NAME

The Swish-e FAQ - Answers to Common Questions

OVERVIEW

List of commonly asked and answered questions. Please review this document before asking questions on the Swish-e discussion list.

General Questions

What is Swish-e?

Swish-e is Simple Web Indexing System for Humans - Enhanced. With it, you can quickly and easily index directories of files or remote web sites and search the generated indexes for words and phrases.

So, is Swish-e a search engine?

Well, yes. Probably the most common use of Swish-e is to provide a search engine for web sites. The Swish-e distribution includes CGI scripts that can be used with it to add a search engine for your web site. The CGI scripts can be found in the example directory of the distribution package. See the README file for information about the scripts.

But Swish-e can also be used to index all sorts of data, such as email messages, data stored in a relational database management system, XML documents, or documents such as Word and PDF documents -- or any combination of those sources at the same time. Searches can be limited to fields or MetaNames within a document, or limited to areas within an HTML document (e.g. body, title). Programs other than CGI applications can use Swish-e, as well.

Should I upgrade if I'm already running a previous version of Swish-e?

A large number of bug fixes, feature additions, and logic corrections were made in version 2.2. In addition, indexing speed has been drastically improved (reports of indexing times changing from four hours to 5 minutes), and major parts of the indexing and search parsers have been rewritten. There's better debugging options, enhanced output formats, more document meta data (e.g. last modified date, document summary), options for indexing from external data sources, and faster spidering just to name a few changes. (See the CHANGES file for more information.

Since so much effort has gone into version 2.2, support for previous versions will probably be limited.

Are there binary distributions available for Swish-e on platform foo?

Foo? Well, yes there are some binary distributions available. Please see the Swish-e web site for a list at http://swish-e.org/.

In general, it is recommended that you build Swish-e from source, if possible.

Do I need to reindex my site each time I upgrade to a new Swish-e version?

At times it might not strictly be necessary, but since you don't really know if anything in the index has changed, it is a good rule to reindex.

What's the advantage of using the libxml2 library for parsing HTML?

Swish-e may be linked with libxml2, a library for working with HTML and XML documents. Swish-e can use libxml2 for parsing HTML and XML documents.

The libxml2 parser is a better parser than Swish-e's built-in HTML parser. It offers more features, and it does a much better job at extracting out the text from a web page. In addition, you can use the ParserWarningLevel configuration setting to find structural errors in your documents that could (and would with Swish-e's HTML parser) cause documents to be indexed incorrectly.

Libxml2 is not required, but is strongly recommended for parsing HTML documents. It's also recommended for parsing XML, as it offers many more features than the internal Expat xml.c parser.

The internal HTML parser will have limited support, and does have a number of bugs. For example, HTML entities may not always be correctly converted and properties do not have entities converted. The internal parser tends to get confused when invalid HTML is parsed where the libxml2 parser doesn't get confused as often. The structure is better detected with the libxml2 parser.

If you are using the Perl module (the C interface to the Swish-e library) you may wish to build two versions of Swish-e, one with the libxml2 library linked in the binary, and one without, and build the Perl module against the library without the libxml2 code. This is to save space in the library. Hopefully, the library will someday soon be split into indexing and searching code (volunteers welcome).

Does Swish-e include a CGI interface?

Yes. Kind of.

There's two example CGI scripts included, swish.cgi and search.cgi. Both are installed at $prefix/lib/swish-e.

Both require a bit of work to setup and use. Swish.cgi is probably what most people will want to use as it contains more features. Search.cgi is for those that want to start with a small script and customize it to fit their needs.

An example of using swish.cgi is given in the INSTALL man page, and it the swish.cgi documentation. Like often is the case, it will be easier to use if you first read the documentation.

Please use caution about CGI scripts found on the Internet for use with Swish-e. Some are not secure.

The included example CGI scripts were designed with security in mind. Regardless, you are encouraged to have your local Perl expert review it (and all other CGI scripts you use) before placing it into production. This is just a good policy to follow.

How secure is Swish-e?

We know of no security issues with using Swish-e. Careful attention has been made with regard to common security problems such as buffer overruns when programming Swish-e.

The most likely security issue with Swish-e is when it is run via a poorly written CGI interface. This is not limited to CGI scripts written in Perl, as it's just as easy to write an insecure CGI script in C, Java, PHP, or Python. A good source of information is included with the Perl distribution. Type perldoc perlsec at your local prompt for more information. Another must-read document is located at http://www.w3.org/Security/faq/wwwsf4.html.

Note that there are many free yet insecure and poorly written CGI scripts available -- even some designed for use with Swish-e. Please carefully review any CGI script you use. Free is not such a good price when you get your server hacked...

Should I run Swish-e as the superuser (root)?

No. Never.

What files does Swish-e write?

Swish writes the index file, of course. This is specified with the IndexFile configuration directive or by the -f command line switch.

The index file is actually a collection of files, but all start with the file name specified with the IndexFile directive or the -f command line switch.

For example, the file ending in .prop contains the document properties.

When creating the index files Swish-e appends the extension .temp to the index file names. When indexing is complete Swish-e renames the .temp files to the index files specified by IndexFile or -f. This is done so that existing indexes remain untouched until it completes indexing.

