CHANGES - List of revisions
This document contains list of bug fixes and feature additions to Swish-e.
Per http://swish-e.org/archive/2008-06/12180.html
A global change was made to normalize index sizes for 64 bits.
This requires reindexing if you are upgrading from a previous version.
Issues included:
* http://dev.swish-e.org/ticket/14The new -D option at indexing time lets you specify an alternate char delimiter when property values are appended to one another at indexing time. The default value is the same as it always has been: a single space.
To match the libswish3 behavior, pass the ETX (\x03) ascii control character like this:
swish-e -D '\x03'It's up to your search code to process the property values accordingly.
The tar file of SWISH::Stemmer in the example dir has been broken for years, but since the module itself is severely deprecated we haven't bothered with it. Till now. Because it was preventing the good people of the fedora project from building their RPMs.
Setting ReturnRawRank to a true value will return the rank score unscaled. Can be set with the -a command line option (mnemonic: "a"bsolute rank score).
The ranking debugging feature using setenv introduced in 2.4.6 was yanked. Some platforms (notably HP-UX and Windows) lack the setenv feature, and the convenience of setting the env var was not worth the limitations.
Clark Vent reported that the query parser was not respecting MinWordLength settings. See http://dev.swish-e.org/changeset/2145
The file.c patch was in response to http://swish-e.org/archive/2007-03/11321.html although that user never responded about that patch.
Set SWISH_DEBUG_RANK to a true value to enable lots of rank debugging on stderr.
Recent versions of ExtUtils::MakeMaker revealed a bug in Makefile.PL. Patch from mschwern via RT, report by mpeters.
jrobinson852@yahoo.com suggest LARGEFILE support be auto-detected since it is needed so often on Linux systems.
Trygve Falch contributed patches to update the Snowball stemmers, including new Hungarian and Romanian stemmers.
Anthony Dovgal patched two leaks. One when there's a failure to open a file the file name was not freed.
SwishSetSearchLimit() was nulling the search limits when an error was found in the parameters, but not freeing the existing limits.
Fixed a leak if a limit was set and then reset but not prepared. Patch provided by Antony Dovgal.
Added SwishGetStructure() and SwishGetPhraseDelimiter() functions which return relevant properties of the search object. Patch provided by Antony Dovgal.
spider.pl was using the wrong method do uncompress HTTP responses that were 'deflate' encoded. Also decode content based on the document's charset and encode back to charset before outputting.
The magic numbers in src/swish.h were changed to require re-indexing from version 2.4.4 indexes. This should have been done in 2.4.4 as well, and anytime the index format changes. -- karman
stemmer.c had a mix up in the deprecated stemmer assignments for "Stemmer_en" and "Stem". Also fixed stemmer.h so that 2.4.3 indexes can be read correctly. -- karman
FileFilter* was using popen to run the filter, which could pass user data though the shell. Now uses fork/exec if fork is available which should be everywhere except Windows. In windows popen is used but all parameters are double-quoted. -- moseley
Cleaned up search.c to catch mismatched signedness warnings from newer GCC versions. This issue pre-existed 2.4.4 but the new wildcard features in search.c made for a lot more warnings. -- karman
Modified root Makefile to include the perl/Makefile.mingw file. -- karman
Release Candidate 1 for 2.4.4, 2 Oct 2006.
Ludovic Drolez contributed a patch to fix a quoting issue with filenames. This affects non-Windows builds only.
SWISH::Filter is now available on http://cpan.org/. The version in the distribution is not kept in sync with the CPAN version. Install the CPAN version if you want the latest and greatest version.
Added several fixes, including:
Herman Knoops contributed these patches. See http://swish-e.org/archive/2006-05/10543.html
Error messages were also changed to better reflect correct use of wildcards.
Fixed problem reported by Andreas Völter where a query that generated a two-word query with DoubleMetaphone fuzzy mode was not working.
Sorithy Seng (pourlassi@gmail.com) submitted a patch against docprop.c to fix an issue on sparc64 platforms. It is unknown whether this bug affected other 64-bit architectures.
Added check in db_native.c to check that some words exist before writing index.
Added doc for -u and -r options.
added fix to SWISH::Filters::pp2html and SWISH::Filters::XLtoHTML to save only filename as title without full path
The legacy Porter stemmer was removed. This had been deprecated some time ago. A warning will issue if the old stemmer is indicated in config file, and Stemmer_en1 will be used instead.
After a source code review, the developers decided to put Swish-e under the GPL with a special exception for linking against libswish-e. See http://swish-e.org/license.html for the details.
Dobrica Pavlinusic reported a segfaut after updating an index multiple times. José provided updated worddata.c. - April 27, 2005
Swish was returning results for deleted files when the NOT operator was used.
Thomas Angst reported swish consuming memory when using -S prog to process large number of empty documents.
