OrientDB使用Lucene的全文索引,并且使用中文分词器

2019-01-20-19:07

文下有官方文档的备份当前数据库版本为3.0.x

# 创建表

CREATE CLASS Item;

# 创建字段

CREATE PROPERTY Item.text STRING;

#  创建索引（默认英文索引

CREATE INDEX Item.text ON Item(text) FULLTEXT ENGINE LUCENE;

#  插入数据

INSERT INTO Item (text) VALUES ('My sister is coming for the holidays.');



#  创建中文分词器

CREATE INDEX Item.text ON Item(text)
            FULLTEXT ENGINE LUCENE METADATA {
                "analyzer": "org.apache.lucene.analysis.cn.smart.SmartChineseAnalyzer"
            }


#  查询


SELECT FROM Item WHERE SEARCH_CLASS("sister") = true

# 搜索包含两个词的 

SELECT FROM Item WHERE SEARCH_CLASS("+sister +coming") = true

#  包含sister但是不包含 coming

SELECT FROM Item WHERE SEARCH_CLASS("+sister -coming") = true


# 通配符

SELECT FROM Item WHERE SEARCH_CLASS('meet*') = true

#  高亮


SELECT FROM City WHERE SEARCH_CLASS("+name:cas*  +description:beautiful", {
    "allowLeadingWildcard": true ,
    "lowercaseExpandedTerms": false,
    "boost": {
        "name": 2
    },
    "highlight": {
        "fields": ["name"],
        "start": "",
        "end": ""
    }
}) = true

SELECT name, $name_hl, description, $description_hl FROM City
WHERE SEARCH_CLASS("+name:cas*  +description:beautiful", {
    "highlight": {
        "fields": ["name", "description"],
        "start": "",
        "end": ""
    }
}) = true


#  指定索引搜索

SELECT FROM City WHERE SEARCH_INDEX("City.name", "cas*") = true

search: keywords: ['index', 'FULLTEXT', 'full text', 'Lucene']

Lucene FullText Index

In addition to the standard FullText Index, which uses the SB-Tree index algorithm, you can also create FullText indexes using the Lucene Engine . Apache LuceneTM is a high-performance, full-featured text search engine library written entirely in Java. Check the Lucene documentation for a full overview of its capabilities.

How Lucene's works?

Let's look at a sample corpus of five documents: My sister is coming for the holidays. The holidays are a chance for family meeting. Who did your sister meet? It takes an hour to make fudge. * My sister makes awesome fudge.

What does Lucene do? Lucene is a full text search library. Search has two principal stages: indexing and retrieval.

During indexing, each document is broken into words, and the list of documents containing each word is stored in a list called the postings list. The posting list for the word my is:

my --> 1,5

Posting list for others terms:

fudge --> 4,5

sister --> 1,2,3,5

fudge --> 4,5

The index consists of all the posting lists for the words in the corpus. Indexing must be done before retrieval, and we can only retrieve documents that were indexed.

Retrieval is the process starting with a query and ending with a ranked list of documents. Say the query is "my fudge". In order to find matches for the query, we break it into the individual words, and go to the posting lists. The full list of documents containing the keywords is [1,4,5]. Note that the query is broken into words (terms) and each term is matched with the terms in the index. Lucene's default operator is OR, so it retrieves the documents tha contain my OR fudge. If we want to retrieve documents that contain both my and fudge, rewrite the query: "+my +fudge".

Lucene doesn't work as a LIKE operator on steroids, it works on single terms. Terms are produced analyzing the provided text, so the right analyzer should be configured. On the other side, it offers a complete query language, well documented here:

Index creation

To create an index based on Lucene

CREATE INDEX  ON  (prop-names) FULLTEXT ENGINE LUCENE [{json metadata}]

The following SQL statement will create a FullText index on the property name for the class City, using the Lucene Engine.

CREATE INDEX City.name ON City(name) FULLTEXT ENGINE LUCENE

Indexes can also be created on n-properties. For example, create an index on the properties name and description on the class City.

CREATE INDEX City.name_description ON City(name, description)
          FULLTEXT ENGINE LUCENE

When multiple properties should be indexed, define a single multi-field index over the class. A single multi-field index needs less resources, such as file handlers. Moreover, it is easy to write better Lucene queries. The default analyzer used by OrientDB when a Lucene index is created is the StandardAnalyzer. The StandardAnalyzer usually works fine with western languages, but Lucene offers analyzer for different languages and use cases.