Swish-e also writes temporary files in some cases during indexing (e.g. -s http, -s prog with filters), when merging, and when using -e). Temporary files are created with the mkstemp(3) function (with 0600 permission on unix-like operating systems).

The temporary files are created in the directory specified by the environment variables TMPDIR and TMP in that order. If those are not set then swish uses the setting the configuration setting TmpDir. Otherwise, the temporary file will be located in the current directory.

Can I index PDF and MS-Word documents?

Yes, you can use a Filter to convert documents while indexing, or you can use a program that "feeds" documents to Swish-e that have already been converted. See Indexing below.

Can I index documents on a web server?

Yes, Swish-e provides two ways to index (spider) documents on a web server. See Spidering below.

Swish-e can retrieve documents from a file system or from a remote web server. It can also execute a program that returns documents back to it. This program can retrieve documents from a database, filter compressed documents files, convert PDF files, extract data from mail archives, or spider remote web sites.

Can I implement keywords in my documents?

Yes, Swish-e can associate words with MetaNames while indexing, and you can limit your searches to these MetaNames while searching.

In your HTML files you can put keywords in HTML META tags or in XML blocks.

META tags can have two formats in your source documents:

    <META NAME="DC.subject" CONTENT="digital libraries">

And in XML format (can also be used in HTML documents when using libxml2):

    <meta2>
        Some Content
    </meta2>

Then, to inform Swish-e about the existence of the meta name in your documents, edit the line in your configuration file:

    MetaNames DC.subject meta1 meta2

When searching you can now limit some or all search terms to that MetaName. For example, to look for documents that contain the word apple and also have either fruit or cooking in the DC.subject meta tag.

What are document properties?

A document property is typically data that describes the document. For example, properties might include a document's path name, its last modified date, its title, or its size. Swish-e stores a document's properties in the index file, and they can be reported back in search results.

Swish-e also uses properties for sorting. You may sort your results by one or more properties, in ascending or descending order.

Properties can also be defined within your documents. HTML and XML files can specify tags (see previous question) as properties. The contents of these tags can then be returned with search results. These user-defined properties can also be used for sorting search results.

For example, if you had the following in your documents

   <meta name="creator" content="accounting department">

and creator is defined as a property (see PropertyNames in SWISH-CONFIG) Swish-e can return accounting department with the result for that document.

    swish-e -w foo -p creator

Or for sorting:

    swish-e -w foo -s creator

What's the difference between MetaNames and PropertyNames?

MetaNames allows keywords searches in your documents. That is, you can use MetaNames to restrict searches to just parts of your documents.

PropertyNames, on the other hand, define text that can be returned with results, and can be used for sorting.

Both use meta tags found in your documents (as shown in the above two questions) to define the text you wish to use as a property or meta name.

You may define a tag as both a property and a meta name. For example:

   <meta name="creator" content="accounting department">

placed in your documents and then using configuration settings of:

    PropertyNames creator
    MetaNames creator

will allow you to limit your searches to documents created by accounting:

    swish-e -w 'foo and creator=(accounting)'

That will find all documents with the word foo that also have a creator meta tag that contains the word accounting. This is using MetaNames.

And you can also say:

    swish-e -w foo -p creator

which will return all documents with the word foo, but the results will also include the contents of the creator meta tag along with results. This is using properties.

You can use properties and meta names at the same time, too:

    swish-e -w creator=(accounting or marketing) -p creator -s creator

That searches only in the creator meta name for either of the words accounting or marketing, prints out the contents of the contents of the creator property, and sorts the results by the creator property name.

(See also the -x output format switch in SWISH-RUN.)

Can Swish-e index multi-byte characters?

No. This will require much work to change. But, Swish-e works with eight-bit characters, so many characters sets can be used. Note that it does call the ANSI-C tolower() function which does depend on the current locale setting. See locale(7) for more information.

Indexing

How do I pass Swish-e a list of files to index?

Currently, there is not a configuration directive to include a file that contains a list of files to index. But, there is a directive to include another configuration file.

    IncludeConfigFile /path/to/other/config

And in /path/to/other/config you can say:

    IndexDir file1 file2 file3 file4 file5 ...
    IndexDir file20 file21 file22

You may also specify more than one configuration file on the command line:

    ./swish-e -c config_one config_two config_three

Another option is to create a directory with symbolic links of the files to index, and index just that directory.

How does Swish-e know which parser to use?

Swish can parse HTML, XML, and text documents. The parser is set by associating a file extension with a parser by the IndexContents directive. You may set the default parser with the DefaultContents directive. If a document is not assigned a parser it will default to the HTML parser (HTML2 if built with libxml2).

You may use Filters or an external program to convert documents to HTML, XML, or text.

Can I reindex and search at the same time?

Yes. Starting with version 2.2 Swish-e indexes to temporary files, and then renames the files when indexing is complete. On most systems renames are atomic. But, since Swish-e also generates more than one file during indexing there will be a very short period of time between renaming the various files when the index is out of sync.

Settings in src/config.h control some options related to temporary files, and their use during indexing.

Can I index phrases?

Phrases are indexed automatically. To search for a phrase simply place double quotes around the phrase.

For example:

    swish-e -w 'free and "fast search engine"'

How can I prevent phrases from matching across sentences?