When -S prog generated a zero length file the old parsers (e.g. TXT) would attempt to read in *all* content from the -S prog program into a buffer. The old parser incorrectly assumed it was reading from a filter and tried to read to eof().
The default value for ParserWarnLevel was changed form zero to two.
The ParserWarnLevel controls the error handling of the libxml2 parser. The higher the setting, the more verbose the output. The change to the default is to report when libxml2 has problems parsing a document (which often times results in processing only part of a document).
To get the old behavior, either set ParserWarnLevel to zero in your config file, or use the new -W command line option to set the ParserWarnLevel at run time. If ParserWarnLevel is set in the config file, it will override the -W option.
Also, to see UTF-8 to 8859-1 conversion errors set ParserWarnLevel to 3 or more. Previously, these warning were issues at ParserWarnLevel of one.
Removed all the target documentation (html, pdf, ps) from cvs. There's now a separate cvs module "swish_website" that is used to generate both the website and the html docs. If building swish-e from cvs please see the README.cvs file for instructions.
Gunnar Mätzler reported a problem with reading the pre-sorted property index tables when running with USE_BTREE (--enable-enremental). Not all entries were being written to disk. There was/is a question if the "array" code used for pre-sorted indexes with USE_BTREE would be slower. So, added a separate define USE_PRESORT_ARRAY to enable that code when USE_BTREE is set. This allows using the old integer arrays with USE_BTREE. Gunnar reported that this is working, but more testing is needed. Need to compare speed of the array code vs. the non-array code, and to verify the workings of USE_PRESORT_ARRAY code.
Andreas Seltenreich provided a patch to use strcoll when sorting properties. strcoll is locale dependent.
Jose fixed a problem with incremental indexing where a file could not be added back to the index once removed.
Patch initially provided by Dobrica Pavlinusic:
http://swish-e.org/Discussion/archive/2004-12/8694.htmlA change in the default way the index is compressed was not documented in 2.4.3. The change resulted in larger indexes. See CompressPositions below and in SWISH-CONFIG.
Fixed issue where a UTF-8 to Latin1 encoding failure would skip more input than just the failed character. Libxml2 passes swish text that is not null terminated, but the libxml2 functions to skip UTF-8 chars expected a null-terminated string. Replace libxml2 call with fixed version.
This option enables zlib compression for word data in the index. Previously word data was always compressed but resulted in slower wildcard searches. The default now is to not compress the word data, but results in larger index files. Set to "YES" to get pre-2.4.3 index sizes.
[This CHANGES entry was added after 2.4.3 was released]
There was a bit of confusion on how to use incremental indexing (still experimental) so added better logic for error messages.
Also fixed a logic error when setting the incremental update mode. Caught by Paul Loner.
Bernhard Weisshuhn supplied a patch to parser.c for checking the return value of UTF8Toisolat1(). Seems that libxml2 now returns the number of characters converted instead of zero for success.
http://bugzilla.gnome.org/show_bug.cgi?id=153937Swish now provides a swish-config script and config file for the pkg-config utility. These tools help when building programs that link with the swish-e library.
The SWISH::API Makefile.PL program uses swish-config to locate the installation directory of swish-e. This should make building SWISH::API easier when swish-e is installed in a non-standard location.
Peter van Dijk noticed that MetaNamesRank settings were not being copied to the output index when merging.
SwishFuzzy function (SWISH::API::Fuzzy) lets you stem a word without first searching. This might be helpful for playing with queries prior to the search.
Michael Levy found an error in the table used to translate 8859-1 to ascii7. Luckily, it was an upper case translation and the table is only used on lower case characters.
Changed the 'not yet implemented' caveat to 'implemented but experimental'.
You can now use continuation lines in the config file:
IgnoreWords \
the \
am \
is \
are \
wasThere may not be any characters following the backslash.
Words entered in config were not converted to lower case before storing in the index.
Peter Karman found an error when merging indexes where the source indexes had the same metanames, but listed in a different order in their config files. Words would then be indexed under the wrong metaID number in the output index.
The web spider spider.pl was updated to work better with SWISH::Filter by default and also make it easier to use the spider default along with a spider config file. See spider.pl for details.
SWISH::Filter was updated. The way filters are created has changed. If you created your own filters you will need to update them. Take a look at SWISH::Filter and the filters included in the distribution.
Richard Morin submitted formatting and punctuation dates to the README and INSTALL docs.
Added Inverse Document Frequency calculation to the getrank() routine. This will allow the relative frequency of a word in relationship to other words in the query to impact the ranking of documents.
Example: if 'foo' is present twice as often as 'bar' in the collection as a whole, a search for 'foo bar' will weight documents with 'bar' more heavily (i.e., higher rank) than those with 'foo'.
The impact is greatest when OR'ing words in a query rather than AND'ing them (which is the default).
Also added Rank discussion to the FAQ.
Updated PhraseHighlight.pm as suggested by Bill Schell for an optimization when all words in a document are highlighted.
