User:Cpiral/CirrusSearch

Historically MediaWiki uses and extends the Lucene search engine, and now with CirrusSearch, it wraps Lucene with ElasticSearch. CirrusSearch became available July 2013, and the Wikimedia Foundation migrated to CirrusSearch from MWSearch by late 2014 because CirrusSearch makes several key advances.
 * Faster updates to the search indexes: changes to articles are reflected in search results much more quickly.
 * Search index for templates: Template wordings and subtemplate naming are attributed at each page.
 * Better support for searching in different languages.
 * Regular expression searches.

There are seven additional search parameters, and new syntax and arguments in the other four parameters and in the search words and phrases. In order to document all this exquisite control we cover the prerequisite concepts of an "indexed search" and of "page ranking". For its renewed importance in regexp searches, we reintroduce the idea of one search creating the search domain for anther term in the same query.

If your question is not answered here, please ask on the talk page.

Full text search
A "full text search" is an "indexed search". All pages are stored in the wiki database, and all the words in them are stored in the search database, which is an index to the full text of the wiki. Each visible word is indexed to the list of pages where it is found, so a search for a word is as fast as looking up a single-record. Furthermore, for any changes in wording, the search index is updated within seconds.

There are actually many indexes on the "full text" of the wiki to facilitate the many types of searches needed. The full wikitext is indexed many times into many special-purpose indexes, each parsing the wikitext in whatever way optimizes their use. Example indexes include
 * "auxiliary" text, includes hatnotes, captions, ToC, and any wikitext classed by an HTML attribute class=searchaux.
 * "Lead-in" text is the wikitext between the top of the page and the first heading.
 * The "category" text indexes the listings at the bottom.
 * Templates are now indexed. If the transcluded words of a template change, then all the pages that transclude it are updated.  (This can take a long while depending on a job queue.) If the subtemplates used by a template change, the index is updated.
 * Document content that is stored in the File/Media namespace are now indexed. Thousands of formats are recognized.

There is support for dozens of languages currently, but inclusion of all languages is desired. You can find a list of currently supported languages at elasticsearch.org. See their documentation on contributing to submit requests or patches.

CirrusSearch will optimize your query, and run it. The resulting titles are weighted by relevance, and heavily post-processed, 20 at a time, for the search results page. For example snippets are garnered from the article, and search terms are highlighted in bold text.

Search results will often be accompanied by various preliminary reports including Did you mean (spelling correction), and, when no results would otherwise be found it will say Showing results for (query correction) search instead for (your query).

Search features also include
 * sorting navigation suggestions by the number of incoming links.
 * Starting with the tilde character ~ to disable navigation and suggestions in such a way that also preserves page ranking.
 * Smart-matching characters by normalizing (or "folding") non-keyboard characters into keyboard characters.
 * Words and phrases that match are highlighted in bold on the search results page.

Words, phrases, and modifiers
The basic search term is a word or a "phrase in quotes". Search recognizes a "word" to be A "stop word" is a word that is ignored (because it is common, or for other reasons).
 * a string of digits
 * a string of letters
 * subwords between letters/digit transitions, such as in txt2regex
 * subwords inside a compoundName using camelCase

A given search term matches against content (rendered on the page). To match against wikitext instead, use the insource search parameter. Each search parameter has its own index, and interpret its given term in its own way.

Spacing between words, phrases, parameters, and input to parameters can include generous instances of whitespace and greyspace characters. "Greyspace characters" are all the non-alphanumeric characters ~!@#$%^&*_+-={}|[]\:";'<>?,./ . A mixed string of greyspace characters and whitespace characters, is "greyspace", and is one big word boundary.  Greyspace is how indexes are made and queries are interpreted.

Two exceptions are where 1) an embedded:colon is one word (it being treated as a letter), and 2) an embedded comma, such as in 1,2,3, is treated as a number Greyspace characters are otherwise ignored unless, due to query syntax, they can be interpreted as modifier characters.