Use the BumpPositionCounterCharacters configuration directive.

Swish-e isn't indexing a certain word or phrase.

There are a number of configuration parameters that control what Swish-e considers a "word" and it has a debugging feature to help pinpoint any indexing problems.

Configuration file directives (SWISH-CONFIG) WordCharacters, BeginCharacters, EndCharacters, IgnoreFirstChar, and IgnoreLastChar are the main settings that Swish-e uses to define a "word". See SWISH-CONFIG and SWISH-RUN for details.

Swish-e also uses compile-time defaults for many settings. These are located in src/config.h file.

Use of the command line arguments -k, -v and -T are useful when debugging these problems. Using -T INDEXED_WORDS while indexing will display each word as it is indexed. You should specify one file when using this feature since it can generate a lot of output.

     ./swish-e -c my.conf -i problem.file -T INDEXED_WORDS

You may also wish to index a single file that contains words that are or are not indexing as you expect and use -T to output debugging information about the index. A useful command might be:

    ./swish-e -f index.swish-e -T INDEX_FULL

Once you see how Swish-e is parsing and indexing your words, you can adjust the configuration settings mentioned above to control what words are indexed.

Another useful command might be:

     ./swish-e -c my.conf -i problem.file -T PARSED_WORDS INDEXED_WORDS

This will show white-spaced words parsed from the document (PARSED_WORDS), and how those words are split up into separate words for indexing (INDEXED_WORDS).

How do I keep Swish-e from indexing numbers?

Swish-e indexes words as defined by the WordCharacters setting, as described above. So to avoid indexing numbers you simply remove digits from the WordCharacters setting.

There are also some settings in src/config.h that control what "words" are indexed. You can configure swish to never index words that are all digits, vowels, or consonants, or that contain more than some consecutive number of digits, vowels, or consonants. In general, you won't need to change these settings.

Also, there's an experimental feature called IgnoreNumberChars which allows you to define a set of characters that describe a number. If a word is made up of only those characters it will not be indexed.

Swish-e crashes and burns on a certain file. What can I do?

This shouldn't happen. If it does please post to the Swish-e discussion list the details so it can be reproduced by the developers.

In the mean time, you can use a FileRules directive to exclude the particular file name, or pathname, or its title. If there are serious problems in indexing certain types of files, they may not have valid text in them (they may be binary files, for instance). You can use NoContents to exclude that type of file.

Swish-e will issue a warning if an embedded null character is found in a document. This warning will be an indication that you are trying to index binary data. If you need to index binary files try to find a program that will extract out the text (e.g. strings(1), catdoc(1), pdftotext(1)).

How to I prevent indexing of some documents?

When using the file system to index your files you can use the FileRules directive. Other than FileRules title, FileRules only works with the file system (-S fs) indexing method, not with -S prog or -S http.

If you are spidering a site you have control over, use a robots.txt file in your document root. This is a standard way to excluded files from search engines, and is fully supported by Swish-e. See http://www.robotstxt.org/

If spidering a website with the included spider.pl program then add any necessary tests to the spider's configuration file. Type <perldoc spider.pl> in the prog-bin directory for details or see the spider documentation on the Swish-e website. Look for the section on callback functions.

If using the libxml2 library for parsing HTML (which you probably are), you may also use the Meta Robots Exclusion in your documents:

    <meta name="robots" content="noindex">

See the obeyRobotsNoIndex directive.

How do I prevent indexing parts of a document?

To prevent Swish-e from indexing a common header, footer, or navigation bar, AND you are using libxml2 for parsing HTML, then you may use a fake HTML tag around the text you wish to ignore and use the IgnoreMetaTags directive. This will generate an error message if the ParserWarningLevel is set as it's invalid HTML.

IgnoreMetaTags works with XML documents (and HTML documents when using libxml2 as the parser), but not with documents parsed by the text (TXT) parser.

If you are using the libxml2 parser (HTML2 and XML2) then you can use the the following comments in your documents to prevent indexing:

       <!-- SwishCommand noindex -->
       <!-- SwishCommand index -->

and/or these may be used also:

       <!-- noindex -->
       <!-- index -->

How do I modify the path or URL of the indexed documents.

Use the ReplaceRules configuration directive to rewrite path names and URLs. If you are using -S prog input method you may set the path to any string.

How can I index data from a database?

Use the "prog" document source method of indexing. Write a program to extract out the data from your database, and format it as XML, HTML, or text. See the examples in the prog-bin directory, and the next question.

How do I index my PDF, Word, and compressed documents?

Swish-e can internally only parse HTML, XML and TXT (text) files by default, but can make use of filters that will convert other types of files such as MS Word documents, PDF, or gzipped files into one of the file types that Swish-e understands.

Please see SWISH-CONFIG and the examples in the filters and filter-bin directory for more information.

See the next question to learn about the filtering options with Swish-e.

How do I filter documents?

The term "filter" in Swish-e means the converstion of a document of one type (one that swish-e cannot index directly) into a type that Swish-e can index, namely HTML, plain text, or XML. To add to the confusion, there are a number of ways to accomplish this in Swish-e. So here's a bit of background.