Updated search.cgi and PhraseHighlight.pm to use the internal stemmers via the SWISH::API module as suggested by Jonas Wolf.
David Windmueller found a memory leak when calling multiple searches on a swish handle. The problem was swish loading the pre-sorted property index on every search, even after the table had been loaded into memory.
The example script swish.cgi uses an alarm (on platforms that support alarm) to abort processing after some number of seconds, but it was not killing the child process, swish-e. Bill Schell submitted a patch to kill the child when the alarm triggers.
The template was renamed because it's used by swish.cgi, not by search.cgi, which was confusing.
The example script search.cgi was updated to work better with mod_perl and to use external template files and style sheets.
James Job provided the SWISH::Filter::Doc2html filter that uses the wvWare (http://wvware.sourceforge.net/) program for filtering MS Word documents. If both catdoc and wvWare are installed then wvWare will be used.
wvWare is reported to do a good job at converting MS Word docs to HTML. In a few tests it did work well, but other cases it failed to generate correct output. It was also much, much slower than catdoc. I tested with wvWare 0.7.3 on Debian Linux. Testing with both is recommended.
John-Marc Chandonia pointed out that if a symlink is skipped by FileRules, then the actual file/directory is marked as "already seen" and cannot be indexed by other links or directly.
Now, files and directories are not marked "already seen" until after passing FileRules (i.e after a file is actually indexed or a directory is processed).
David Windmueller found a problem when trying to set the sort order more than once on an existing search object. Memory was not correctly reset after clearing the previous sort values.
Patch provided by Jamie Herre to access the MetaNames and PropertyNames via the C API and to test via the testlib program. Swish::API also updated to access this data.
David Windmueller reported that SwishResultPropertyULong() was returning ULONG_MAX on all calls. This was fixed.
Bill Schell with the help of valgrind found a null written past the end of a buffer in file.c in the code that supports the old parsers. This resulted in a segfault while indexing a large set of XML documents.
Steve Harris reported a problem when indexing a very large document that caused an integer overflow. José Ruiz updated to used unsigned integers.
Peter Karman pointed out the the libxml2 HTML parser was allowing phrase matches across block level html elements. Swish now bumps the word position on these elements.
UseStemming didn't take no for an answer
UseStemming was coded as an alias for FuzzyIndexingMode when Snowball was compiled in (the default), but "no" doesn't always mean no when the Norwegian stemmer is available.
Fixed problem building incremental version
Fixed compile problem with building incremental indexing mode. This is an experimental option with swish-e to allow adding files to an index. See configure --help for build option. Incremental indexes are not compatible with standard indexes.
Updated build instructions in INSTALL
Added a few comments about use of CPPFLAGS and LDFLAGS.
Updated the index_hypermail.pl
Updated to work with latest version of hypermail (pre-2.1.9).
Time zone in ResultPropertyStr()
Format string for generating date did not include the time zone in location. Add strftime format string to config.h
Undefined and Blank Properties and (NULL)
Fixed a few problems with printing properties:
1) Using -p and -x showed different results if a bad property value was given:
$ swish-e -w not dkdk -p badname -H0
err: Unknown Display property name "badname"
.
$ swish-e -w not dkdk -x '<badname>\n' -H0
(NULL)Now both return an error.
2) Fixed bug where using a "fmt" string with -x output generated (bad) output if the result did not have the specified property.
$ swish-e -w not dkdk -x '<somedate>\n' -H0 # undefined value
$ swish-e -w not dkdk -x '<somedate fmt="%Y %B %d">\n' -H0
%Y %B 1075353525Now nothing is printed if the property does not exist.
3) Updated SWISH::API to croak() on invalid property names, and to return undefined values for missing properties.
4) Updated swish.cgi and search.cgi to not generate warnings on undefined values return as properties. Note that swish.cgi will now die on undefined properties. Previously would just display (NULL).
Fixed segfault when generating warnings while parsing
Parser.c was incorrectly calling warning() incorrectly. And -Wall was not catching this!
Added check for internal property names.
Parser was not checking for use of Swish-e reserved property names.
<swishrank>foo</swishrank>This will now generate a warning.
Added new example CGI script
search.cgi is a new skeleton CGI script that uses SWISH::API for searching. It is installed in the same location as swish.cgi.
Add Fuzzy access to C and Perl interfaces
Added a number of functions to the C API (and SWISH::API) to access the stemmer used when indexing a given index.
Commas in numbers
Added commas to summary display at end of indexing.
Insert whitespace between tags
Parser.c was updated to flush the text buffer before and after every (non-inline HTML) tag.
The problem was that:
foo<tag>bar</tag>bazwould index as a single word "foobarbaz".
DirTree.pl
DirTree.pl was updated to work with SWISH::Filter and to work on Windows. DirTree.pl is a program to fetch files from the file system and works with the -S prog input method.
