Skip to content

Latest commit

 

History

History
537 lines (422 loc) · 40.9 KB

fulltext.md

File metadata and controls

537 lines (422 loc) · 40.9 KB
Raw

Full text search with Reindexer

Reindexer has builtin full text search engine. This document describes usage of full text search.

LIKE

For simple search in text can be used operator LIKE. It search strings which match a pattern. In the pattern _ means any char and % means any sequence of chars.

    In Go:
    query := db.Query("items").
        Where("field", reindexer.LIKE, "pattern")

    In SQL:
    SELECT * FROM items WHERE fields LIKE 'pattern'

    'me_t' corresponds to 'meet', 'meat', 'melt' and so on
    '%tion' corresponds to 'tion', 'condition', 'creation' and so on

Define full text index fields

Full text search is performed in fields marked with text tag:

type Item struct {
    ID          int64  `reindex:"id,,pk"`
    Description string `reindex:"description,text"`
}

Full text search is also available for multiple fields of composite index marked with text tag:

type Item struct {
    ID          int64  `reindex:"id,,pk"`
    Name        string `reindex:"name,-"`
    Description string `reindex:"description,-"`
    _ struct{}         `reindex:"name+description=text_search,text,composite`
}

In this example full text index will include fields name and description,text_search is short alias of composite index name for using in Queries.

Full text index is case insensitive. The source text is tokenized to set of words. Word is sequence of any UTF-8 letters, digits or +, - or / symbols. First symbol of word is letter or digit.

Query to full text index

Queries to full text index are constructed by usual query interface

    query := db.Query ("items").
        Match ("name+description","text query","<stemmers>")

Or equivalent query using name alias:

    query := db.Query ("items").
        Match ("text_search","text query","<stemmers>")

Queries to full text index can be combined with conditions on another fields. e.g:

    query := db.Query ("items").
        Match ("description","text query").
        WhereInt("year",reindexer.GT,2010)

Each result of query contains rank of match. Rank is integer from 0 to 255. 0 - lowest relevancy, 255 - best relevancy. The query Iterator has method Rank(), which returns rank of current result

Note: <stemmers> see more Stemming.

Text query format

The format of query is:

query := [@[+]field[^boost][,field2[^boost]]] [=][*]term1[*][~][^boost] [+|-][*]term2[*][~][@][^boost] ...

Patterns

  • * - match any symbols. For example, termina* matches words terminator and terminal. * Can be at word start or at word end, but can not be at word middle.
  • ~ - fuzzy match misspelled word by typos dictionary. For example, black~ watches words block, blck, or blask. ~. The typos dictionary is contains all words in index with 1 possible mistake. If ~ combined with *, then it means mistake match or prefix match, but not prefix with mistake. For example, turmin*~ is matches turminals, termin, but not terminal
  • ^x - boost matches term by x. default boost value is 1.

Field selection

  • @ - comma separated list of fields to search
  • * - search in all fields
  • ^x - boost matches in field by x. default boost value is 1.
  • + - when a term matches in several fields final rank for this term will be maximum rank among fields. If some fields are marked with + ranks of matches in these fields will be added with ratio specified in config as SumRanksByFieldsRatio. All not skipped rank will be sorted and summed as Rmax + K*R1 + K*K*R2 + ... + K^n * Rn where + is SumRanksByFieldsRatio. For example, value of SumRanksByFieldsRatio is K, query is @f1,+f2,f3,+f4 term, if ranks of matches in the fields are R1 < R2 < R3 < R4 then R = R4 + K*R2 (R1 and R3 are skipped as fields f1 and f3 are not marked with +), if R2 < R3 < R4 < R1 then R = R1 + K*R4 + K*K*R2 (R1 is not skipped as it is maximum).

Binary operators

  • + - next pattern must present in found document
  • - - next pattern must not present in found document

Escape character

\ character allows to include any special character (+,-,@,*,^,~) to a word to be searched: \*crisis means to search literally *crisis and not all words that end with 'crisis'.

