CHANGES
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 results0.14 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 bug0.09
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 masks0.07
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 operations0.06
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 <FONT COLOR="red">
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
highligting 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:
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"
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.org0.05
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.