Full-Text Search Internals
A presentation at SAP Inside Track Vienna in in Vienna, Austria by Philipp Krenn
Full-Text Search Internals Philipp Krenn @xeraa
Who uses a Database?
Who uses Search?
Developer
Store
Apache Lucene Elasticsearch
Example These are <em>not</em> the droids you are looking for.
html_strip Char Filter These are not the droids you are looking for.
standard Tokenizer These are not the droids you looking for are
lowercase Token Filter these are not the droids looking for you are
stop Token Filter droids you looking
snowball Token Filter droid you look
Analyze
GET /_analyze { "analyzer": "english", "text": "These are not the droids you are looking for." }
{ } "tokens": [ { "token": "droid", "start_offset": 18, "end_offset": 24, "type": "<ALPHANUM>", "position": 4 }, { "token": "you", "start_offset": 25, "end_offset": 28, "type": "<ALPHANUM>", "position": 5 }, ... ]
GET /_analyze { "char_filter": [ "html_strip" ], "tokenizer": "standard", "filter": [ "lowercase", "stop", "snowball" ], "text": "These are <em>not</em> the droids you are looking for." }
{ } "tokens": [ { "token": "droid", "start_offset": 27, "end_offset": 33, "type": "<ALPHANUM>", "position": 4 }, { "token": "you", "start_offset": 34, "end_offset": 37, "type": "<ALPHANUM>", "position": 5 }, ... ]
Stop Words a an and are as at be but by for if in into is it no not of on or such that the their then there these they this to was will with https://github.com/apache/lucene-solr/blob/master/lucene/ analysis/common/src/java/org/apache/lucene/analysis/en/ EnglishAnalyzer.java#L44-L50
Always Use Stop Words?
To be, or not to be.
Languages Arabic, Armenian, Basque, Brazilian, Bulgarian, Catalan, CJK, Czech, Danish, Dutch, English, Finnish, French, Galician, German, Greek, Hindi, Hungarian, Indonesian, Irish, Italian, Latvian, Lithuanian, Norwegian, Persian, Portuguese, Romanian, Russian, Sorani, Spanish, Swedish, Turkish, Thai
Language Rules English: Philipp's → philipp French: l'église → eglis German: äußerst → ausserst
German GET /_analyze { "analyzer": "german", "text": "Das sind nicht die Droiden, nach denen du suchst." }
{ } "tokens": [ { "token": "droid", "start_offset": 19, "end_offset": 26, "type": "<ALPHANUM>", "position": 4 }, { "token": "den", "start_offset": 33, "end_offset": 38, "type": "<ALPHANUM>", "position": 6 }, { "token": "such", "start_offset": 42, "end_offset": 48, "type": "<ALPHANUM>", "position": 8 } ]
German with the English Analyzer da sind nicht die droiden denen du suchst nach
German Stop Words https://github.com/apache/lucene-solr/blob/master/lucene/ analysis/common/src/resources/org/apache/lucene/analysis/ snowball/german_stop.txt
Another Example Obi-Wan never told you what happened to your father.
Another Example obi wan never told you what happen your father
Another Example <b>No</b>. I am your father.
Another Example i am your father
Inverted Index am droid father happen i look never obi told wan what you your ID 1 0 1[4] 0 0 0 1[7] 0 0 0 0 0 1[5] 0 ID 2 0 0 1[9] 1[6] 0 0 1[2] 1[0] 1[3] 1[1] 1[5] 1[4] 1[8] ID 3 1[2] 0 1[4] 0 1[1] 0 0 0 0 0 0 0 1[3]
To / The Index
PUT /starwars { "settings": { "number_of_shards": 1, "analysis": { "filter": { "my_synonym_filter": { "type": "synonym", "synonyms": [ "father,dad", "droid => droid,machine" ] } },
}, } "analyzer": { "my_analyzer": { "char_filter": [ "html_strip" ], "tokenizer": "standard", "filter": [ "lowercase", "stop", "snowball", "my_synonym_filter" ] } }
} "mappings": { "_doc": { "properties": { "quote": { "type": "text", "analyzer": "my_analyzer" } } } }
Synonyms Index synonym or query time synonym_graph
GET /starwars/_mapping GET /starwars/_settings
PUT /starwars/_doc/1 { "quote": "These are <em>not</em> the droids you are looking for." } PUT /starwars/_doc/2 { "quote": "Obi-Wan never told you what happened to your father." } PUT /starwars/_doc/3 { "quote": "<b>No</b>. I am your father." }
GET /starwars/_doc/1 GET /starwars/_doc/1/_source
Search
POST /starwars/_search { "query": { "match_all": { } } }
{ "took": 1, "timed_out": false, "_shards": { "total": 5, "successful": 5, "failed": 0 }, "hits": { "total": 3, "max_score": 1, "hits": [ { "_index": "starwars", "_type": "_doc", "_id": "2", "_score": 1, "_source": { "quote": "Obi-Wan never told you what happened to your father." } }, ...
