Right now, I have successfully configured a basic Hibernate Search index to be able to search for full words on various fields of my JPA entity:
#Entity
#Indexed
class Talk {
#Field String title
#Field String summary
}
And my query looks something like this:
List<Talk> search(String text) {
FullTextEntityManager fullTextEntityManager = Search.getFullTextEntityManager(entityManager)
QueryBuilder queryBuilder = fullTextEntityManager.getSearchFactory().buildQueryBuilder().forEntity(Talk).get()
Query query = queryBuilder
.keyword()
.onFields("title", "summary")
.matching(text)
.createQuery()
FullTextQuery jpaQuery = fullTextEntityManager.createFullTextQuery(query, Talk)
return jpaQuery.getResultList()
}
Now I would like to fine-tune this setup so that when I search for "test" it still finds talks where title or summary contains "test" even as the prefix of another word. So talks titled "unit testing", or whose summary contains "testicle" should still appear in the search results, not just talks whose title or summary contains "test" as a full word.
I've tried to look at the documentation, but I can't figure out if I should change something to the way my entity is indexed, or whether it has something to do with the query. Note that I wanted to do something like the following, but then it's hard to search on several fields:
Query query = queryBuilder
.keyword().wildcard()
.onField("title")
.matching(text + "*")
.createQuery()
EDIT:
Based on Hardy's answer, I configured my entity like so:
#Indexed
#Entity
#AnalyzerDefs([
#AnalyzerDef(name = "ngram",
tokenizer = #TokenizerDef(factory = StandardTokenizerFactory.class),
filters = [
#TokenFilterDef(factory = LowerCaseFilterFactory.class),
#TokenFilterDef(factory = NGramFilterFactory.class,
params = [
#Parameter(name = "minGramSize",value = "3"),
#Parameter(name = "maxGramSize",value = "3")
])
])
])
class Talk {
#Field(analyzer=#Analyzer(definition="ngram")) String title
#Field(analyzer=#Analyzer(definition="ngram")) String summary
}
Thanks to that configuration, when I search for 'arti', I get Talks where title or summary contains words whose 'arti' is a subword of (artist, artisanal, etc.). Unfortunately, after those I also get Talks where title or summary contain words that contains subwords of my search term (arts, fart, etc.). There's probably some fine-tuning to eliminate those, but at least I get results sooner now, and they are in a sensible order.
There are multiple things you can do here. A lot can be done via the proper analyzing during index time.
For example, you want to apply a stemmer appropriate for your language. For English this is generally the Snowball stemmer.The idea is that during indexing all words are reduced to their stem, testing and tested to _test for example. This gets you a bit along your way.
The other thing you can look into is ngramm indexing. According to your description you want to find matching in unrelated words as well. The idea here is to index "subwords" of each words, so that they later can be found.
Regarding analyzers you want to look at the named analyzerssection of the Hibernate Search docs. The key here is the #AnalyzerDef annotation.
On the query side you can also apply some "tricks". Indeed you can use wildcard queries, however, if you are using the Hibernate Search query DSL, you cannot use a keyword query, but you need to use a wildcard query. Again, check the Hibernate Search docs.
You should use Ngram or EdgeNGram Filter for indexin as you correctly noted in your answer. But you should use different analyzer for your queries as suggested in lucene documentation (see search_analyzer):
https://www.elastic.co/guide/en/elasticsearch/guide/current/_index_time_search_as_you_type.html
This way your search query wouldn't be tokenized to ngrams and your results would be more like %text% or text% in SQL.
Unfortunately for unknown reasons Hibernate Search currently doesn't support search_analyzer specification on fields. You can only specific analyzer for indexing, which would be also used for search query analysis.
I plan to implement this functionality myself.
EDIT:
You can specify search-time analyzer (search_analyzer) like this:
List<Talk> search(String text) {
FullTextEntityManager fullTextEntityManager = Search.getFullTextEntityManager(entityManager)
EntityContext entityContext = fullTextEntityManager.getSearchFactory().buildQueryBuilder().forEntity(Talk);
entityContext.overridesForField("myField", "myNamedAnalyzerDef");
QueryBuilder queryBuilder = ec.get()
Query query = queryBuilder
.keyword()
.onFields("title", "summary")
.matching(text)
.createQuery()
FullTextQuery jpaQuery = fullTextEntityManager.createFullTextQuery(query, Talk)
return jpaQuery.getResultList()
}
I have used this technique to effectively simulate Lucene search_analyzer property.
In Lucene version 4.9 I used the EnglishAnalyzer for this. I think it is a English only implementation of the SnowballAnalyzer, but not 100% certain. I used it for both creating and searching the indexes. There is nothing special needed to use it.