The modifiers are ~ * ? - " ! . Depending on their placement in the syntax they can apply to a term, a parameter, or to an entire query.  Word and phrase modifiers are the wildcard, proximity, and fuzzy searches.  Each parameter can have their own modifiers, but in general
 * A fuzzy-word or fuzzy-phrase search can suffix a tilde ~ character (and a number telling the degree).
 * A tilde ~ character prefixing a query guarantees search results instead of any possible navigation.
 * A wildcard character inside a word can be a question ? mark for one character or an asterisk * character for more.
 * Truth-logic can interpret AND and OR, but parameters cannot.
 * Truth-logic understands - or ! prefixed to a term to invert the usual meaning of the term from "match" to "exclude".
 * Quotes around words mark an "exact phrase" search. For parameters they are also needed to delimit multi-word input.
 * Stemming is automatic but can be turned off using an "exact phrase".

A phrase search can be initiated by various hints to the search engine. Each method of hinting has a side-effect of how tolerant the matching of the word sequence will be. For greyspace, camelCase, or txt2number hints A "search instead" report is triggered when a universally unknown word is ignored in a phrase.
 * given words-joined_by_greyspace(characters) or wordsJoinedByCamelCaseCharacters it finds words joined by ... characters, in their bare forms or greyspace forms.
 * txt2number will match  or.
 * Stop words are enabled for the edge cases (in the periphery) of a grey_space or camelCase phrase. For example the_invisible_hand_of_a will match.

Each one of the following types of phrase-matching contains and widens the match-tolerances of the previous one: A word search will "additionally" find the words anywhere on the page.
 * An "exact phrase" "in quotes" will tolerate (match) with greyspace. Given "exact_phrase" or "exact phrase" it matches.
 * A greyspace_phrase initiates stemming and stop word checks.
 * Given CamelCase it will additionally match, all lowercase.  CirrusSearch is case insensitive.

Some parameters interpret greyspace phrases, but other paramters, like insource only interpret the usual "phrase in quotes".

Note how the "exact phrase" search interpreted the embedded:colon character as a letter, but not the embedded_underscore character. A similar event occurs with the comma, character inside a number.

Given, CirrusSearch, when in an "exact phrase" context, (which includes the insource parameter context), will not match  ,  , or  , but will then only match.

Otherwise, remember that for CirrusSearch words are letters, numbers, or a combination of the two, and case does not matter.

The common word search employs the space character and is aggressive with stemming, and when the same words are joined by greyspace characters or camelCase they are aggressive with phrases and subwords.

When common words like "of" or "the" are included in a greyspace-phrase, they are ignored, so as to match more aggressively.

A greyspace_phrase search term, or a camelCase, or a txt2number term, match the signified words interchangeably. You can use any of those three forms.

Now camelcase matches camelCase because Search is not case sensitive, but camelCase matches camelcase because camelCase is more aggressive. Like the rest of Search, subword "words" are not case-sensitive. By comparison the "exact phrase" is greyspace oriented and ignores numeric or letter-case transitions, and stemming. "Quoted phrases" are not case sensitive.

From the table we can surmise that the basic search parser_function -"parser function" is the sum of the basic searches  and.

Making inquiries with numbers, we would find that
 * Plan9 or Plan_9 matches any of:,  ,  ,  ,
 * "plan9" only matches  (case insensitive)
 * Plan*9 matches  or.

The wildcard * character is basic and matches within a rendered word. It ignores the first (one or more) letters, then it can match a string of letters and digits. Or it ignores the first (one or more) digits, then it can match a sequence of zero or more numbers; capital letters, ordinal letters (st, nd, rd), or time abbreviations (am or pm); or parts of decimal numbers. For the wildcard character: Or *? is also accepted, but ?* is not recognized.
 * The comma is considered part of one number, but the decimal point is considered a greyspace character, and will delimit two numbers.
 * Inside an "exact phrase" it matches stemming plus compounding.
 * It may sometimes serve as an alternative to regex (an advanced search covered later);
 * Wildcards work for word, phrase, and insource searches, but not in the intitle parameter.
 * ? can represent one letter or number;

Putting a tilde ~ character after a word or phrase activates a fuzzy search.
 * For a phrase it is termed a proximity search, because proximal words are tolerated to an approximate rather than exact phrase.
 * For example, "exact one two phrase"~2 matches.
 * For a word it means extra characters or changed characters.
 * For an phrase a fuzzy search requires a whole number telling how many extra words to fit in, but
 * for a word a fuzzy search is a decimal fraction, defaulting to word~0.5 (word~.5), where at most two letters can be found swapped, changed, or added, but never the first two letters.
 * For a proximity phrase, a large number can be used, but that is an "expensive" (slow) search.
 * For a word word~.1 is most fuzzy, and word~.9 is least fuzzy, and word~1 is not fuzzy at all.