Phrase search

The DSL's operand can be a phrase: "word1 word2 ..."[~N]. In this case word1, word2, ... are the words that make up the phrase in the given sequence. Words order is matter, i.e. sequence word2 word1 will not be found with "word1 word2" DSL.

~N - max distance (in words) between adjacent word positions. This argument is optional, default value for N is 1.

Synonyms of multiple words are not supported in the phrase.

Examples of text queris

  • termina* -genesis - find documents contains words begins with termina, exclude documents contains word genesis
  • black~ - find documents contains word black with 1 possible mistake. e.g block, blck, or blask
  • tom jerry cruz^2 - find documents contains at least one of word tom, cruz jerry. relevancy of documents, which contains tom cruz will be greater, than tom jerry
  • fox +fast - find documents contains both words: fox and fast
  • "one two" - find documents with phrase one two
  • "one two"~5 - find documents with words one and two with distance between terms < 5. Order of terms is matter. If you need to search those terms in any order, all the required permutations must be explicitly specified in DSL: "one two"~5 "two one"~5
  • @name rush - find docuemnts with word rush only in name field
  • @name^1.5,* rush - find documents with word rush, and boost 1.5 results from name field
  • =windows - find documents with exact term windows without language specific term variants (stemmers/translit/wrong kb layout)
  • one -"phrase example" - find documents containing the word one and not containing the phrase "phrase example"
  • one "phrase example" - find documents containing the word one or the phrase "phrase example" or both
  • +one +"phrase example" - find documents containing the word one and the phrase "phrase example"
  • one "phrase example"~3 - find documents containing the word one or the phrase "phrase example" (with max distance equals of 3 word between phrase and example) or both

Natural language processing

There are built in stemmers support in full text search. It enables natural language search of words with same stem. For example, query users will also match user. Stemmer is language specific, so it is neccessary to specify language of used stemmer.

All the available stemmers are in this [directory].(cpp_src/vendor/libstemmer/src_c)

Merging queries results

It is possible to merge multiple queries results and sort final result by relevancy.

    query := db.Query ("items").
        Match ("description","text query1")
    q2 := db.Query ("another_items").
        Match ("description","text query2")
    query.Merge (q2)
    iterator = query.Exec ()
    // Check the error
    if err := iterator.Error(); err != nil {
        panic(err)
    }
    // Iterate over results
    for iterator.Next() {
        // Get the next document and cast it to a pointer
        switch elem := iterator.Object().(type) {
            case Item:
                fmt.Printf ("%v,rank=%d\n",*elem,iterator.Rank())
            case AnotherItem:
                fmt.Printf ("%v,rank=%d\n",*elem,iterator.Rank())
        }
    }

Using select functions

It is possible to use select functions to process result data. For now you can use snippet, snippet_n and highlight, debug_rank. For composite indexes the result of the function will be written in to corresponding subfields. You can not put [,)\0] symbols in functions params. If the value contains special characters, it must be enclosed in single quotes.

Notice: although text indexes may be created over numeric fields, select functions can not be applied to any non-string field.

For all the functions there are two types of supported syntax with the same behavior: field.func_name(...) and field = func_name(...).

Highlight

This functions just highlights text area that was found. It has two arguments -

  • first string that will be inserted before found text area
  • second string that will be inserted after found text area

Example: word: "some text"

b.Query("items").Match("text", query).Limit(limit).Offset(offset).Functions("text.highlight(<b>,</b>)")

result: "some text"

Snippet

Snippet highlights text area and erase other text. It has six arguments - last two is default

  • first string that will be inserted before found text area
  • second string that will be inserted after found text area
  • third number of symbols that will be placed before area
  • fourth number of symbols that will be placed after area
  • fifth delimiter before snippets, default nothing
  • sixth delimiter after snippets, default space

Example: word: "some text"

b.Query("items").Match("text", query).Limit(limit).Offset(offset).Functions("text.snippet(<b>,</b>,2,0)")

result: "e text"

Snippet_n

Snippet_n highlights text area and erase other text. More flexible version of the snippet function. It has 4 position arguments and 5 named arguments. The named arguments are optional and can be passed in any order.