POST /starwars/_search { "query": { "match": { "quote": "droid" } } }
{ } "took": 2, "timed_out": false, "_shards": { "total": 5, "successful": 5, "failed": 0 }, "hits": { "total": 1, "max_score": 0.39556286, "hits": [ { "_index": "starwars", "_type": "_doc", "_id": "1", "_score": 0.39556286, "_source": { "quote": "These are <em>not</em> the droids you are looking for." } } ] }
POST /starwars/_search { "query": { "match": { "quote": "dad" } } }
... "hits": { "total": 2, "max_score": 0.41913947, "hits": [ { "_index": "starwars", "_type": "_doc", "_id": "3", "_score": 0.41913947, "_source": { "quote": "<b>No</b>. I am your father." } }, { "_index": "starwars", "_type": "_doc", "_id": "2", "_score": 0.39291072, "_source": { "quote": "Obi-Wan never told you what happened to your father." } } ] } }
POST /starwars/_doc/0/_explain { "query": { "match": { "quote": "dad" } } }
{ } "_index": "starwars", "_type": "_doc", "_id": "0", "matched": false
POST /starwars/_doc/1/_explain { "query": { "match": { "quote": "dad" } } }
{ } "_index": "starwars", "_type": "_doc", "_id": "1", "matched": false, "explanation": { "value": 0, "description": "no matching term", "details": [] }
POST /starwars/_doc/2/_explain { "query": { "match": { "quote": "dad" } } }
{ "_index": "starwars", "_type": "_doc", "_id": "2", "matched": true, "explanation": { ...
POST /starwars/_search { "query": { "match": { "quote": "machine" } } }
{ } "took": 2, "timed_out": false, "_shards": { "total": 1, "successful": 1, "skipped": 0, "failed": 0 }, "hits": { "total": 1, "max_score": 1.2499592, "hits": [ { "_index": "starwars", "_type": "_doc", "_id": "1", "_score": 1.2499592, "_source": { "quote": "These are <em>not</em> the droids you are looking for." } } ] }
POST /starwars/_search { "query": { "match_phrase": { "quote": "I am your father" } } }
{ } "took": 3, "timed_out": false, "_shards": { "total": 5, "successful": 5, "failed": 0 }, "hits": { "total": 1, "max_score": 1.5665855, "hits": [ { "_index": "starwars", "_type": "_doc", "_id": "3", "_score": 1.5665855, "_source": { "quote": "<b>No</b>. I am your father." } } ] }
POST /starwars/_search { "query": { "match_phrase": { "quote": { "query": "I am father", "slop": 1 } } } }
{ } "took": 16, "timed_out": false, "_shards": { "total": 5, "successful": 5, "failed": 0 }, "hits": { "total": 1, "max_score": 0.8327639, "hits": [ { "_index": "starwars", "_type": "_doc", "_id": "3", "_score": 0.8327639, "_source": { "quote": "<b>No</b>. I am your father." } } ] }
POST /starwars/_search { "query": { "match_phrase": { "quote": { "query": "I am not your father", "slop": 1 } } } }
{ } "took": 5, "timed_out": false, "_shards": { "total": 5, "successful": 5, "failed": 0 }, "hits": { "total": 1, "max_score": 1.0409548, "hits": [ { "_index": "starwars", "_type": "_doc", "_id": "3", "_score": 1.0409548, "_source": { "quote": "<b>No</b>. I am your father." } } ] }
POST /starwars/_search { "query": { "match": { "quote": { "query": "van", "fuzziness": "AUTO" } } } }
{ } "took": 14, "timed_out": false, "_shards": { "total": 5, "successful": 5, "failed": 0 }, "hits": { "total": 1, "max_score": 0.18155496, "hits": [ { "_index": "starwars", "_type": "_doc", "_id": "2", "_score": 0.18155496, "_source": { "quote": "Obi-Wan never told you what happened to your father." } } ] }
POST /starwars/_search { "query": { "match": { "quote": { "query": "ovi-van", "fuzziness": 1 } } } }
{ } "took": 109, "timed_out": false, "_shards": { "total": 5, "successful": 5, "failed": 0 }, "hits": { "total": 1, "max_score": 0.3798467, "hits": [ { "_index": "starwars", "_type": "_doc", "_id": "2", "_score": 0.3798467, "_source": { "quote": "Obi-Wan never told you what happened to your father." } } ] }
FuzzyQuery History http://blog.mikemccandless.com/2011/03/lucenes-fuzzyquery-is-100-times-faster.html Before: Brute force Now: Levenshtein Automaton
http://blog.notdot.net/2010/07/Damn-Cool-Algorithms-Levenshtein-Automata
SELECT * FROM starwars WHERE quote LIKE "?an" OR quote LIKE "V?n" OR quote LIKE "Va?"