Problem with --enable-incremental option
Fixed configure script to build incremental option. Note that this is still experimental. But testers are welcome.
headers.c bug
Mark Fletcher with the help of valgrind found a bug in headers.c function SwishIndexHeaderNames used by the C API.
Clarify documentation regarding search order
At the prompting of Doralyn Rossmann updated SEARCH.pod to try and make the explanation of searching clearer, and to fix an error in the description of nested searches.
Note: Different Index Format
Swish-e version 2.4.0 has a different index file format from previous versions of Swish-e. Upgrading will require reindexing -- version 2.4.0 cannot read indexes created with previous versions.
robots.txt not closed correctly
When using -S http method robots.txt was not closed and that caused the (last) .contents file to not be unlinked under Windows. Windows seems to think filenames are related to files.
SWISH::Filter and locating programs on Windows
SWISH::Filter now scans $libexecdir in addition to the PATH for programs (such at catdoc and pdftotext), and also checks for programs by adding the extensions ".exe" and ".bat" to the program name.
Install sample templates
The sample templates included with swish.cgi are now installed in $pkgdatadir (typically /usr/local/share/swish-e).
Fix parser bug meta=(foo*)
Fixed bug in query parser caused in rc2's (pr2) attempt to catch wildcards errors.
Indexing HTML title
Fixed a problem when these were used in combination:
MetaNames swishtitle
MetaNameAlias swishtitle titleThat failed to correctly reset the metaname stack and indexed text under the wrong metaID.
Single Wildcards
Due to the way the query parser "works" a search of
"foo *"would result in a search of "foo*". Now that results in:
err: Single wildcard not allowed as word Fixed search parsing bug
Brad Miele reported that the word "andes" was not being found. It was being stemmed to "and" when was then considered an operator. [moseley]
Add new directive PropertyNamesSortKeyLength
PropertyNamesSortKeyLength sets the sort key length to use when sorting string properties. The default is 100 characters. There was a hard-coded 100 char limit before, but that was a problem where people were not building from source (Windows). The value of this is questionable -- it's intended to limit how much memory is used when sorting while indexing and searching. [moseley]
Fixed sorting issues with multiple indexes and reverse sorting
Reworked much of the sorting code. Still to do is setting the character sort order. [moseley]
Fixed minor memory leak
Fixed leak of not releasing memory of index file name and swish_handle destroy, and fixed SwishStemWord to default to the Stemmer_en. [moseley]
Fixed libtest.c example program that was not cleaning up memory after an error condition.
Replaced Swish-e's Porter Stemmer with Snowball
Swish-e now has support for Snowball stemmers (http://snowball.tartarus.org/). The stemmers are enabled for an index with FuzzyIndexingMode Stemming_* where "*" can be:
de, dk, en1, en2, es, fi, fr, it, nl, no, pt, ru, seIn addition, UseStemming yes or FuzzyIndexingMode Stemming_en will use the old stemmer.
Security Fix: swish.cgi
The swish.cgi script was not correctly escaping HTML when searching by the right combination of metanames and highlighting module. This could lead to cross-site scripting if indexing un-trusted documents. [moseley]
Added Support for building a Debian Package
To build as a .deb unpack the distribution and chdir then run
$ fakeroot debian/build binaryThen install the generated .deb file with dpkg -i
Use SWISH::Filter by default with spider.pl
spider.pl is installed in the libexecdir directory as well as the SWISH::Filter modules. PDF, MS Word, MP3, and XML documents will be indexed automatically if the required helper applications (e.g. catdoc, pdftotext) or scripts (e.g. MP3::Tag) are installed.
Swish also knows about libexecdir, so you you specify a relative path with -S prog swish-e will look for the program in libexecdir. This is mostly for spider.pl so indexing only requires:
IndexDir spider.pl
SwishProgParameters default http://localhost/index.htmlAnd swish-e will find spider.pl and SWISH::Filter will be used to convert docs.
Fixed Document-Type bug
Document-Type was not being reset after set input from a -S prog program causing the wrong parser to be used. [moseley]
New Directive: PropertyNamesNoStripChars
Swish replaces all series of low ASCII chars with a single space character. This option instructs swish to store all chars in the property. [moseley]
Change HTTP access defaults
Defaults used with -S http access method were changed.
Delay was reduced from one minute between start of each request to five seconds between requests.
MaxDepth was changed from five to zero, meaning there is no limit to depth indexed by default. [moseley]
swishspider location and SpiderDirectory
The swishspider program is now installed in $prefix/lib/swish-e by default. This can be changed by the --libexecdir option to configure.
The SpiderDirectory option now defaults to the value of libexecdir instead of the current directory. [moseley]
Added libtool and automake support
Replaces the build system with Autotools. Now builds libswish-e as a shared library on systems that support shared libraries. The swish-e binary links against this shared library. Can also build outside the source tree on platforms with GNU make. [moseley]
Updates to installation
Running "make install" now installs additional files. Files include the swish-e binary, the libswish-e search library, swish-e.h header, documentation files, the swishspider program, and Perl modules used for the example swish.cgi search script. Directories will be created if they do not already exist. Installation directories can be specified at build time.