  • first String that will be inserted before found text area
  • second String that will be inserted after found text area
  • third Number of symbols that will be placed before area
  • fourth Number of symbols that will be placed after area
  • pre_delim Named argument. Delimiter before snippets, default nothing
  • post_delim Named argument. Delimiter after snippets, default space
  • with_area Named argument. Takes the value 0 or 1. Print the beginning and end of the fragment relative to the beginning of the document in characters (UTF-8) in the format [B,E] after pre_delim
  • left_bound Named argument. UTF-8 character string. The beginning of the fragment is the character from the string left_bound if it occurs before third
  • right_bound Named argument. UTF-8 character string. The end of the fragment is a character from the string right_bound if it occurs before fourth

String values must be enclosed into single quotes.

Parameters' names may be specified without quotes or in double quotes.

Numbers may be passed without quotes or in single quotes.

Examples: word: "some text string"

b.Query("items").Match("text", query).Limit(limit).Offset(offset).Functions("text.snippet_n('<b>','</b>',2,2,pre_delim='{',post_delim='}',with_area=1)")

result: "{[3,11]e text s}"

b.Query("items").Match("text", query).Limit(limit).Offset(offset).Functions("text.snippet_n('<b>','</b>',5,5,pre_delim='{',post_delim='}',left_bound='o',right_bound='i')")

result: "{me text str}"

Debug_rank

This function outputs additional information about ranking of the found word in the text in the key-value format. Returned format and content may vary depending on reindexer's version. Works with text-index only.

Typos algorithm

Reindexer finds misspelled words by matching terms with deleted symbols. MaxTypos in fulltext index configuration limits how many symbols could be deleted in two words (in query and searching document). In each word could be deleted up to MaxTypos / 2 with rounding to a large value.

Typos handling details

Typos handling algorithm is language independant and simply checks possible word permutations. There are a banch of parameters to tune it:

  • MaxTypos - configures possible number of typos. Available values: [0, 4]. Larger values allow to handle more permutations, however will require much more RAM and will decrease search speed. Recommended (and default) value is 2. Behavior, depending on MaxTypos value:
    • 0 - typos are disabled.
    • 1 - allows to find words with 1 missing or 1 extra symbol. For example, query world~ will match to world, word and worlds.
    • 2 - same as '1', but also allows to find words with 1 changed symbol. For example, query sward~ will match to sward, sword, ward and swards.
    • 3 - same as '2', but also allows to find words with 1 changed symbol AND 1 extra/missing symbol at the same time. For example, query sward~ will match to sward, sword, swords, ward, wards, war and swards.
    • 4 - same as '3', but also allows to find words with 2 changed symbols at the same time. For example, query sward~ will match to dword (in addition to all the results, which it would match to with MaxTypos == 3).
  • MaxTypoLen - Maximum word length for building and matching variants with typos. Larger values will also require more RAM for typos mappings.
  • TyposDetailedConfig - those configs are serving to tune tune typos algorithm more precisely.
    • MaxMissingLetters - allows to configure maximum number of symbols, which may be removed from the initial term to transform it into the result word. Posible values are: [-1, 2]. Values, larger than (MaxTypos / 2) + (MaxTypos % 2) will be reduced to this value.
    • MaxExtraLetters - allows to configure maximum number of symbols, which may be added to the initial term to transform it into the result word. Posible values are: [-1, 2]. Values, larger than (MaxTypos / 2) + (MaxTypos % 2) will be reduced to this value.
    • MaxTypoDistance - allows to configure maximum distance between removed and added symbol in case of the symbols switch (requires MaxTypos >= 2). For example, with MaxTypoDistance = -1 (means 'no limitations') and MaxTypos == 2 query dword~ will match to both sword and words (i.e. initial symbol may not only be changed, but also may be moved to any distance). And with MaxTypoDistance = 0 (default) and MaxTypos == 2 positions of the initial and result symbol must be the same. I.e. query dword~ will match to sword, but will not match to words. Possible values for MaxTypoDistance: [-1,100].
    • MaxSymbolPermutationDistance - by design this parameter goes in pair with MaxTypoDistance. It allows to relax restrictions from MaxTypoDistance to be able to handle repositioning of the same symbol. For example, with MaxSymbolPermutationDistance = 0, MaxTypoDistance = 0 and MaxTypos == 2 query wsord~ will not match to the word sword, because it requires to either switch 2 letters with each other or change 2 letters on the same positions. With MaxSymbolPermutationDistance = 1, this symbol perutation will be handled independant from MaxTypoDistance = 0 and query wsord~ will be able to recieve sword as a result.