Analyzer analyzer = new EnglishAnalyzer(Version.LUCENE_4_9);
IndexWriterConfig iwc = new IndexWriterConfig(Version.LUCENE_4_9, analyzer);
and
analyzer = new EnglishAnalyzer(Version.LUCENE_4_9);
parser = new StandardQueryParser(analyzer);
You can see it in action at Guided Code Search. This runs exclusively off Lucene.
Lucene can be integrated into Hibernate searches, but I haven't yet tried to do that myself. I seems like it would be powerful, but I don't know: See Apache Lucene™ Integration.
I've also read that lucene can be patched into SQL engines, but I haven't tried that either. Example: Indexing Databases with Lucene.
Related
I would like to be able to find an entity based on any part of its indexed fields, and the fields must not loose any content while indexing.
Lets say I have the following sample entity class:
#Entity
public class E {
private String f;
// ...
}
And if the value of f in one entity is "This is a nice field!", I would like to be able to find it by any of these queries:
"this"
"a"
"IC"
"!"
"This is a nice field!"
The most obvious decision is to annotate the entity this way:
#Entity
#Indexed
#AnalyzerDef(name = "a",
tokenizer = #TokenizerDef(factory = KeywordTokenizerFactory.class),
filters = #TokenFilterDef(factory = LowerCaseFilterFactory.class)
)
#Analyzer(definition = "a")
public class E {
#Field
private String f;
// ...
}
And then search the following way:
String queryString;
// ...
org.apache.lucene.search.Query query = queryBuilder
.keyword()
.wildcard()
.onField("f")
.matching("*" + queryString.toLowerCase() + "*")
.createQuery();
But it is stated in the documentation that for performance purposes, it is recommended that the query does not start with either ? or *.
So as I understand, this method is ineffective.
The other idea is to use n-grams like this:
#Entity
#Indexed
#AnalyzerDef(name = "a",
tokenizer = #TokenizerDef(factory = KeywordTokenizerFactory.class),
filters = {
#TokenFilterDef(factory = LowerCaseFilterFactory.class),
#TokenFilterDef(factory = NGramFilterFactory.class,
params = {
#Parameter(name = "minGramSize", value = "1"),
#Parameter(name = "maxGramSize", value = E.MAX_LENGTH)
})
}
)
#Analyzer(definition = "a")
public class E {
static final String MAX_LENGTH = "42";
#Field
private String f;
// ...
}
And create queries this way:
String queryString;
// ...
org.apache.lucene.search.Query query = queryBuilder
.keyword()
.onField("f")
.ignoreAnalyzer()
.matching(queryString.toLowerCase())
.createQuery();
This time no wildcard queries are used and the analyzer in the query is ignored. I'm not sure whether ignoring the analyzer is good or bad, but it works with analyzer ignored.
Other possible solution would be to use WhitespaceTokenizerFactory instead of KeywordTokenizerFactory when using n-grams, then split queryString by spaces and combine searches for each substring using MUST.
In this approach, as I understand, I will get a lot less n-grams built, if the length of the string contained in f is E.MAX_LENGTH, what must be good for performance. And I will also be able to find the previously described entity by, for example, "hi ield" query. And that would be ideal.
So what would be the best way to deal with my problem? Or are all my ideas bad?
P.S. Should one ignore analyzer in queries when using n-grams?
Other possible solution would be to use WhitespaceTokenizerFactory instead of KeywordTokenizerFactory when using n-grams, then split queryString by spaces and combine searches for each substring using MUST. In this approach, as I understand, I will get a lot less n-grams built, if the length of the string contained in f is E.MAX_LENGTH, what must be good for performance. And I will also be able to find the previously described entity by, for example, "hi ield" query. And that would be ideal.
This is more or less the ideal solution, except for one thing: you shouldn't ignore the analyzer when querying. What you should do is define another analyzer without the ngram filter, but with the tokenizer, lowercase filter, etc., and explicitly instruct Hibernate Search to use that analyzer at query time.
The other solutions are too expensive, either in I/O and CPU at query time (first solution) or in storage space (second solution). Note that this third solution may still be rather expensive in storage space, depending on the value of E.MAX_LENGTH. It's generally recommended to only have a difference of one or two between minGramSize and maxGramSize, to avoid the indexing of too many grams.
Just define another analyzer, name it something like "ngram_query", and when you need to build the query, create the query builder like this:
QueryBuilder queryBuilder = fullTextEntityManager.getSearchFactory().buildQueryBuilder().forEntity(EPCAsset.class)
.overridesForField( "f" /* name of the field */, "ngram_query" )
.get();
Then create your query as usual.