Score
Term Frequency / Inverse Document Frequency (TF/IDF) Search one term
BM25 Default in Elasticsearch 5.0 https://speakerdeck.com/elastic/improved-text-scoring-withbm25
Term Frequency
Inverse Document Frequency
Field-Length Norm
POST /starwars/_search?explain=true { "query": { "match": { "quote": "father" } } }
... "_explanation": { "value": 0.41913947, "description": "weight(Synonym(quote:dad quote:father) in 0) [PerFieldSimilarity], result of:", "details": [ { "value": 0.41913947, "description": "score(doc=0,freq=2.0 = termFreq=2.0\n), product of:", "details": [ { "value": 0.2876821, "description": "idf(docFreq=1, docCount=1)", "details": [] }, { "value": 1.4569536, "description": "tfNorm, computed from:", "details": [ { "value": 2, "description": "termFreq=2.0", "details": [] }, ...
Score 0.41913947: i am your father 0.39291072: obi wan never told what happen your father you
Vector Space Model Search multiple terms
Search your father
Coordination Factor Reward multiple terms
Search for 3 terms 1 term: 2 terms: 3 terms:
Practical Scoring Function Putting it all together
score(q,d) = queryNorm(q) · coord(q,d) · ∑ ( tf(t in d) · idf(t)² · t.getBoost() · norm(t,d) ) (t in q)
Function Score Script, weight, random, field value, decay (geo or date)
POST /starwars/_search { "query": { "function_score": { "query": { "match": { "quote": "father" } }, "random_score": {} } } }
Compare Scores "100% perfect" vs a "50%" match
Don't do this. Seriously. Stop trying to think about your problem this way, it's not going to end well. — https://wiki.apache.org/lucene-java/ ScoresAsPercentages
GET /starwars/_analyze { "analyzer" : "my_analyzer", "text": "These are my father's machines." }
{ "tokens": [ { "token": "my", "start_offset": 10, "end_offset": 12, "type": "<ALPHANUM>", "position": 2 }, { "token": "father", "start_offset": 13, "end_offset": 21, "type": "<ALPHANUM>", "position": 3 }, { "token": "dad", "start_offset": 13, "end_offset": 21, "type": "SYNONYM", "position": 3 }, { "token": "machin", "start_offset": 22, "end_offset": 30, "type": "<ALPHANUM>", "position": 4 } ] }
PUT /starwars/_doc/4 { "quote": "These are my father's machines." }
POST /starwars/_search { "query": { "match": { "quote": "my father machine" } } }
"hits": { "total": 4, "max_score": 2.92523, "hits": [ { "_index": "starwars", "_type": "_doc", "_id": "4", "_score": 2.92523, "_source": { "quote": "These are my father's machines." } }, { "_index": "starwars", "_type": "_doc", "_id": "1", "_score": 0.8617505, "_source": { "quote": "These are <em>not</em> the droids you are looking for." } }, ...
2.92523 == 100%
DELETE /starwars/_doc/4 POST /starwars/_search { "query": { "match": { "quote": "my father machine" } } }
"hits": { "total": 3, "max_score": 1.2499592, "hits": [ { "_index": "starwars", "_type": "_doc", "_id": "1", "_score": 1.2499592, "_source": { "quote": "These are <em>not</em> the droids you are looking for." } }, ...
1.2499592 == 43% or 100%?
PUT /starwars/_doc/4 { "quote": "These droids are my father's father's machines." } POST /starwars/_search { "query": { "match": { "quote": "my father machine" } } }
"hits": { "total": 4, "max_score": 3.0068164, "hits": [ { "_index": "starwars", "_type": "_doc", "_id": "4", "_score": 3.0068164, "_source": { "quote": "These droids are my father's father's machines." } }, { "_index": "starwars", "_type": "_doc", "_id": "1", "_score": 0.89701396, "_source": { "quote": "These are <em>not</em> the droids you are looking for." } }, ...
3.0068164 == 103%?
Performance
Conclusion
Indexing Formatting Tokenize Lowercase, Stop Words, Stemming Synonyms
Scoring Term Frequency Inverse Document Frequency Field-Length Norm Vector Space Model
There is more...
Today’s applications are expected to provide powerful full-text search. But how does that work in general and how do I implement it in my application? This talk shows you how with Elasticsearch, the most widely used search engine available today.
Video
Buzz and feedback
Here’s what was said about this presentation on social media.