Fixed bug when searching at end of inverted index
Swish was not correctly detecting the end of the inverted index when searching a wildcard word that was past the last word in the index. Caught by Frank Heasley. [moseley]
Increase sort key length from 50 to 100 characters
The setting MAX_SORT_STRING_LEN in src/config.h sets the max length used when sorting in swish-e. You may reduce this number to save memory while sorting, or increase it if you have very long properties to sort.
Remove " entity from -p output
The -p option to print properties was escaping double quotes in properties with the &quot; entity. -x does not do that, so inconsistent. -p no longer converts double quotes. The user should pick a good delimiter with -d or preferably use the -x method for generating output.
XML parser and Windows
The XML parser was being passed the incorrect buffer length when used on Windows platform causing the parser to abort with an error.
Version Numbering
SWISH-E versions starting with 2.3.4 use kernel version numbering. Versions are in the form: Major.Minor.Build. Odd minor versions are development. Even minor versions are releases. 2.3.4 would be a development version. 2.4.0 would be a release version. 2.3.20 would be the 20th build of 2.3.
Added RPM support
RPMs can be built with:
./configure
make distCopy the resulting tarball to RPM's SOURCES directory and then run as a superuser:
rpmbuild -ba rpm/swish-e.specYou should have swish-e packages in your RPMS/$arch directory. [augur]
Changed default perl binary location
Most perl scripts provided with SWISH-E now use /usr/bin/perl by default. Note that some scripts are generated at build time, so those will look in the path for the location of the perl binary.
New Feature: MetaNamesRank
MetaNamesRank can be used to adjust the ranking for words based on the word's MetaName.
New Swish Library API and Perl Module
The Swish-e C library interface was rewritten to provide better memory management and better separation of data. Most indexing related code has been removed from the library. A new header file is provided for the API: swish-e.h.
The Perl module SWISHE was replaced with the SWISH::API module in the Swish-e distribution.
Previous versions of the SWISHE module will not work with this version of Swish-e.
If you are using the SWISHE module from a previous version of Swish then you must either rewrite your code to use the new SWISH::API module (highly recommended) or use the replacement SWISHE module. The replacement SWISHE module is a thin interface to the SWISH::API module. It can be downloaded from
http://swish-e.org/Download/old/SWISHE-0.03.tar.gzNoContents not working with libxml2 parser
Corrected problem when using NoContents with binary files and the HTML2 parser.
Trying to index image file names with:
IndexOnly .gif .jpeg
NoContents .gif .jpegfailed to index the path names because the default parser (HTML2 when libxml2 is linked with swish-e) was not finding any text in the binary files. [moseley]
Updates to swish.cgi
The example/swish.cgi script can now use the SWISH::API module for searching an index. Combined with mod_perl this module can improve search performance considerably.
The Perl modules used with the swish.cgi script have all been moved into the SWISH::* namespace. Hence, files in the modules directory were moved into the modules::SWISH directory.
Multiple -L options were ORing instead of ANDing. Catch by Patrick Mouret. [moseley]
Pass non- text/* files onto indexing code IF there is a FileFilter associated with the *extension* of the URL. Fixes the problem of not being able to index, say, pdf files by using the FileFilter configuation option.
Fixed bug where nulls were stripped when using FileFilter with -S prog. Catch by Greg Fenton. [moseley]
NoContents with -S prog
Failed to use the correct default parser when using the No-Contents header and libxml2 linked in. [moseley]
Add tests for IRIX and sparc machines
8-byte alignment in mem_zones is is required for these machine [moseley]
Fixed code when removing files
Was not correctly removing words from index when parser aborted [jmruiz]
Merge segfault
Fixed segfault caused by trying to print null dates while merging duplicate files. [moseley]
Documentation patches
Spelling corrections to the SWISH-CONFIG pod page [Steve Eckert]
Configure corrections
Fixed a zlib test error that used "==" in a test [Steve Eckert]
Updates to VMS build
The VMS build was updated [Jean-François PIÉRONNE]
MANIFEST corrections
Added missing filters and vms build file into MANIFEST [moseley]
Default parser
Swish-e will now use the HTML2 (libxml2) parser by default if libxml2 is installed and DefaultContents or IndexContents is not used.
Selecting parsers
Allow HTML*, XML*, and TXT* to automatically select the libxml2-based parsers if libxml2 is linked with Swish-e, otherwise fallback to the built-in parsers.
SwishSpider and Filters
Filters (FileFilter directive) did not work correctly when spidering with the -S http method. A new filter system was developed and now filtering of documents (e.g. pdf->html or MSWord->text) is handled by the src/SwishSpider program.