Note that, if you rely on Hibernate Search to push the index schema and analyzers to Elasticsearch, you will have to use a hack in order for the query-only analyzer to be pushed: by default only the analyzers that are actually used during indexing are pushed. See https://discourse.hibernate.org/t/cannot-find-the-overridden-analyzer-when-using-overridesforfield/1043/4
When we searching like "prd gem". It returns all results names with prd gem.
but when we search only "prd", it returns all results with prd in it like prd,prd gem, prd time etc. Why not exact search now?
Following was code in picture:
luceneQuery = queryBuilder.phrase()
.onField("productId")
.andField("productName").andField("refId")
.sentence(searchText)
.createQuery();
Exact search is working fine with the name having space in it like if i search "Prd Gem", it shows only one product with name "Prd Gem", but when i search only a word like "prd", exact search is not working, it shows all product like "prd","prd gem"
So what changes need to be done with above code to implement the same?
Thats because you are "tokenizing" the data in the lucene index.
Lucene by default will try to break the strings into tokens in order to allow and speed up such searches.
I assume you are using the latest hibernate. Can you try to annotate the "productName" field with the following:
#Field(name = "productName", index = Index.YES, analyze = Analyze.NO, norms = Norms.NO)
The "analyze = Analyze.no" part should disable this feature.
I have integrated the hibernate search 3.1.1 with my existing application with Spring 2.5 and Hibernate core 3.3.2 GA. With hibernate-search-3.1.1, I am using apache lucene 2.4.1.
The problem I am facing is when I search a single word or multiple words in order, it searches perfectly and return the result but when I search multiple words out of order with blank spaces, it does not return any result. For Example, If I have a text indexed as
"Hello great world!"
Now If I search "Hello" or "great world", it returns result successfully but if I search "world Hello", it returns no result.
What I want is to be able return result if any of the complete or partial words matches on the indexed text. My source code is as below:
FullTextEntityManager fullTextEntityManager = Search.getFullTextEntityManager(this.entityManager);
// create native Lucene query
String[] fields = new String[] { "text", "description", "standard.title", "standard.briefPurpose", "standard.name" };
MultiFieldQueryParser parser = new MultiFieldQueryParser(fields, new StandardAnalyzer());
org.apache.lucene.search.Query query = null;
try {
query = parser.parse(searchTerm);
} catch (ParseException e) {
e.printStackTrace();
}
// wrap Lucene query in a javax.persistence.Query
FullTextQuery persistenceQuery = fullTextEntityManager.createFullTextQuery(query, Requirement.class);
// execute search
#SuppressWarnings("unchecked")
List<Requirement> result = persistenceQuery.getResultList();
return result;
Please help if I need to add any thing to support what I desire.
I know its very old question. But for other this short description can be helpful.
You should use analyze = Analyze.YES attribute with #Field annotation in your pojo class. It divides the string value of that field into tokens i.e. into single words. So you can search in any sequence. But note that you have to enter whole world. For example to search 'United States of America' can be found with 'America', 'States', 'States United', 'America United', etc. But whole word will be required by Hibernate search. Only 'Uni' will not work.
EDIT:
For older version of Hibernate Search Apis:
One has to use Index.TOKENIZED instead of analyze = Analyze.YES with #Field annotation.
Have you tried to use PhraseQuery.setSlop(int)? This should allow word reordering. Check out the Javadoc for more information.
I am using Hibernate Search to search for titles of tv shows on my web app.
I can use the method fuzzy() on keyword() in order to perfom fuzzy searches on keywords, but I need to take into account the whole title, so I am using phrase() instead of keyword(). The method fuzzy() is not defined for phrase(), so I was wondering if there is an easy way to achieve fuzzy searches on phrases using Hibernate Search.
If you just need a PhraseQuery with slop (that is, extra words thrown in), then you can set the slop on a phrase query like:
queryBuilder.phrase()
.setSlop(2)
.onField("myField")
.sentance("this sentence missing something")
.createQuery();
However, I'm not aware of anything in the Hibernate APIs that supports embedding fuzzy queries in phrases, but in Lucene, you can work with the SpanQuery API to build that. SpanMultiTermQueryWrapper and SpanNearQuery, in particular, are what you would need. Something like:
FuzzyQuery query1 = new FuzzyQuery(new Term("field", "fuzy"));
FuzzyQuery query2 = new FuzzyQuery(new Term("field", "phrse"));
Query wrappedQuery1 = new SpanMultiTermQueryWrapper<FuzzyQuery>(query1);
Query wrappedQuery2 = new SpanMultiTermQueryWrapper<FuzzyQuery>(query2);
SpanQuery[] clauses = {wrappedQuery1, wrappedQuery2};
SpanNearQuery(clauses, 0, true);
I'm learning the Hibernate Search Query DSL, and I'm not sure how to construct queries using boolean arguments such as AND or OR.