For the closeness value necessary to match in reverse (right to left) order, count and discard all the extra words, then add twice the the total count of remaining words minus one. (In other words, add twice the number of segments). For the full proximity algorithm, see Elastic Search. An explicit AND is required between two phrases because otherwise the two "inner" "quotation marks" are confused.

Quotes turn off stemming, "but appending"~ the tilde reactivates the stemming.

Insource
Insource could be used to find any one word rendered on a page, but it's made for finding any odd phrase you might find including MediaWiki markup, and this phrase completely ignores greyspace: insource: "state state autocollapse" matches.

Insource complements itself. On the one hand it has full text search for any word in the wikitext, instantly. On the other hand it can process a regexp search for any string of characters. Regex scan all the textual characters in a given list of pages; they don't have a word index to speed things up, and the process is interrupted if it must run more than twenty seconds. Regex run last, so to limit needless character-level scanning, you advance it a list of pages (a search domain) selected by a indexed search added to the query as a "clause", and you do this to every single regex query. . Insource can play both roles, and the best candidate for insource:/arg/ is often insource: arg, where arg is the same.

The syntax for the regexp is insource: no space, and then /regexp/. (No other parameter disallows a space. All the parameters except insource:/regexp/ generously accept space after their colon. )

Insource indexed-search and regexp-search roles are similar in many respects:
 * Both search wikitext only.
 * Neither finds things "sourced" by a transclusion.
 * Neither does stemmed, fuzzy, or proximity searches.
 * Both want the fewest results, and both work faster when accompanied by another clause.

But indexed searches all ignore greyspace; wildcards searches do not match greyspace, so regex are the only way to find an exact string of any characters, for example a sequence of two spaces. Regex are an entirely different class of search tool that make matching a literal string in a regexp exact string search, a basic, easy search. Advanced regex are an entirely different endeavor than matching a literal string. See below.

Prefix and namespace
One namespace can be specified at the beginning of a search. Two or more namespaces may be set from the search results page, Special:Search, in the Advanced pane of the search bar. Furthermore, this search domain "profile" can be set and remembered as a user preference there. Setting a namespace in the search box overrides all search bar settings or indications.

Enter a namespace name, or. Enter a colon   for the default namespace. Namespace aliases are accepted.

When the the file namespace, is involved a namespace modifier  has an effect, otherwise it is ignored.

You can now use an interwiki prefix as a namespace to search other projects.

The prefix syntax in its current form is relied upon for a great deal of functionality so it's been recreated as exactly as possible.

Note that the old rule of having to put prefix: at the end of the query still applies.

Prefix and namespace are used to set the initial search domain, but each is also a query. Like prefix, namespace can run alone, and it will return the the top twenty pages, and show the number of total pages.

Filters
Filters are required to accompany a bare regex search.

Any word or phrase is a filter because A filter returns a Y/N for every page in its given search domain. The filters If it can run as a standalone, a filter is also a query.

Intitle and incategory
Word and phrase searches match in a title and match in the category box on bottom of the page. But with these parameters you can select titles only or category only.


 * cow*
 * Find articles whose title or text contains words that start with cow
 * intitle:foo
 * Find articles whose title contains foo. Stemming is enabled for foo.
 * intitle:"fine line"
 * Find articles whose title contains fine then line. Stemming is disabled.
 * intitle:foo bar
 * Find articles whose title contains foo and whose title or text contains bar.
 * -intitle:foo bar
 * Find articles whose title does not contain foo and whose title or text contains bar.
 * intitle: foo bar
 * Syntax error, devolves into searching for articles whose title or text contains intitle:, foo, and bar.
 * incategory:Music
 * Find articles that are in Category:Music
 * incategory:"music history"
 * Find articles that are in Category:Music_history
 * incategory:"musicals" incategory:"1920"
 * Find articles that are in both Category:Musicals and Category:1920
 * incategory:Felis_silvestris_catus|Dogs
 * Find articles that are either in category:Felis_silvestris_catus or in Category:Dogs
 * -incategory:"musicals" incategory:"1920"
 * Find articles that are not in Category:Musicals but are in Category:1920

