2 green ... 3 blue ... 4 red ... 5 blue ... 6 green ... Masks that represent doc ids for in each the three categories can be created: $index->add_mask($mask_name, \@doc_ids); $index->add_mask('green_category', [ 1, 2, 6 ]); $index->add_mask('blue_category', [ 3, 5 ]); $index->add_mask('red_category', [ 4 ]); The first argument is an arbitrary string, and the second is a reference to any array of doc ids that the mask name identifies. mask operations are passed in a second argument hash reference to $index->search(): %query_args = ( first_field => '+andword -notword orword "phrase words"', second_field => ... ... ); %args = ( not_mask => \@not_mask_list, and_mask => \@and_mask_list, or_mask => \@or_mask_list, or_mask_set => [ \@or_mask_list_1, \@or_mask_list_2, ... ], ); $index->search(\%query_args, \%args); not_mask For each mask in the not_mask list, the intersection of the search query results and all documents not in the mask is calculated. From our example above, to narrow search results to documents not in green category: $index->search(\%query_args, { not_mask => ['green_category'] }); and_mask For each mask in the and_mask list, the intersection of the search query results and all documents in the mask is calculated. This would give return results only in blue category: $index->search(\%query_args, { and_mask => ['blue_category'] }); Instead of using named masks, lists of doc ids can be passed on the fly as array references. This would give the same results as the previous example: my @blue_ids = (3, 5); $index->search(\%query_args, { and_mask => [ \@blue_ids ] }); or_mask_set With the or_mask_set argument, the union of all the masks in each list is computed individually, and then the intersection of each union set with the query results is calculated. or_mask An or_mask is treated as an or_mask_set with only one list. In this example, the union of blue_category and red_category is taken, and then the intersection of that union with the query results is calculated: $index->search(\%query_args, { or_mask => [ 'blue_category', 'red_category' ] }); $index->delete_mask($mask_name); Deletes a single mask from the mask table in the database. PARTIAL PATTERN MATCHING USING WILDCARDS You can use wildcard characters "*" or "?" at end of a word to match all words that begin with that word. Example: the "*" character means "match any characters" car* ==> matches "car", "cars", "careful", "cartel", .... the "?" character means "match also the plural form" car? ==> matches only "car" or "cars" The option min_wildcard_length is used to set the minimum number or characters appearing before the "*" wildcard character. The default is four characters to avoid selection of excessive amounts of word combinations. Unless this option is set to a lower value, the example above (car*) wouldn't produce any results. HIGHLIGHTING OF QUERY WORDS OR PATTERNS IN RESULTING DOCUMENTS A module HTML::Highlight can be used either independently or together with DBIx::TextIndex for this task. The HTML::Highlight module provides a very nice Google-like highligting using different colors for different words or phrases and also can be used to preview a context in which the query words appear in resulting documents. The module works together with DBIx::TextIndex using its new method html_highlight(). Check example script 'html_search.cgi' in the 'examples/' directory of DBIx::TextIndex distribution or refer to the documentation of HTML::Highlight for more information. CZECH LANGUAGE SUPPORT For czech diacritics insensitive operation you need to set the language option to 'cz'. my $index = DBIx::TextIndex->new({ .... language => 'cz', .... }); This option MUST be set for correct czech language proccessing. Diacritics sensitive operation is not possible. Requires the module "CzFast" that is available on CPAN in directory of author "TRIPIE". AUTHOR Daniel Koch, dkoch@bizjournals.com. COPYRIGHT Copyright 1997-2004 by Daniel Koch. All rights reserved. LICENSE This package is free software; you can redistribute it and/or modify it under the same terms as Perl itself, i.e., under the terms of the "Artistic License" or the "GNU General Public License". DISCLAIMER This package is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the "GNU General Public License" for more details. ACKNOWLEDGEMENTS Thanks to Jim Blomo, for PostgreSQL patches. Thanks to the lucy project (http://www.seg.rmit.edu.au/lucy/) for ideas and code for the Okapi scoring function. Simon Cozens' Lucene::QueryParser module was adapted to create the DBIx::TextIndex QueryParser module. Special thanks to Tomas Styblo, for first version of proximity index, Czech language support, stoplists, highlighting, document removal and many