More examples

MaxTypos = 1 - 1 symbol could be deleted. black and blaack matches if delete excess a in the second word. black and block do not match.

MaxTypos = 2 - by 1 symbol could be deleted in each word (up to 2 symbols in sum). black and blaack match if delete excess a in the second word. black and block match if delete a in the first word and o in the second word. black and blok do not match.

MaxTypos = 3 - 1 symbol in one word and 2 symbols in another one could be deleted (up to 3 symbols in sum). black and blok match if delete ac in the first word and o in the second word.

Performance and memory usage

Internally reindexer uses enhanced suffix array of unique words, and compressed reverse index of documents. Typically size of index is about 30%-80% of source text. But can vary in corner cases.

The Upsert operation does not perform actual indexing, but just stores text. There are lazy indexing is implemented. So actually, full text index is building on first Query on fulltext field. The indexing is uses several threads, so it is efficiently utilizes resources of modern multi core CPU. Therefore the indexing speed is very high. On modern hardware indexing speed is about ~50MB/sec

But on huge text size lazy indexing can seriously slow down first Query to text index. To avoid this side-effect it is possible to warmup text index: just by dummy Query after last Upsert

Configuration

Several parameters of full text search engine can be configured from application side. To setup configuration use db.AddIndex or db.UpdateIndex methods:

...
    ftconfig := reindexer.DefaultFtFastConfig()
    // Setup configuration
    ftconfig.LogLevel = reindexer.TRACE
    // Setup another parameters
    // ...
    // Create index definition
    indexDef := reindexer.IndexDef {
        Name: "description",
        JSONPaths: []string{"description"},
        IndexType: "text",
        FieldType: "string",
        Config: ftconfig,
    }
    // Add index with configuration
    return db.AddIndex ("items",indexDef)