When indexing with the -S http method only documents of content-type "text/*" are indexed. Other documents must be converted to text by using the filter system.
Buffer overflow in xml.c
Fixed bug in xml.c reported by Rodney Barnett when very long words were indexed. [moseley]
configure script updates
Updated from _WIN32 checks to feature checks using autoconf [moseley, norris]
updates to run on Alpha (Linux 2.4 (Debian 3.0))
Fixed a cast error when calling zlib, and the calls to read/write a packed longs to disk. [jmruiz, moseley]
COALESCE_BUFFER_MAX_SIZE
Some people were seeing the following error:
err: Buffer too short in coalesce_word_locations.
Increase COALESCE_BUFFER_MAX_SIZE in config.h and rebuild.This was due to indexing binary data or files with very large number of words. The best solution is to not index binary data or files with a very large number of words.
Swish-e will now automatically reallocate the buffer as needed. [jmruiz]
Many large changes were made internally in the code, some for performance reasons, some for feature changes and additions, and some to prepare for new features in later versions of Swish-e.
Documentation!
Documentation is now included in the source distribution as .pod (perldoc) files, and as HTML files. In addition, the distribution can now generate PDF, postscript, and unix man pages from the source .pod files. See README for more information.
Indexing and searching speed
The indexing process has been imporoved. Depending on a number of factors, you may see a significant improvement in indexing speed, especially if upgrading from version 1.x.
Searching speed has also been improved. Properties are not loaded until results are displayed, and properties are pre-sorted during indexing to speed up sorting results by properties while searching.
Properties are written to a sepearte file
Swish-e now stores document properties in a separate file. This means there are now two files that make up a Swish-e index. The default files are index.swish-e and index.swish-e.prop.
This change frees memory while indexing, allowing larger collections to be indexed in memory.
Internal data stored as Properties
Pre 2.2 some internal data was stored in fixed locations within the index, namely the file name, file size, and title. 2.2 introduced new internal data such as the last modified date, and document summaries. This data is considered meta data since it is data about a document.
Instead of adding new data to the internal structure of the index file, it was decided to use the MetaNames and PropertyNames feature of Swish-e to store this meta information. This allows for new meta data to be added at a later time (e.g. Content-type), and provides an easy and customizable way to print results with the -p switch and the new -x switch. In addition, search results can now be sorted and limited by properties.
For example, to sort by the rank and title:
swish-e -w foo -s swishrank desc swishtitle ascThe header display has been slightly reorganized.
If you are parsing output headers in a program then you may need to adjust your code. There's a new switch '-H' to control the level of header output when searching.
Results are now combined when searching more than one index.
Swish-e now merges (and sorts) the results from multiple indexes when using -f to specify more than one index. This change effects the way maxhits (-m) works. Here's a summary of the way it works for the different versions.
1.3.2 - MaxHits returns first N results starting from the first index.
e.g. maxhits=20; 15 hits Index1, 40 hits Index2
All 15 from Index1 plus first five from Index2 = 20 hits.
2.0.0 - MaxHits returns first N results from each index.
e.g. Maxhits=20; 15 hits Index1, 40 hits Index2
All 15 from Index1 plus 15 from Index2.
2.2.0 - Results are merged and first N results are returned.
e.g. Maxhits=20; 15 hits Index1, 40 hits Index2
Results are merged from each index and sorted
(rank is the default sort) and only the first
20 are returned.New prog document source indexing method
You can now use -S prog to use an external program to supply documents to Swish-e. This external program can be used to spider web servers, index databases, or to convert any type of document into html, xml, or text, so it can be indexed by Swish-e. Examples are given in the prog-bin directory.
The indexing parser was rewritten to be more logical.
TranslateCharacters now is done before WordCharacters is checked. For example,
WordCharacters abcdefghijklmnopqrstuvwxyz
TranslateCharacters ñ nNow El Niño will be indexed as El Nino (el and nino), even though ñ is not listed in WordCharacters.
Previously, stopwords were checked after stemming and soundex conversions, as well as most of the other word checks (WordCharacters, min/max length and so on). This meant that the stopword list probably didn't work as expected when using stemming.
The search parser was rewritten to be more logical
The search parser was rewritten to correct a number of logic errors. Swish-e did not differentiate between meta names, Swish-e operators and search words when parsing the query. This meant, for example, that metanames might be broken up by the WordCharacters setting, and that they could be stemmed.
Swish-e operator characters "*()= can now be searched by escaping with a backslash. For example:
./swish-e -w 'this\=odd\)word'will end up searching for the word this=odd)word. To search for a backslash character preceed it with a backslash.
Currently, searching for:
./swish-e -w 'this\*'is the same as a wildcard search. This may be fixed in the future.
Searching for buzzwords with those characters will still require backslashing. This also may change to allow some un-escaped operator characters, but some will always need to be escaped (e.g. the double-quote phrase character).