other improvements. Thanks to Ulrich Pfeifer for ideas and code from Man::Index module in "Information Retrieval, and What pack 'w' Is For" article from The Perl Journal vol. 2 no. 2. Thanks to Steffen Beyer for the Bit::Vector module, which enables fast set operations in this module. Version 5.3 or greater of Bit::Vector is required by DBIx::TextIndex. BUGS Documentation is not complete. Please feel free to email me (dkoch@bizjournals.com) with any questions or suggestions. SEE ALSO perl(1). Revision history for Perl extension DBIx::TextIndex. =head1 CHANGES 0.22 WARNING: indexes must be recreated from scratch. Rewrote proximity index code from scratch using XS. New index format is faster, more compact, and scaleable. Flag proximity_index now defaults to 1. Removed legacy_tf_idf scoring method. Fixed division-by-zero error with wildcard searches Added support for PostgreSQL 0.21 Fixed bug where proximity search was case-sensitive instead of case-insensitive Fixed problem where some punctuation characters in query were treated as query terms instead of ignoring the characters 0.20 Added inline fielded queries, e.g.: $index->search( field1 => 'field2:this field3:"that phrase" foo bar' ); In previous version, this search could only be represented like this: $index->search( field1 => 'foo bar', field2 => 'this', field3 => '"that phrase"'); The advantage of the new syntax is that fields can be combined in arbitary boolean expressions, where previously only the union of the fields was returned. For example: (field1:this OR field2:that) AND (field3:"foo bar" OR field1:foo) Wild and plural search bug fixes Turned proximity search back on, changed syntax. For example, to find "foo" followed by "bar" with 0, 1 or 2 words between: "foo bar"~3 0.19 Rewrote query parser. Now we support nested subqueries in parentheses, AND/OR/NOT operators, and existing '+/-' operators (this OR that) AND (foo OR bar) +"this phrase" -"not this phrase" Changed prefix query syntax slightly: '*' is now prefix operator, '?' is plural operator: dog* matches dog, dogs, dogma, dogmatic, dogmatically cat? matches cat, cats Rewrote critical loop of Okapi scoring as XSUB. Search over large collections is much faster. 0.18 Fixed bug where Okapi function did not score multiple fields Tweaks to improve boolean search performance. Added periodic flush of hash tables caches to save memory in long-running instances. Used a more compact represention of docweights, index needs to be recreated. 0.17 Reduced memory usage by TermDocsCache Implemented Okapi BM25 weighting function [S.E. Robertson et al., Google for description] for scoring documents. Pass scoring_method => 'legacy_tfidf' to new() or search() to use older method. Thanks to the lucy project (http://www.seg.rmit.edu.au/lucy/) for inspiration. Added docweights table to support Okapi BM25 scoring. Indexes need to be dropped and recreated. Pass "update_commit_interval" to new() to control memory usage in index updates. As the index is built, postings lists are built up in memory; this parameter controls the number of documents held in memory before being flushed to the database. Default is 20000. To disable, pass 0, and all documents passed to add_doc() will indexed in memory before writing to the database. Fixed problem with inaccurate search results if multiple searches were performed on a single instance while another instance was updating index. 0.16 Added XSUB pack_term_docs_append_vint to speed up _commit_docs when adding new postings to existing compressed postings 0.15 WARNING: indexes will have to be recreated after upgrading to this version Changed inverted list tables: renamed "docs" to "term_docs" and removed redundant column "docs_vector". Bit vectors are now built from postings stored in term_docs column. Total index size should be reduced significantly with removal of docs_vector column. Implemented DBIx::TextIndex::TermDocsCache to reduce database queries Fixed bug where single word queries for high doc frequency terms would return no results 0.14 Fixed memory leak in integer decoding 0.13 Changed format for inverted list postings to compressed binary format. Index sizes should be 50% smaller than with previous versions. Implemented XS code for fast coding and decoding. This format is similar to the format in Lucene (http://jakarta.apache.org/lucene/). Renamed 'freq_d' to 'docfreq_t'. Bug fix: add_doc will only set bits in all_docs_vector if document was actually added UPGRADE warning, indexes will have to be rebuilt. 