Intitle and incategory are old search parameters. Incategory no longer searches any subcategory automatically, but but you can now add multiple category pagenames manually. To get the search parameter [//wikitech.wikimedia.org/wiki/Nova_Resource:Catgraph/Deepcat deepcat] , to automatically add up to 70 subcategories onto an incategory parameter, incategory:category1|category2|...|category70 , you can add a line to your user-customized javascript.

Linksto
Linksto finds wikilinks to a given name, not links to content. The input is the canonical, case sensitive, page name. It must match the title line of the content page, exactly, before any title modifications of the letter-case. (It must match its { {FULLPAGENAME}}, e.g. .)

Linksto does not find redirects. It only finds [ [wikilinks]], not internal URL links. It does find wikilinks made by a template.

To find all wikilinks to a "Help:Cirrus Search", if "Help:Searching" and "H:S" are redirects to it:
 * 1) linksto: "Help:Cirrus Search"
 * 2) linksto: Help:Searching
 * 3) linksto: H:S

finds articles that mention "CirrusSearch" but not in a wikilink.

Hastemplate
You can specify template usage with. Input the canonical pagename to find all usage of the template, but use any of its redirect pagenames finds just that naming. Namespace aliases are accepted, capitalization is entirely ignored, and redirects are found, all in one name-search. The namespace defaults to Template. (Compare boost-template no default namespace; linksto no namespace aliases, case-sensitive, no redirects; intitle no redirects.)

Hastemplate finds secondary (or meta-template) usage on a page: it searches the post-expansion inclusion. This is the same philosophy as for words and phrases from a template, but here it for templates from a template. The page will be listed as having that content even though that content is not seen in the wikitext.


 * , finds "Template:Quality image" usage in your default search domain (namespaces).
 * , finds mainspace usage of a "Contents/TOCnavbar" template in the Portal namespace.

Page weighting
Weighting determines snippet, suggestions, and page relevance. The normal weight is one. Additional weighting is given through multipliers.

If the query is just words, pages that match them in order are given a boost. If you add any explicit phrases to your search, or for certain other additions, this "prefer phrase" feature is not applied.

Morelike
Find articles about stinging insects. Find templates about regex searching for template usage on the wiki.
 * Find articles whose text is most similar to the text of the given articles.
 * Find articles whose text is most similar to the text of the given articles.

The  query works by choosing a set of words in the input articles and run a query with the chosen words. You can tune the way it works by adding the following parameters to the search results URL: These settings can be made persistent by overriding  in Help:System message.
 * : Minimum number of documents (per shard) that need a term for it to be considered.
 * : Maximum number of documents (per shard) that have a term for it to be considered.
 * : Maximum number of terms to be considered.
 * : Minimum number of times the term appears in the input to doc to be considered. For small fields this value should be 1.
 * : Minimal length of a term to be considered. Defaults to 0.
 * : The maximum word length above which words will be ignored. Defaults to unbounded (0).
 * (comma separated list of values): These are the fields to use. Allowed fields are,  ,  ,  ,   and.
 * ( | ): use only the field data. Defaults to : the system will extract the content of the   field to build the query.
 * : The percentage of terms to match on. Defaults to 0.3 (30 percent).
 * Example:

Prefer-recent
You can give recently edited articles a boost in the search results. It goes anywhere in the query. It defaults to 160 days as recent. If you're interested in the last week, use 7 instead. All articles older than seven days are boosted half as much, and all articles older than 14 days are boosted half as much again, and so on. The boost is more than the usual multiplier, it is exponential. The factor used in the exponent is the time since the last edit.
 * prefer-recent: anywhere in the query.
 * prefer-recent:recent,boost

and a score boost of 60% of the It takes a comma-separated pair of numbers defining "recent" and the boost. The default behavior for a bar adding "prefer-recent:" to the beginning of your search. By default this will scale 60% of the score exponentially with the time since the last edit, with a half life of 160 days. This can be modified like this: prefer-recent:proportion_of_score_to_scale,half_life_in_days This number works pretty well if very small. I've tested it around .0001, which is 8.64 seconds.
 * Proportion_of_score_to_scale must be a number between 0 and 1 inclusive.
 * Half_life_in_days must be greater than 0 but allows decimal points.