Quotes and Backslash escapes in strings
A bug was fixed in the parse_line() function (in string.c) where backslashes were not escaping the next character. parse_line() is used to parse a string of text into tokens (words). Normally splitting is done at whitespace. You may use quotes (single or double) to define a string (that might include whitespace) as a single parameter. The backslash can also be used to escape the following character when *within* quotes (e.g. to escape an embedded quote character).
ReplaceRules append "foo bar" <- define "foo bar" as a single word
ReplaceRules append "foo\"bar" <- escape the quotes
ReplaceRules append 'foo"bar' <- same thingExample user.config file removed.
Previous versions of Swish-e included a configuration file called user.config which contained examples of all directives. This has been replaced by a series of example configuration files located in the conf directory. The configuration directives are now described in SWISH-CONFIG.
Ports to Win32 and VMS
David Norris has included the files required to build Swish-e under Windows. See src/win32. A self-extracting Windows version is available from the Download page of the swish-e.org web site.
Jean-François Piéronne has provided the files required to build Swish-e under OpenVMS. See src/vms for more information.
String properties are concatenated
Multiple string properties of the same name in a document are now concatenated into one property. A space character is added between the strings if needed. A warning will be generated if multiple numeric or date properties are found in the same document, and the additional properties will be ignored.
Previously, properties of the same name were added to the index, but could not be retrieved.
To do: remove the next pointer, and allow user-defined character to place between properties.
regex type added to ReplaceRules
A more general purpose pattern replacement syntax.
New Parsers
Swish-e's XML parser was replaced with James Clark's expat XML parser library.
Swish-e can now use Daniel Veillard's libxml2 library for parsing HTML and XML. This requires installation of the library before building Swish-e. See the INSTALL document for information. libxml2 is not required, but is strongly recommended for parsing HTML over Swish-e's internal HTML parser, and provides more features for both HTML and XML parsing.
Support for zlib
Swish-e can be compiled with zlib. This is useful for compressing large properties. Building Swish-e with zlib is stronly recommended if you use its StoreDescription feature.
LST type of document no longer supported
LST allowed indexing of files that contained multiple documents.
Temporary files
To improve security Swish-e now uses the mkstemp(3) function to create temporary files. Temporary files are used while indexing only. This may result in some portability issues, but the security issues were overriding.
(Currently this does not apply to the -S http indexing method.)
mkstemp opens the temporary with O_EXCL|O_CREAT flags. This prevents overwriting existing files. In addition, the name of the file created is a lot harder to guess by attackers. The temporary file is created with only owner permissions.
Please report any portability issues on the Swish-e discussion list.
Temporary file locations
Swish-e now uses the environment variables TMPDIR, TMP, and TEMP (in that order) to decide where to write temporary files. The configuration setting of TmpDir will be used if none of the environment variables are set. Swish-e uses the current directory otherwise; there is no default temporary directory.
Since the environment variables override the configuration settings, a warning will be issued if you set TmpDir in the configuration file and there's also an environment variable set.
Temporary files begin with the letters "swtmp" (which can be changed in config.h), followed by two or more letters that indicate the type of temporary file, and some random characters to complete the file name. If indexing is aborted for some reason you may find these temporary files left behind.
New Fuzzy indexing method Double Metaphone
Based on Lawrence Philips' Metaphone algorithm, add two new methods of creating a fuzzy index (in addition to Stemming and Soundex).
Changes to Configuration File Directives. Please see SWISH-CONFIG for more info.
New directives: IndexContents and DefaultContents
The IndexContents directive assigns internal Swish-e document parsers to files based on their file type. The DefaultContents directive assigns a parser to be used on file that are not assigned a parser with IndexContents.
New directive: UndefinedMetaTags [error|ignore|index|auto]
This describes what to do when a meta tag is found in a document that is not listed in the MetaNames directive.
New directive: IgnoreTags
Will ignore text with the listed tags.
New directive: SwishProgParameters *list of words*
Passes words listed to the external Swish-e program when running with -S prog document source method.
New directive: ConvertHTMLEntities [yes|no]
Controls parsing and conversion of HTML entities.
New directive: DontBumpPositionOnMetaTags
The word position is now bumped when a new metatag is found -- this is to prevent phrases from matching across meta tags. This directive will disable this behavior for the listed tags.
This directive works for HTML and XML documents.
Changed directive: IndexComments
This has been changed such that comments are not indexed by default.
Changed directive: IgnoreWords
The builtin list of stopwords has been removed. Use of the SwishDefault word will generate a warning, and no stop words will be used. You must now specify a list of stopwords, or specify a file of stopwords.
A sample file stopwords.txt has been included in the conf/stopwords directory of the distribution, and can be used by the directive:
IgnoreWords File: /path/to/stopwords.txtChange of the default for IgnoreTotalWordCountWhenRanking
The default is now "yes".