Base config parameters

Parameter name Type Description Default value
Bm25Boost float Boost of bm25 ranking 1
Bm25Weight float Weight of bm25 rank in final rank 0: bm25 will not change final rank. 1: bm25 will affect to fin l rank in 0 - 100% range. 0.1
DistanceBoost float Boost of search query term distance in found document. 1
DistanceWeight float Weight of search query terms distance in found document in final rank 0: distance will not change final rank. 1: distance will affect to final rank in 0 - 100% range. 0.5
TermLenBoost float Boost of search query term length 1
TermLenWeight float Weight of search query term length in final rank. 0: term length will not change final rank. 1: term length will affect to final rank in 0 - 100% range 0.3
PositionBoost float Boost of search query term position 1.0
PositionWeight float Weight of search query term position in final rank. 0: term position will not change final rank. 1: term position will affect to final rank in 0 - 100% range 0.1
FullMatchBoost float Boost of full match of search phrase with doc 1.1
PartialMatchDecrease int Decrease of relevancy in case of partial match by value: partial_match_decrease * (non matched symbols) / (matched symbols) 15
MinRelevancy float Minimum rank of found documents. 0: all found documents will be returned 1: only documents with relevancy >= 100% will be returned 0.05
MaxTypos int Maximum possible typos in word. 0: typos are disabled, words with typos will not match. N: words with N possible typos will match. Check typos handling section for detailed description. 2
MaxTyposInWord int Deprecated, use MaxTypos instead of this. Cannot be used with MaxTypos. Maximum possible typos in word. 0: typos is disabled, words with typos will not match. N: words with N possible typos will match. It is not recommended to set more than 1 possible typo -It will seriously increase RAM usage, and decrease search speed -
MaxTypoLen int Maximum word length for building and matching variants with typos. 15
FtTyposDetailedConfig struct Config for more precise typos algorithm tuning
MaxRebuildSteps int Maximum steps without full rebuild of ft - more steps faster commit slower select - optimal about 15. 50
MaxStepSize int Maximum unique words to step 4000
MergeLimit int Maximum documents count which will be processed in merge query results. Increasing this value may refine ranking of queries with high frequency words, but will decrease search speed 20000
Stemmers []string List of stemmers to use. Available values: "en", "ru", "nl", "fin", "de", "da", "fr", "it", "hu", "no", "pt", "ro", "es", "sv", "tr" "en","ru"
EnableTranslit bool Enable russian translit variants processing. e.g. term "luntik" will match word "лунтик" true
EnableKbLayout bool Enable wrong keyboard layout variants processing. e.g. term "keynbr" will match word "лунтик" true
StopWords []struct List of objects of stopwords. Words from this list will be ignored when building indexes, but may be used in fulltext queries (such as 'word*', 'word~' etc) and produce non-empty search results. More...
SumRanksByFieldsRatio float Ratio of summation of ranks of match one term in several fields 0.0
LogLevel int Log level of full text search engine 0
FieldsCfg []struct Configs for certain fields. Overlaps parameters from main config. Contains parameters: FieldName, Bm25Boost, Bm25Weight, TermLenBoost, TermLenWeight, PositionBoost, PositionWeight. empty
ExtraWordSymbols string Extra symbols, which will be threated as parts of word to addition to letters and digits "+/-"
MaxAreasInDoc int Max number of highlighted areas for each field in each document (for snippet() and highlight()). '-1' means unlimited 5
MaxTotalAreasToCache int Max total number of highlighted areas in ft result, when result still remains cacheable. '-1' means unlimited -1
Optimization string Optimize the index by 'memory' or by 'cpu' "memory"
FtBaseRanking struct Relevance of the word in different forms
Bm25Config struct Document ranking function parameters More...
SplitterType string Text breakdown algorithm. Available values: 'friso' and 'fast' "fast"

Stopwords details

The list item can be either a string or a structure containing a string (the stopword) and a bool attribute (is_morpheme) indicating whether the stopword can be part of a word that can be shown in query-results. If the stopword is set as a string, then the is_morpheme attribute is false by default and following entries are equivalent:

"stop_words":[
    {
        "word": "some_word",
        "is_morpheme": false
    },
    ///...
]

,

"stop_words":[
    "some_word",
    ///...
]

Example:

If the list of stopwords looks like this:

"stop_words":[
    {
        "word": "under",
        "is_morpheme": true
    },
    ///...
]

and there are pair of documents containing this word: {"...under the roof ..."}, {"... to understand and forgive..."}. Then for the query 'under*' we will get as a result only document {"... to understand and forgive..."} and for the query 'under' we will get nothing as a result.

If the "StopWords" section is not specified in the config, then the default stopwords list will be used, and if it is explicitly specified empty, it means that there are no stopwords.

Detailed typos config

FtTyposDetailedConfig: config for more precise typos algorithm tuning.

Parameter name Type Description Default value
MaxTypoDistance int Maximum distance between symbols in initial and target words to perform substitution. Check typos handling section for detailed description. 0
MaxSymbolPermutationDistance int aximum distance between same symbols in initial and target words to perform substitution (to handle cases, when two symbolws were switched with each other). Check typos handling section for detailed description. 1
MaxMissingLetters int Maximum number of symbols, which may be removed from the initial term to transform it into the result word. Check typos handling section for detailed description. 2
MaxExtraLetters int Maximum number of symbols, which may be added to the initial term to transform it into the result word. Check typos handling section for detailed description. 2

Base ranking config

FtBaseRanking: config for the base relevancy of the word in different forms.