This will eventually be on by default for Wikinews, but there is no reason why you can't activate it in any of your searches.

Boost-templates
You can boost pages' scores based on what templates they contain. This can be done directly in the search via  or you can set the default for all searches via the new   message. replaces the contents of  if the former is specified. The syntax is a bit funky but was chosen for simplicity. Some examples:


 * Find files in the China category sorting quality images first.
 * Find files in the China category sorting quality images first.


 * Find files in the China category sorting quality images first and low quality images last.
 * Find files in the China category sorting quality images first and low quality images last.


 * Find files about popcorn sorting quality images first and low quality images last. Remember that through the use of the  message this can be reduced to just.
 * Find files about popcorn sorting quality images first and low quality images last. Remember that through the use of the  message this can be reduced to just.

Don't try to add decimal points to the percentages. They don't work and search scoring is such that they are unlikely to matter much.

A word of warning about : if you add really really big or small percentages they can poison the full text scoring. Think, for example, if enwiki boosted featured articles by a million percent. Then searches for terms mentioned in featured articles would find the featured articles before exact title matches of the terms. Phrase matching would be similarly blown away so a search like  would find a featured article with those words scattered throughout it instead of the article for Brave New World.

Sorry for the inconsistent  in the name. Sorry again but the quotes are required on this one. Sorry also for the funky syntax. Sorry we don't try to emulate the template transclusion syntax like we do with.

Regular expression searches
A basic indexed-search finds words rendered visible on a page. Hyphenation and punctuation marks and bracketing, slash and other math and computing symbols, are merely boundaries for the words. It is not possible to include them in an indexed search.

because its possible that this could block other regexp searches. Other regex users are probably blocked when you query takes too long If ever a regexp search takes more than These return much much faster when you limit the regexp search-domain to the results of one or more index-based searches.

Warning: Do not run a bare insource:/regexp/ search. It will probably timeout after 20 seconds anyway, while blocking responsible users.

An "exact string" regexp search is a basic search; it will simply "quote" the entire regexp, or "backslash-escape" all non-alphanumeric characters in the string. All regexp searches also require that the user develop a simple filter to generate the search domain for the regex engine to search:

The last example works from a link on a page, but { {FULLPAGENAME}} doesn't function in the search box. For example
 * ' [[Special:Search/insource:/regex/ prefix:| finds the term regex'' on this page ]].

Any search with no namespace specified (or prefix specified) searches your default search domain, settable on any search-results page, i.e. settable at Special:Search. The default search domain is commonly reset by power users to All namespaces, i.e. the entire wiki, but if this occurs for a bare regexp search, then on a large wiki it will probably incur an HTML timeout before completing the search.

A regex search actually scours each page in the search domain character-by character. By contrast, an indexed search actually queries a few records from a database separately maintained from the wiki database, and provides nearly instant results. So when using using an insource:// (a regexp of any kind), consider creating one the other search terms that will limit the regex search domain as much as possible. There are many search terms that use an index and so instantly provide a more refined search domain for the /regexp/. In order of general effectiveness:


 * insource:"" with quotation marks, duplicating the regexp except without the slashes or escape characters, is ideal.
 * intitle, incategory, and linksto are excellent filters.
 * hastemplate: is a very good filter.
 * "word1 word2 word3", with or without the quotation marks, are good.
 * namespace: is practically useless, but may enable a slow regexp search to complete.

The prefix operator is especially useful with a { {FULLPAGENAME}} in a search template, a search link, or an input box, because it automatically searches any subdirectories. To develop a new regexp, or refine a complex regexp, use  on a page with a sample of the target data.

Search terms that do not increase the efficiency of a regexp search are the page-scoring operators: morelike, boost-template, and prefer-recent.