New directive: Buzzwords
Buzzwords are words that should be indexed as-is, without checking for stopwords, word length, WordCharacters, or any other of the word limiting features. This allows indexing of things like C++ when "+" is not listed in WordCharacters.
Currenly, IgnoreFirstChar and IgnoreLastChar will be stripped before processing Buzzwords.
In the future we may use separate IgnoreFirst/Last settings for buzzwords since, for example, you may wish to index all + within Swish-e words, but strip + from the start/end of Swish-e words, but not from the buzzword C++.
New directives: PropertyNamesNumeric PropertyNamesDate
Before Swish-e 2.2 all user-defined document properties were stored in the index as strings. PropertyNamesNumeric and PropertyNamesDate tell it that a property should be stored in binary format. This allows for correct sorting of numeric properties.
Currenly, only integers can be stored, such as a unix timestamp. (Swish-e uses strtoul to convert the number to an unsigned long internally.)
PropertyNamesDate only indicates to Swish-e that a number is a unix timestamp, and to display the property as a formatted time when printing results. Swish does not currently parse date strings; you must provide a unix timestamp.
New directive: MetaNameAlias
You may now create alias names for MetaNames. This allow you to map or group multiple names to the same MetaName.
New directive: PropertyNameAlias
Creates aliases for a PropertyName.
New directive: PropertyNamesMaxLength
Sets the max length of a text property.
New directive: HTMLLinksMetaName
Defines a metaname to use for indexing href links in HTML documents. Available only with libxml2 parser.
New directive: ImageLinksMetaName
Defines a metaname to use for indexing src links in <img> tags. Allow you to search image pathnames within HTML pages. Available only with libxml2 parser.
New directive: IndexAltTagMetaName
Allows indexing of image ALT tags. Only available when using the libxml2 parser.
New directive: AbsoluteLinks
Attempts to convert relative links indexed with HTMLLinksMetaName and ImageLinksMetaName to absolute links. Available only with libxml2 parser.
New directive: ExtractPath
Allows you to use a regular expression to extract out part of the path of each file and index it with a meta name. For example, this allows searches to be limited to parts of your file tree.
New directive: FileMatch
FileMatch is similar to FileRules. Where FileRules is used to exclude files and directoires, FileMatch is used to include files.
New directive: PreSortedIndex
Controls which properties are pre-sorted while indexing. All properties are sorted by default.
New directive: ParserWarnLevel
Sets the level of warning printed when using libxml2.
New directive: obeyRobotsNoIndex [yes|NO]
When using libxml2 to parse HTML, Swish-e will skip files marked as NOINDEX.
<meta name="robots" content="noindex">Also, comments may be used within HTML and XML source docs to block sections of content from indexing:
<!-- SwishCommand noindex -->
<!-- SwishCommand index -->and/or these may be used also:
<!-- noindex -->
<!-- index -->New directive: UndefinedXMLAttributes
This describes how the content of XML attributes should be indexed, if at all. This is similar to UndefinedMetaTags, but is only for XML attributes and when parsed by libxml2. The default is to not index XML attributes.
New directive: XMLClassAttributes
XMLClassAttributes can specify a list of attribute names whose content is combined with the element name to form metanames.
New directive: PropCompressionLevel [0-9]
If compiled with zlib, Swish-e uses this setting to control the level of compression applied to properties. Properties must be long enough (defined in config.h) to be compressed. Useful for StoreDescription.
Experimental directive: IgnoreNumberChars
Defines a set of characters. If a word is made of of *only* those characters the word will not be indexed.
New directive: FuzzyIndexingMode
This configuration directive is used to define the type of "fuzzy" index to create. Currently the options are:
None
Stemming
Soundex
Metaphone
DoubleMetaphoneChanges to command line arguments. See SWISH-RUN for documentation on these switches.
New command line argument -H
Controls the level (verbosity) of header information printed with search results.
New command line argument -x
Provides additional header output and allows for a format string to describe what data to print.
New command line argument -k
Prints words stored in the Swish-e index.
New command line argument -N
Provides a way to do incremental indexing by comparing last modification dates. You pass -N a path to a file and only files newer than the last modified date of that file will be indexed.
Removed command line argument -D
-D no longer dumps the index file data. Use -T instead.
New command line argument -T
-T is used for debugging indexing and searching.
Enhanced command line argument -d
Now -d can accept some back-slashed characters to be used as output separators.
Enhanced command line argument -P
Now -P sets the phrase delimiter character in searches.
New command line argument -L
Swish-e 2.2 contains an experimental feature to limit results by a range of property values. This behavior of this feature may change in the future.
Modified command line argument -v
Now the argument -v 0 results in *no* output unless there is an error. This is a bit more handy when indexing with cron.
Hey! The above document had some coding errors, which are explained below:
You forgot a '=back' before '=head2'
Non-ASCII character seen before =encoding in 'Völter'. Assuming ISO8859-1