The FileFilter directive was added to swish first. This feature allows you to specify a program to run for documents that match a given file extension. For example, to filter PDF files (files that end in .pdf) you can specify the configuation setting of:

    FileFilter .pdf pdftotext   "'%p' -"

which says to run the program "pdftotext" passing it the pathname of the file (%p) and a dash (which tells pdftotext to output to stdout). Then for each .pdf file Swish-e runs this program and reads in the filtered document from the output from the filter program.

This has the advantage that it is easy to setup -- a single line in the config file is all that is needed to add the filter into Swish-e. But it also has a number of problems. For example, if you use a Perl script to do your filtering it can be very slow since the filter script must be run (and thus compiled) for each processed document. This is exacerbated when using the -S http method since the -S http method also uses a Perl script that is run for every URL fetched. Also, when using -S prog method of input (reading input from a program) using FileFilter means that Swish-e must first read the file in from the external program and then write the file out to a temporary file before running the filter.

With -S prog it makes much more sense to filter the document in the program that is fetching the documents than to have swish-e read the file into memory, write it to a temporary file and then run an external program.

The Swish-e distribution contains a couple of example -S prog programs. spider.pl is a reasonably full-featured web spider that offers many more options than the -S http method. And it is much faster than running -S http, too.

The spider has a perl configuration file, which means you can add programming logic right into the configuration file without editing the spider program. One bit of logic that is provided in the spider's configuration file is a "call-back" function that allows you to filter the content. In other words, before the spider passes a fetched web document to swish for indexing the spider can call a simple subroutine in the spider's configuration file passing the document and its content type. The subroutine can then look at the content type and decide if the document needs to be filtered.

For example, when processing a document of type "application/msword" the call-back subroutine might call the doc2txt.pm perl module, and a document of type "appliation/pdf" could use the pdf2html.pm module. The prog-bin/SwishSpiderConfig.pl file shows this usage.

This system works reasonably well, but also means that more work is required to setup the filters. First, you must explicitly check for specific content types and then call the appropriate Perl module, and second, you have to know how each module must be called and how each returns the possibly modified content.

In comes SWISH::Filter.

To make things easier the SWISH::Filter Perl module was created. The idea of this module is that there is one interface used to filter all types of documents. So instead of checking for specific types of content you just pass the content type and the document to the SWISH::Filter module and it returns a new content type and document if it was filtered. The filters that do the actual work are designed with a standard interface and work like filter "plug-ins". Adding new filters means just downloading the filter to a directory and no changes are needed to the spider's configuation file. Download a filter for Postscript and next time you run indexing your Postscript files will be indexed.

Since the filters are standardized, hopefully when you have the need to filter documents of a specific type there will already be a filter ready for your use.

Now, note that the perl modules may or may not do the actual conversion of a document. For example, the PDF conversion module calls the pdfinfo and pdftotext programs. Those programs (part of the Xpfd package) must be installed separately from the filters.

The SwishSpiderConfig.pl examle spider configuration file shows how to use the SWISH::Filter module for filtering. This file is installed at $prefix/share/doc/swish-e/examples/prog-bin, where $prefix is normally /usr/local on unix-type machines.

The SWISH::Filter method of filtering can also be used with the -S http method of indexing. By default the swishspider program (the Perl helper script that fetches documents from the web) will attempt to use the SWISH::Filter module if it can be found in Perls library path. This path is set automatically for spider.pl but not for swishspider (because it would slow down a method that's already slow and spider.pl is recommended over the -S http method).

Therefore, all that's required to use this system with -S http is setting the @INC array to point to the filter directory.

For example, if the swish-e distribution was unpacked into ~/swish-e:

   PERL5LIB=~/swish-e/filters swish-e -c conf -S http

will allow the -S http method to make use of the SWISH::Filter module.

Note that if you are not using the SWISH::Filter module you may wish to edit the swishspider program and disable the use of the SWISH::Filter module using this setting:

    use constant USE_FILTERS  => 0;  # disable SWISH::Filter

This prevents the program from attempting to use the SWISH::Filter module for every non-text URL that is fetched. Of course, if you are concerned with indexing speed you should be using the -S prog method with spider.pl instead of -S http.

If you are not spidering, but you still want to make use of the SWISH::Filter module for filtering you can use the DirTree.pl program (in $prefix/lib/swish-e). This is a simple program that traverses the file system and uses SWISH::Filter for filtering.

Here's two examples of how to run a filter program, one using Swish-e's FileFilter directive, another using a prog input method program. See the SwishSpiderConfig.pl file for an example of using the SWISH::Filter module.

These filters simply use the program /bin/cat as a filter and only indexes .html files.

First, using the FileFilter method, here's the entire configuration file (swish.conf):

    IndexDir .
    IndexOnly .html
    FileFilter .html "/bin/cat"   "'%p'"

and index with the command

    swish-e -c swish.conf -v 1

Now, the same thing with using the -S prog document source input method and a Perl program called catfilter.pl. You can see that's it's much more work than using the FileFilter method above, but provides a place to do additional processing. In this example, the prog method is only slightly faster. But if you needed a perl script to run as a FileFilter then prog will be significantly faster.