Parameter name Type Description Default value
FullMatch int Relevancy of full word match 100
PrefixMin int Mininum relevancy of prefix word match. 50
SuffixMin int Mininum relevancy of suffix word match. 10
Typo int Base relevancy of typo match 85
TypoPenalty int Extra penalty for each word's permutation (addition/deletion of the symbol) in typo algorithm. The minimum rank after applying penalties will be at least 1. 15
StemmerPenalty int Penalty for the variants, created by stemming. The minimum rank after applying penalties will be at least 1. 15
Kblayout int Relevancy of the match in incorrect kblayout 90
Translit int Relevancy of the match in translit 90
Synonyms int Relevancy of synonyms match 95

Text splitters

Reindexer supports two algorithms to break texts into words: fast and friso.

Default fast algorithm is based on the definition of a word in the form of a alpha (from supported Unicode subset), number and an extended character, everything else (whitespaces, special characters, unsopported Unicode subsets, etc) will be threated as a delimiters.

Reindexer supports the following unicode block codes / extra symbols:

  • Basic Latin
  • Latin-1 Supplement
  • Latin Extended-A
  • Latin Extended-B
  • Latin Extended Additional
  • IPA Extensions
  • Greek and Coptic
  • Cyrillic
  • Armenian
  • Hebrew
  • Arabic
  • Devanagari
  • Gujarati
  • Georgian
  • Hangul Jamo
  • Greek Extended
  • Enclosed Alphanumerics
  • Hiragana
  • Katakana
  • Hangul Compatibility Jamo
  • CJK Unified Ideographs
  • Hangul Jamo Extended-A
  • Hangul Syllables
  • Hangul Jamo Extended-B
  • Fullwidth Latin Forms
  • Digits: '0'-'9'
  • Extended characters: defined in the ExtraWordSymbols field of the text index config.

This algorithm is simple and provides high performance, but it can not handle texts without delimiters (for example, Chinese language does not requires whitespaces between words, so fast-splitter will not be able to index it properly).

Alternative friso-splitter is based on friso implementation of mmseg algorithm and uses dictionaries for tokenization. Currently this splitter supports only Chinese and English languages.

Basic document ranking algorithms

For basic document ranking, one of the following algorithms can be used:

  • bm25
  • bm25_rx
  • word_count

Calculation formula for bm25:

R = (log(totalDocCount / (matchedDocCount + 1)) + 1) * termCountInDoc / wordsInDoc * (k1 + 1.0) / (termCountInDoc / wordsInDoc + k1 * (1.0 - b_ + b_ * wordsInDoc / avgDocLen))

Calculation formula for bm25_rx:

R = (log(totalDocCount / (matchedDocCount + 1)) + 1) * termCountInDoc * (k1 + 1.0) / (termCountInDoc + k1 * (1.0 - b_ + b_ * wordsInDoc / avgDocLen))

Calculation formula for word_count:

R = termCountInDoc
  • totalDocCount - total number of documents
  • matchedDocCount - number of documents in which a subform of the word was found
  • termCountInDoc - number of words found in the document for the query word subform
  • wordsInDoc - number of words in document
  • k1 - free coefficient (Сoefficient that sets the saturation threshold for the frequency of the term. The higher the coefficient, the higher the threshold and the lower the saturation rate.)
  • b - free coefficient (If b is bigger, the effects of the length of the document compared to the average length are more amplified.)
Parameter name Type Description Default value
Bm25k1 float Coefficient k1 in the formula for calculating bm25 (used only for rx_bm25, bm25). 2.0
Bm25b float Coefficient b in the formula for calculating bm25 (used only for rx_bm25, bm25). 0.75
Bm25Type string Formula for calculating document relevance (rx_bm25, bm25, word_count) "rx_bm25"

Limitations and know issues

  • Results of full text search is always sorted by relevancy.
  • It is not possible to combine 2 Match in single Query. (but actually this can be done by composite indexes or by Merge functional)