For example, let's say that I want to return all person records that have a firstName value of "bill" or "bob".
Following the hibernate docs, one example uses the bool() method w/ two subqueries, such as:
QueryBuilder b = fts.getSearchFactory().buildQueryBuilder().forEntity(Person.class).get();
Query luceneQuery = b.bool()
.should(b.keyword().onField("firstName").matching("bill").createQuery())
.should(b.keyword().onField("firstName").matching("bob").createQuery())
.createQuery();
logger.debug("query 1:{}", luceneQuery.toString());
This ultimately produces the lucene query that I want, but is this the proper way to use boolean logic with hibernate search? Is "should()" the equivalent of "OR" (similarly, does "must()" correspond to "AND")?.
Also, writing a query this way feels cumbersome. For example, what if I had a collection of firstNames to match against? Is this type of query a good match for the DSL in the first place?
Yes your example is correct. The boolean operators are called should instead of OR because of the names they have in the Lucene API and documentation, and because it is more appropriate: it is not only influencing a boolean decision, but it also affects scoring of the result.
For example if you search for cars "of brand Fiat" OR "blue", the cars branded Fiat AND blue will also be returned and having an higher score than those which are blue but not Fiat.
It might feel cumbersome because it's programmatic and provides many detailed options. A simpler alternative is to use a simple string for your query and use the QueryParser to create the query. Generally the parser is useful to parse user input, the programmatic one is easier to deal with well defined fields; for example if you have the collection you mentioned it's easy to build it in a for loop.
You can also use BooleanQuery. I would prefer this beacuse You can use this in loop of a list.
org.hibernate.search.FullTextQuery hibque = null;
org.apache.lucene.search.BooleanQuery bquery = new BooleanQuery();
QueryBuilder qb = fulltextsession.getSearchFactory().buildQueryBuilder()
.forEntity(entity.getClass()).get();
for (String keyword : list) {
bquery.add(qb.keyword().wildcard().onField(entityColumn).matching(keyword)
.createQuery() , BooleanClause.Occur.SHOULD);
}
if (!filterColumn.equals("") && !filterValue.equals("")) {
bquery.add(qb.keyword().wildcard().onField(column).matching(value).createQuery()
, BooleanClause.Occur.MUST);
}
hibque = fulltextsession.createFullTextQuery(bquery, entity.getClass());
int num = hibque.getResultSize();
To answer you secondary question:
For example, what if I had a collection of firstNames to match against?
I'm not an expert, but according to (the third example from the end of) 5.1.2.1. Keyword queries in Hibernate Search Documentation, you should be able to build the query like so:
Collection<String> namesCollection = getNames(); // Contains "billy" and "bob", for example
StringBuilder names = new StringBuilder(100);
for(String name : namesCollection) {
names.append(name).append(" "); // Never mind the space at the end of the resulting string.
}
QueryBuilder b = fts.getSearchFactory().buildQueryBuilder().forEntity(Person.class).get();
Query luceneQuery = b.bool()
.should(
// Searches for multiple possible values in the same field
b.keyword().onField("firstName").matching( sb.toString() ).createQuery()
)
.must(b.keyword().onField("lastName").matching("thornton").createQuery())
.createQuery();
and, have as a result, Persons with (firstName preferably "billy" or "bob") AND (lastName = "thornton"), although I don't think it will give the good ol' Billy Bob Thornton a higher score ;-).
I was looking for the same issue and have a somewhat different issue than presented. I was looking for an actual OR junction. The should case didn't work for me, as results that didn't pass any of the two expressions, but with a lower score. I wanted to completely omit these results. You can however create an actual boolean OR expression, using a separate boolean expression for which you disable scoring:
val booleanQuery = cb.bool();
val packSizeSubQuery = cb.bool();
packSizes.stream().map(packSize -> cb.phrase()
.onField(LUCENE_FIELD_PACK_SIZES)
.sentence(packSize.name())
.createQuery())
.forEach(packSizeSubQuery::should);
booleanQuery.must(packSizeSubQuery.createQuery()).disableScoring();
fullTextEntityManager.createFullTextQuery(booleanQuery.createQuery(), Product.class)
return persistenceQuery.getResultList();