    #!/usr/local/bin/perl -w
    use strict;
    use File::Find;  # for recursing a directory tree

    $/ = undef;
    find(
        { wanted => \&wanted, no_chdir => 1, },
        '.',
    );

    sub wanted {
        return if -d;
        return unless /\.html$/;

        my $mtime  = (stat)[9];

        my $child = open( FH, '-|' );
        die "Failed to fork $!" unless defined $child;
        exec '/bin/cat', $_ unless $child;

        my $content = <FH>;
        my $size = length $content;

        print <<EOF;
    Content-Length: $size
    Last-Mtime: $mtime
    Path-Name: $_

    EOF

        print <FH>;
    }

And index with the command:

    swish-e -S prog -i ./catfilter.pl -v 1

This example will probably not work under Windows due to the '-|' open. A simple piped open may work just as well:

That is, replace:

    my $child = open( FH, '-|' );
    die "Failed to fork $!" unless defined $child;
    exec '/bin/cat', $_ unless $child;

with this:

    open( FH, "/bin/cat $_ |" ) or die $!;

Perl will try to avoid running the command through the shell if meta characters are not passed to the open. See perldoc -f open for more information.

Eh, but I just want to know how to index PDF documents!

See the examples in the conf directory and the comments in the SwishSpiderConfig.pl file.

See the previous question for the details on filtering. The method you decide to use will depend on how fast you want to index, and your comfort level with using Perl modules.

Regardless of the filtering method you use you will need to install the Xpdf packages available from http://www.foolabs.com/xpdf/.

I'm using Windows and can't get Filters or the prog input method to work!

Both the -S prog input method and filters use the popen() system call to run the external program. If your external program is, for example, a perl script, you have to tell Swish-e to run perl, instead of the script. Swish-e will convert forward slashes to backslashes when running under Windows.

For example, you would need to specify the path to perl as (assuming this is where perl is on your system):

    IndexDir e:/perl/bin/perl.exe

Or run a filter like:

    FileFilter .foo e:/perl/bin/perl.exe 'myscript.pl "%p"'

It's often easier to just install Linux.

How do I index non-English words?

Swish-e indexes 8-bit characters only. This is the ISO 8859-1 Latin-1 character set, and includes many non-English letters (and symbols). As long as they are listed in WordCharacters they will be indexed.

Actually, you probably can index any 8-bit character set, as long as you don't mix character sets in the same index and don't use libxml2 for parsing (see below).

The TranslateCharacters directive (SWISH-CONFIG) can translate characters while indexing and searching. You may specify the mapping of one character to another character with the TranslateCharacters directive.

TranslateCharacters :ascii7: is a predefined set of characters that will translate eight-bit characters to ascii7 characters. Using the :ascii7: rule will, for example, translate "Ääç" to "aac". This means: searching "Çelik", "çelik" or "celik" will all match the same word.

Note: When using libxml2 for parsing, parsed documents are converted internally (within libxml2) to UTF-8. This is converted to ISO 8859-1 Latin-1 when indexing. In cases where a string can not be converted from UTF-8 to ISO 8859-1 (because it contains non 8859-1 characters), the string will be sent to Swish-e in UTF-8 encoding. This will results in some words indexed incorrectly. Setting ParserWarningLevel to 1 or more will display warnings when UTF-8 to 8859-1 conversion fails.

Can I add/remove files from an index?

Try building swish-e with the --enable-incremental option.

The rest of this FAQ applies to the default swish-e format.

Swish-e currently has no way to add or remove items from its index. But, Swish-e indexes so quickly that it's often possible to reindex the entire document set when a file needs to be added, modified or removed. If you are spidering a remote site then consider caching documents locally compressed.

Incremental additions can be handled in a couple of ways, depending on your situation. It's probably easiest to create one main index every night (or every week), and then create an index of just the new files between main indexing jobs and use the -f option to pass both indexes to Swish-e while searching.

You can merge the indexes into one index (instead of using -f), but it's not clear that this has any advantage over searching multiple indexes.

How does one create the incremental index?

One method is by using the -N switch to pass a file path to Swish-e when indexing. It will only index files that have a last modification date newer than the file supplied with the -N switch.

This option has the disadvantage that Swish-e must process every file in every directory as if they were going to be indexed (the test for -N is done last right before indexing of the file contents begin and after all other tests on the file have been completed) -- all that just to find a few new files.

Also, if you use the Swish-e index file as the file passed to -N there may be files that were added after indexing was started, but before the index file was written. This could result in a file not being added to the index.

Another option is to maintain a parallel directory tree that contains symlinks pointing to the main files. When a new file is added (or changed) to the main directory tree you create a symlink to the real file in the parallel directory tree. Then just index the symlink directory to generate the incremental index.

This option has the disadvantage that you need to have a central program that creates the new files that can also create the symlinks. But, indexing is quite fast since Swish-e only has to look at the files that need to be indexed. When you run full indexing you simply unlink (delete) all the symlinks.

Both of these methods have issues where files could end up in both indexes, or files being left out of an index. Use of file locks while indexing, and hash lookups during searches can help prevent these problems.

I run out of memory trying to index my files.

It's true that indexing can take up a lot of memory! Swish-e is extremely fast at indexing, but that comes at the cost of memory.

The best answer is install more memory.

Another option is use the -e switch. This will require less memory, but indexing will take longer as not all data will be stored in memory while indexing. How much less memory and how much more time depends on the documents you are indexing, and the hardware that you are using.