Metacharacters
This section covers how to escape the metacharacters used in rexexp searches. For the actual meaning of the metacharacters see the explanation of the syntax.

The use of an exact string requires a regexp, but the regexp term obligates the search to limit itself. Add a regexp term, never search a bare regexp. Start by noting the number of pages in a previous search before committing an exact string search. Querying with an exact string requires a filtered search domain.

For example, Refining with an exact string. because it adds it as a single regexp term while refining a search, the limited number of pages the regexp must crawl is can be seen.
 * to search a namespace, gague the number of pages with a single term that is a namespace. This will list the number of pages in that namespace.
 * starting out to find again what you may have seen, like "wiki-link" or "(trans[in]clusion)" start with namespace and insource filters.
 * refinining an ongoing search process with what you want to see, like "2 + 2 = 4", or "site.org" This is ideally the best use of regex,

You can start out intending an exact string search, but keep in mind
 * regex only search the wikitext not the rendered text, so there are some differences around the markup, and even the number of space characters must match precisely.
 * You are obligated to supply an accompanying filter.
 * You must learn how to escape regex metacharacters.

There are two ways to escape metacharacters, and they are both useful at times, and sometimes concatenated side-by-side in the escaping of a string.
 * Backslash-escape one of them \char. The insource:/regexp/ uses slashes to delimit the regexp. Giving /reg/exp/ is ambiguous, so you must write /reg\/exp/.
 * Put a string of them in double quotes "string". Because escaping a character can't hurt, you can escape any character along with any possible metacharacters in there. Escaping with quotes is cleaner.
 * You can't mix methods, but you can concatenate them.

Double-quotes escaping using insource:/"regexp"/ is an easy way to search for many kinds of strings, but you can't backslash-escape anything inside a double-quoted escape.
 * instead of
 * is as good as
 * But  always.
 * And .  It finds the   literally, which is not the   you probably wanted.

Backslash-escape using insource:/regexp/ allows escaping the " and / delimiters, but requires taking into account metacharacters, and escaping any:
 * To match a  delimiter character use.
 * To match a  delimiter character use.
 * The metacharacters would be.
 * The equivalent expression is.

The simplest algorithm to create the basic string-finding expression using insource:/"regexp"/, need not take metacharacters into account except for the " and / characters:
 * 1) Write   out. (The /" delimiters "/ are not shown.)
 * 2) Replace   with   (previous double-quote: stop, concatenate, quote restart).
 * 3) Replace   with   (stop, concatenate, start).
 * 4) You get , showing concatenation of the two methods.

The square-bracket notation for creating your own character-class also escapes its metacharacters. To target a literal right square bracket in your character-class pattern, it must be backslash escaped, otherwise it can be interpreted as the closing delimiter of the character-class pattern definition. The first position of a character class will also escape the right square bracket. Inside the delimiting square brackets of a character class, the dash character also has special meaning (range) but can it too can be included literally in the class the same way as the right square bracket can. For example both of these patterns target character that is either a dash or a right square bracket or a dot:  or.

For general examples using metacharacters
 * matches "2 + 2 = 4", with zero spaces between the characters.
 * match with zero or one space in between. The equals = sign is not a metacharacter, but the plus + sign is.

There are some notable differences from standard regex metacharacters:
 * The dot . metacharacter stands for any character including a newline, so .* matches across lines.
 * The number # sign means something, and must be escaped.
 * The ^ and $ are not needed. Like "grep" (global per line, regular expression, print each line), each insource:// is a "global per document, regular expression, search-results-list each document" per document.
 * support a multi-digit numeric range like [0-9] does, but without regard to the number of character positions, or the range in each position, so <9-10> works, and even <1-111> works.

Advanced example
For example, using metacharacters to find the usage of a template called Val having, inside the template call, an unnamed parameter containing a possibly signed, three to four digit number, possibly surrounded by space characters, AND on the same page, inside a template Val call, a named argument "fmt=commas" having any allowable spaces around it, (it could be the same template call, or a separate one):



It is fast because it uses two filters so that every page the regexp crawls has the highest possible potential. Assuming your search domain is set to ALL, it searches the entire wiki, because it offers no namespace or prefix.