0.12 Added _optimize_or_search(). If query contains just OR words, turn the rarest words into AND words to reduce result set size before scoring. Practically, this makes queries behave more like "all of the words" instead of "any of the words," which seems to be what the average user expects. Changed all symbols 'occurence' to 'freq_d' (frequency of docs in entire collection that contain term) Multi-field queries will return the intersection of all fields instead of the union. Added method all_doc_ids(). There is no longer any need to sort doc_ids before passing to add_doc(). Added more tests. WARNING: MySQL database 'test' must be available on localhost for tests to succeed. UPGRADE WARNING: collection table format changed, new index table collection_all_docs_vector added, some field names have changed in inverted tables. Any indexes created with 0.11 or earlier will have to be deleted and recreated. UPGRADE WARNING: all symbols with "document" have been renamed to "doc" for brevity. Methods have also been renamed, e.g. add_document() is now add_doc(). The old method names will work, but are deprecated. Replaced option 'language' with 'charset'. iso-8859-1 is the default charset. Added call to Text::Unaccent::unac_string (www.senga.org) to replace accented characters with plain ASCII equivalent. Uses 'charset' option to determine mapping. UPGRADE WARNING: added structured exceptions using Exception::Class. Calls to search() now have to be wrapped with eval blocks to catch query exceptions. 0.11 Bug fix: HTML tags are now changed to a single space, instead of empty string when indexing document. Prevents concatenation of words in some cases. 0.10 Fixed collection table upgrade bug 0.09 Changed $MAX_WORD_LENGTH default to 20 Allow numbers to be indexed as words Use HTML::Entities to decode entities in indexed documents, on by default. Set option decode_html_entities to 0 to disable. Use $dbh->tables to check for existence of tables (caution, may not work in DBI > 1.30). 0.08 Bug fix: add_mask() was not inserting masks 0.07 UPGRADE WARNING: collection table format changed, use new method $index->upgrade_collection_table() to recreate collection table. Calling initialize() method for a new collection will also upgrade collection table. Index backup recommended. Added error_ prefix to error message column names in collection table Added version column to collection table Added language column to collection table, removed czech_language column UPGRADE WARNING: instead of new({ czech_language => 1}), use new({ language => 'cz' }) Bug fix: _store_collection_info() error if stop lists are not used unscored_search() will now return a scalar error message if an error occurs in search search() will croak if passed an invalid field name to search on Added documentation for mask operations 0.06 tripie's patch v2 updates: - a bug in document removing proccess related to incorrect 'occurency' data updates when multi-field documents were removed, was fixed. The methods remove_document() and _inverted_remove() were affected. - a bug related to wildcards in queries in form of "+word +next%" or "+word% +next%" was fixed - a bug related to "%" wildcards used while searching of multi-field documents was fixed - a bug related to stoplists and phrases that contain a non-stoplisted word together with a stoplisted word was fixed - a new full-featured solution of highligting of query words or patterns in content of resulting documents was added I've written a new module HTML::Highlight, that can be used either independently or together with DBIx::TextIndex. Its advantages include: - it makes highlighting very easy - takes phrases and wildcards into account - supports diacritics insensitive highlighting for iso-8859-2 languages - takes HTML tags into account. That means when a user searches for for example 'font', than a FONT element in does not get "highlighted". The module provides a very nice Google-like highlighting using different colors for different words or phrases. The module works together with DBIx::TextIndex using its new method html_highlight(). The module can also be used to preview a context in which query words appear in resulting documents. - the old method highlight() was not changed nor removed for sake of compatibility with old code - I put a new 'html_search.cgi' script to examples/ to show how the new highlighting and context previewing works. - the HTML::Highlight module can be found on CPAN - http://www.cpan.org/authors/id/T/TR/TRIPIE/ - the new highlighting solution has been documented in a new section of the documentation tripie v1 changes: NOTE: tripie proximity index is replaced with newer compact index structure as of version 0.22 added proximity indexing - based on positions of words in a document - by default it is disabled, activate it by new option "proximity_index" - very efficient for bigger documents, much worse for small ones - it's very big (approx. 