Here's an example of indexing all .html files in /usr/doc on Linux. This first example is without -e and used about 84M of memory:

    270279 unique words indexed.
    23841 files indexed.  177640166 total bytes.
    Elapsed time: 00:04:45 CPU time: 00:03:19

This is with -e, and used about 26M or memory:

    270279 unique words indexed.
    23841 files indexed.  177640166 total bytes.
    Elapsed time: 00:06:43 CPU time: 00:04:12

You can also build a number of smaller indexes and then merge together with -M. Using -e while merging will save memory.

Finally, if you do build a number of smaller indexes, you can specify more than one index when searching by using the -f switch. Sorting large results sets by a property will be slower when specifying multiple index files while searching.

"too many open files" when indexing with -e option

Some platforms report "too many open files" when using the -e economy option. The -e feature uses many temporary files (something like 377) plus the index files and this may exceed your system's limits.

Depending on your platform you may need to set "ulimit" or "unlimit".

For example, under Linux bash shell:

  $ ulimit -n 1024

Or under an old Sparc

  % unlimit openfiles

My system admin says Swish-e uses too much of the CPU!

That's a good thing! That expensive CPU is supposed to be busy.

Indexing takes a lot of work -- to make indexing fast much of the work is done in memory which reduces the amount of time Swish-e is waiting on I/O. But, there's two things you can try:

The -e option will run Swish-e in economy mode, which uses the disk to store data while indexing. This makes Swish-e run somewhat slower, but also uses less memory. Since it is writing to disk more often it will be spending more time waiting on I/O and less time in CPU. Maybe.

The other thing is to simply lower the priority of the job using the nice(1) command:

    /bin/nice -15 swish-e -c search.conf

If concerned about searching time, make sure you are using the -b and -m switches to only return a page at a time. If you know that your result sets will be large, and that you wish to return results one page at a time, and that often times many pages of the same query will be requested, you may be smart to request all the documents on the first request, and then cache the results to a temporary file. The perl module File::Cache makes this very simple to accomplish.

Spidering

How can I index documents on a web server?

If possible, use the file system method -S fs of indexing to index documents in you web area of the file system. This avoids the overhead of spidering a web server and is much faster. (-S fs is the default method if -S is not specified).

If this is impossible (the web server is not local, or documents are dynamically generated), Swish-e provides two methods of spidering. First, it includes the http method of indexing -S http. A number of special configuration directives are available that control spidering (see "Directives for the HTTP Access Method Only" in SWISH-CONFIG). A perl helper script (swishspider) is included in the src directory to assist with spidering web servers. There are example configurations for spidering in the conf directory.

As of Swish-e 2.2, there's a general purpose "prog" document source where a program can feed documents to it for indexing. A number of example programs can be found in the prog-bin directory, including a program to spider web servers. The provided spider.pl program is full-featured and is easily customized.

The advantage of the "prog" document source feature over the "http" method is that the program is only executed one time, where the swishspider.pl program used in the "http" method is executed once for every document read from the web server. The forking of Swish-e and compiling of the perl script can be quite expensive, time-wise.

The other advantage of the spider.pl program is that it's simple and efficient to add filtering (such as for PDF or MS Word docs) right into the spider.pl's configuration, and it includes features such as MD5 checks to prevent duplicate indexing, options to avoid spidering some files, or index but avoid spidering. And since it's a perl program there's no limit on the features you can add.

Why does swish report "./swishspider: not found"?

Does the file swishspider exist where the error message displays? If not, either set the configuration option SpiderDirectory to point to the directory where the swishspider program is found, or place the swishspider program in the current directory when running swish-e.

If you are running Windows, make sure "perl" is in your path. Try typing perl from a command prompt.

If you not running windows, make sure that the shebang line (the first line of the swishspider program that starts with #!) points to the correct location of perl. Typically this will be /usr/bin/perl or /usr/local/bin/perl. Also, make sure that you have execute and read permissions on swishspider.

The swishspider perl script is only used with the -S http method of indexing.

I'm using the spider.pl program to spider my web site, but some large files are not indexed.

The spider.pl program has a default limit of 5MB file size. This can be changed with the max_size parameter setting. See perldoc spider.pl for more information.

I still don't think all my web pages are being indexed.

The spider.pl program has a number of debugging switches and can be quite verbose in telling you what's happening, and why. See perldoc spider.pl for instructions.

Swish cannot follow links generated by Javascript, as they are generated by the browser and are not part of the document.

How do I spider other websites and combine it with my own (filesystem) index?

You can either merge -M two indexes into a single index, or use -f to specify more than one index while searching.

You will have better results with the -f method.

Searching

How do I limit searches to just parts of the index?

If you can identify "parts" of your index by the path name you have two options.

The first options is by indexing the document path. Add this to your configuration:

    MetaNames swishdocpath

Now you can search for words or phrases in the path name:

    swish-e -w 'foo AND swishdocpath=(sales)'

So that will only find documents with the word "foo" and where the file's path contains "sales". That might not works as well as you like, though, as both of these paths will match:

    /web/sales/products/index.html
    /web/accounting/private/sales_we_messed_up.html

This can be solved by searching with a phrase (assuming "/" is not a WordCharacter):

    swish-e -w 'foo AND swishdocpath=("/web/sales/")'
    swish-e -w 'foo AND swishdocpath=("web sales")'  (same thing)

The second option is a bit more powerful. With the ExtractPath directive you can use a regular expression to extract out a sub-set of the path and save it as a separate meta name:

    MetaNames department
    ExtractPath department regex !^/web/([^/]+).+$!$1/

Which says match a path that starts with "/web/" and extract out everything after that up to, but not including the next "/" and save it in variable $1, and then match everything from the "/" onward. Then replace the entire matches string with $1. And that gets indexed as meta name "department".

Now you can search like:

    swish-e -w 'foo AND department=sales'

and be sure that you will only match the documents in the /www/sales/* path. Note that you can map completely different areas of your file system to the same metaname:

    # flag the marketing specific pages
    ExtractPath department regex !^/web/(marketing|sales)/.+$!marketing/
    ExtractPath department regex !^/internal/marketing/.+$!marketing/

    # flag the technical departments pages
    ExtractPath department regex !^/web/(tech|bugs)/.+$!tech/

Finally, if you have something more complicated, use -S prog and write a perl program or use a filter to set a meta tag when processing each file.

How is ranking calculated?

The swishrank property value is calculated based on which Ranking Scheme (or algorithm) you have selected. In this discussion, any time the word fancy is used, you should consult the actual code for more details. It is open source, after all.

Things you can do to affect ranking:

MetaNamesRank

You may configure your index to bias certain metaname values more or less than others. See the MetaNamesRank configuration option in SWISH-CONFIG.

IgnoreTotalWordCountWhenRanking

Set to 1 (default) or 0 in your config file. See SWISH-CONFIG. NOTE: You must set this to 0 to use the IDF Ranking Scheme.

structure

Each term's position in each HTML document is given a structure value based on the context in which the word appears. The structure value is used to artificially inflate the frequency of each term in that particular document. These structural values are defined in config.h:

 #define RANK_TITLE             7
 #define RANK_HEADER            5
 #define RANK_META              3
 #define RANK_COMMENTS          1
 #define RANK_EMPHASIZED        0

For example, if the word foo appears in the title of a document, the Scheme will treat that document as if foo appeared 7 additional times.

All Schemes share the following characteristics:

AND searches

The rank value is averaged for all AND'd terms. Terms within a set of parentheses () are averaged as a single term (this is an acknowledged weakness and is on the TODO list).

OR searches

The rank value is summed and then doubled for each pair of OR'd terms. This results in higher ranks for documents that have multiple OR'd terms.

scaled rank

After a document's raw rank score is calculated, a final rank score is calculated using a fancy log() function. All the documents are then scaled against a base score of 1000. The top-ranked document will therefore always have a swishrank value of 1000.

Here is a brief overview of how the different Schemes work. The number in parentheses after the name is the value to invoke that scheme with swish-e -R or RankScheme().

Default (0)

The default ranking scheme considers the number of times a term appears in a document (frequency), the MetaNamesRank and the structure value. The rank might be summarized as:

 DocRank = Sum of ( structure + metabias )

Consider this output with the DEBUG_RANK variable set at compile time:

 Ranking Scheme: 0 
 Word entry 0 at position 6 has struct 7
 Word entry 1 at position 64 has struct 41
 Word entry 2 at position 71 has struct 9
 Word entry 3 at position 132 has struct 9
 Word entry 4 at position 154 has struct 9
 Word entry 5 at position 423 has struct 73
 Word entry 6 at position 541 has struct 73
 Word entry 7 at position 662 has struct 73
 File num: 1104.  Raw Rank: 21.  Frequency: 8 scaled rank: 30445
  Structure tally:
  struct 0x7 = count of 1 ( HEAD TITLE FILE ) x rank map of 8 = 8

  struct 0x9 = count of 3 ( BODY FILE ) x rank map of 1 = 3

  struct 0x29 = count of 1 ( HEADING BODY FILE ) x rank map of 6 = 6

  struct 0x49 = count of 3 ( EM BODY FILE ) x rank map of 1 = 3

Every word instance starts with a base score of 1. Then for each instance of your word, a running sum is taken of the structural value of that word position plus any bias you've configured. In the example above, the raw rank is 1 + 8 + 3 + 6 + 3 = 21.

Consider this line:

  struct 0x7 = count of 1 ( HEAD TITLE FILE ) x rank map of 8 = 8

That means there was one instance of our word in the title of the file. It's context was in the <head> tagset, inside the <title>. The <title> is the most specific structure, so it gets the RANK_TITLE score: 7. The base rank of 1 plus the structure score of 7 equals 8. If there had been two instances of this word in the title, then the score would have been 8 + 8 = 16.

IDF (1)

IDF is short for Inverse Document Frequency. That's fancy ranking lingo for taking into account the total frequency of a term across the entire index, in addition to the term's frequency in a single document. IDF ranking also uses the relative density of a word in a document to judge its relevancy. Words that appear more often in a doc make that doc's rank higher, and longer docs are not weighted higher than shorter docs.