20 bytes for each word) - allows fast proximity based searches in form of: ":2 some phrase" => matches "some nice phrase" ":1 some phrase" => matches only exact "some phrase" ":10 some phrase" => matches "some [1..9 words] phrase" defaults to ":1" when omitted - the proximity matches work only forwards, not backwards, that means: ":3 some phrase" does not match "phrase nice some" or "phrase some" NOTE: as of version 0.20, this syntax has changed: "some phrase"~3 rewrote the word splitter and query parser - added support for czech language diacritics insensitive indexing and searching (option "czech_language") (note: changed option to "charset", pass value "iso-8859-2" to enable -dkoch) that is implemented by converting both the indexed data and the query from iso-8859-2 to ASCII - this can also be used for other iso-8859-2 based Slavic languages - the above is performed by my module "CzFast" that can be found on CPAN (my CPAN id is "TRIPIE"), and is optional added partial pattern matching using wildcards "%" or "*" - these wildcards can be used at end of a word to match all words that begin with that word, ie. the "%" character means "match any characters" car% ==> matches "car", "cars", "careful", "cartel", .... the "*" character means "match also the plural form" car* ==> matches only "car" or "cars" - added option "min_wildcard_length" to specify minimal length of a word base appearing before the "%" wildcard character to avoid selection of excessive amount of results added a database abstraction layer - all SQL queries were moved to separate module (see lib/ and the docs for new "db" option) and polished a bit for better maintainability and possible support of other SQL dialects added stoplists - some words that are too common (are present in almost every document) are not indexed and are removed from the search query before processing to avoid expensive processing of excessively huge result sets - user is notified when a search does not produce any results because some words he used in his query were stoplisted (no_results_stop) - stoplists can be easily localized or modified - default is not to use any stoplist, one or more stoplists can be selected using the "stoplist" option - stoplist data files are in lib/ - english (en) and czech (cz) stoplists are included - more than one stoplist can be used added a facility to remove documents from the index - check the documentation for new method "remove_document" for more info. There is no way to recover all the space taken by the documents that are being removed. This method manages to recover approx. 80% of the space. It's recommended to rebuild the index when you remove a significant amount of documents. added a facility to obtain some statistical information about the index - check the documentation for new method "stat" max_word_length, phrase_threshold and result_threshold are now configurable options added configuration options to customize/localize the error messages (no_results etc.) all new configuration options are properly stored in collection's data max_word_length limit now works much better - all words are stripped down to the maximum size before they are stored to the index, and also all query words are stripped down to the maximum word size before they are proccessed. Now when the max word length is set to, say, six and a user searches for for example "consciousness", all documents containing any words beginning with "consci" are returned Therefore the new max_word_length option is not a limit of a word size, but rather a "resolution" preference. added some comments and occasionally corrected indentation documented the enhancements bugfix: when RaiseError was set on a DBI connection, then one query which only switched off PrintError to avoid some problems, failed note: the interface was not changed - old code using this module should run without any changes Thanks for this excellent module and please excuse my inferior English ! Tomas Styblo, tripie@cpan.org 0.05 Added unscored_search() which returns a reference to an array of doc_ids, without scores. Should be much faster than scored search. Added error handling in case _occurence() doesn't return a number. 0.04 Bug fix: add_document() will return if passed empty array ref instead of producing error. Changed _boolean_compare() and _phrase_search() so and_words and phrases behave better in multiple-field searches. Result set for each field is calculated first, then union of all fields is taken for final result set. Scores are scaled lower in _search(). 0.03 Added example scripts in examples/. 0.02 Added or_mask_set. 0.01 Initial public release. Should be considered beta, and methods may be added or changed until the first stable release.