The IDF Scheme might be summarized as:

  DocRank = Sum of ( density * idf * ( structure + metabias ) )

Consider this output from DEBUG_RANK:

 Ranking Scheme: 1 
 File num: 1104  Word Score: 1  Frequency: 8  Total files: 1451   
 Total word freq: 108   IDF: 2564  
 Total words: 1145877   Indexed words in this doc: 562   
 Average words: 789   Density: 1120    Word Weight: 28716   
 Word entry 0 at position 6 has struct 7
 Word entry 1 at position 64 has struct 41
 Word entry 2 at position 71 has struct 9
 Word entry 3 at position 132 has struct 9
 Word entry 4 at position 154 has struct 9
 Word entry 5 at position 423 has struct 73
 Word entry 6 at position 541 has struct 73
 Word entry 7 at position 662 has struct 73
 Rank after IDF weighting: 574321  
 scaled rank: 132609
  Structure tally:
  struct 0x7 = count of  1 ( HEAD TITLE FILE ) x rank map of 8 = 8

  struct 0x9 = count of  3 ( BODY FILE ) x rank map of 1 = 3

  struct 0x29 = count of  1 ( HEADING BODY FILE ) x rank map of 6 = 6

  struct 0x49 = count of  3 ( EM BODY FILE ) x rank map of 1 = 3

It is similar to the default Scheme, but notice how the total number of files in the index and the total word frequency (as opposed to the document frequency) are both part of the equation.

Ranking is a complicated subject. SWISH-E allows for more Ranking Schemes to be developed and experimented with, using the -R option (from the swish-e command) and the RankScheme (see the API documentation). Experiment and share your findings via the discussion list.

How can I limit searches to the title, body, or comment?

Use the -t switch.

I can't limit searches to title/body/comment.

Or, I can't search with meta names, all the names are indexed as "plain".

Check in the config.h file if #define INDEXTAGS is set to 1. If it is, change it to 0, recompile, and index again. When INDEXTAGS is 1, ALL the tags are indexed as plain text, that is you index "title", "h1", and so on, AND they loose their indexing meaning. If INDEXTAGS is set to 0, you will still index meta tags and comments, unless you have indicated otherwise in the user config file with the IndexComments directive.

Also, check for the UndefinedMetaTags setting in your configuration file.

I've tried running the included CGI script and I get a "Internal Server Error"

Debugging CGI scripts are beyond the scope of this document. Internal Server Error basically means "check the web server's log for an error message", as it can mean a bad shebang (#!) line, a missing perl module, FTP transfer error, or simply an error in the program. The CGI script swish.cgi in the example directory contains some debugging suggestions. Type perldoc swish.cgi for information.

There are also many, many CGI FAQs available on the Internet. A quick web search should offer help. As a last resort you might ask your webadmin for help...

When I try to view the swish.cgi page I see the contents of the Perl program.

Your web server is not configured to run the program as a CGI script. This problem is described in perldoc swish.cgi.

How do I make Swish-e highlight words in search results?

Short answer:

Use the supplied swish.cgi or search.cgi scripts located in the example directory.

Long answer:

Swish-e can't because it doesn't have access to the source documents when returning results, of course. But a front-end program of your creation can highlight terms. Your program can open up the source documents and then use regular expressions to replace search terms with highlighted or bolded words.

But, that will fail with all but the most simple source documents. For HTML documents, for example, you must parse the document into words and tags (and comments). A word you wish to highlight may span multiple HTML tags, or be a word in a URL and you wish to highlight the entire link text.

Perl modules such as HTML::Parser and XML::Parser make word extraction possible. Next, you need to consider that Swish-e uses settings such as WordCharacters, BeginCharacters, EndCharacters, IgnoreFirstChar, and IgnoreLast, char to define a "word". That is, you can't consider that a string of characters with white space on each side is a word.

Then things like TranslateCharacters, and HTML Entities may transform a source word into something else, as far as Swish-e is concerned. Finally, searches can be limited by metanames, so you may need to limit your highlighting to only parts of the source document. Throw phrase searches and stopwords into the equation and you can see that it's not a trivial problem to solve.

All hope is not lost, thought, as Swish-e does provide some help. Using the -H option it will return in the headers the current index (or indexes) settings for WordCharacters (and others) required to parse your source documents as it parses them during indexing, and will return a "Parsed Words:" header that will show how it parsed the query internally. If you use fuzzy indexing (word stemming, soundex, or metaphone) then you will also need to stem each word in your document before comparing with the "Parsed Words:" returned by Swish-e.

The Swish-e stemming code is available either by using the Swish-e Perl module (SWISH::API) or the C library (included with the swish-e distribution). Also on CPAN is the module Text::DoubleMetaphone. Using SWISH::API probably provides the best stemming support.

No. Filters (FileFilter or via "prog" method) are only used for building the search index database. During search requests there will be no filter calls.

I have read the FAQ but I still have questions about using Swish-e.

The Swish-e discussion list is the place to go. http://swish-e.org/. Please do not email developers directly. The list is the best place to ask questions.

Before you post please read QUESTIONS AND TROUBLESHOOTING located in the INSTALL page. You should also search the Swish-e discussion list archive which can be found on the swish-e web site.

In short, be sure to include in the following when asking for help.

Document Info

$Id$

.

POD ERRORS

Hey! The above document had some coding errors, which are explained below:

Around line 733:

Non-ASCII character seen before =encoding in '"Ääç"'. Assuming ISO8859-1