Two minutes tutorial

Open studio or console and create a sample dataset:

CREATE CLASS Item;
CREATE PROPERTY Item.text STRING;
CREATE INDEX Item.text ON Item(text) FULLTEXT ENGINE LUCENE;
INSERT INTO Item (text) VALUES ('My sister is coming for the holidays.');
INSERT INTO Item (text) VALUES ('The holidays are a chance for family meeting.');
INSERT INTO Item (text) VALUES ('Who did your sister meet?');
INSERT INTO Item (text) VALUES ('It takes an hour to make fudge.');
INSERT INTO Item (text) VALUES ('My sister makes awesome fudge.');

Search all documents that contain sister:

SELECT FROM Item WHERE SEARCH_CLASS("sister") = true

Search all documents that contain sister AND coming:

SELECT FROM Item WHERE SEARCH_CLASS("+sister +coming") = true

Search all documents that contain sister but NOT coming:

SELECT FROM Item WHERE SEARCH_CLASS("+sister -coming") = true

Search all documents that contain the phrase sister meet:

SELECT FROM Item WHERE SEARCH_CLASS(' "sister meet" ') = true

Search all documents that contain terms starting with meet:

SELECT FROM Item WHERE SEARCH_CLASS('meet*') = true

To better understand how the query parser work, read carefully the official documentation and play with the above documents.

Customize Analyzers

In addition to the StandardAnalyzer, full text indexes can be configured to use different analyzer by the METADATA operator through CREATE INDEX.

Configure the index on City.name to use the EnglishAnalyzer:

CREATE INDEX City.name ON City(name)
            FULLTEXT ENGINE LUCENE METADATA {
                "analyzer": "org.apache.lucene.analysis.en.EnglishAnalyzer"
            }

Configure the index on City.name to use different analyzers for indexing and querying.

CREATE INDEX City.name ON City(name)
            FULLTEXT ENGINE LUCENE METADATA {
                "index": "org.apache.lucene.analysis.en.EnglishAnalyzer",
                "query": "org.apache.lucene.analysis.standard.StandardAnalyzer"
          }

EnglishAnalyzer will be used to analyze text while indexing and the StandardAnalyzer will be used to analyze query text.

A very detailed configuration, on multi-field index configuration, could be:

CREATE INDEX Song.fulltext ON Song(name, lyrics, title, author, description)
            FULLTEXT ENGINE LUCENE METADATA {
                "default": "org.apache.lucene.analysis.standard.StandardAnalyzer",
                "index": "org.apache.lucene.analysis.core.KeywordAnalyzer",
                "query": "org.apache.lucene.analysis.standard.StandardAnalyzer",
                "name_index": "org.apache.lucene.analysis.standard.StandardAnalyzer",
                "name_query": "org.apache.lucene.analysis.core.KeywordAnalyzer",
                "lyrics_index": "org.apache.lucene.analysis.en.EnglishAnalyzer",
                "title_index": "org.apache.lucene.analysis.en.EnglishAnalyzer",
                "title_query": "org.apache.lucene.analysis.en.EnglishAnalyzer",
                "author_query": "org.apache.lucene.analysis.core.KeywordAnalyzer",
                "description_index": "org.apache.lucene.analysis.standard.StandardAnalyzer",
                "description_index_stopwords": [
                  "the",
                  "is"
                ]
            }

With this configuration, the underlying Lucene index will works in different way on each field:

name: indexed with StandardAnalyzer, searched with KeywordAnalyzer (it's a strange choice, but possible)
lyrics: indexed with EnglishAnalyzer, searched with default query analyzer StandardAnalyzer
title: indexed and searched with EnglishAnalyzer
author: indexed and searched with KeywordAnalyzer
description: indexed with StandardAnalyzer with a given set of stop-words that overrides the internal set

Analysis is the foundation of Lucene. By default the StandardAnalyzer removes english stop-words and punctuation and lowercase the generated terms:

The holidays are a chance for family meeting!

Would produce

holidays
are
chance
for
family
meeting

Each analyzer has its set of stop-words and tokenize the text in a different way. Read the full (documentation)[http://lucene.apache.org/core/6_6_0/].

Query parser

It is possible to configure some behavior of the Lucene query parser Query parser's behavior can be configured at index creation time and overridden at runtime.

Allow Leading Wildcard

Lucene by default doesn't support leading wildcard: Lucene wildcard support

It is possible to override this behavior with a dedicated flag on meta-data:

{
  "allowLeadingWildcard": true
}

CREATE INDEX City.name ON City(name)
            FULLTEXT ENGINE LUCENE METADATA {
                "allowLeadingWildcard": true
            }

Use this flag carefully, as stated in the Lucene FAQ:

Note that this can be an expensive operation: it requires scanning the list of tokens in the index in its entirety to look for those that match the pattern.

Disable lower case on terms

Lucene's QueryParser applies a lower case filter on expanded queries by default. It is possible to override this behavior with a dedicated flag on meta-data:

{
  "lowercaseExpandedTerms": false
}

It is useful when used in pair with keyword analyzer:

CREATE INDEX City.name ON City(name)
            FULLTEXT ENGINE LUCENE METADATA {
              "lowercaseExpandedTerms": false,
              "default" : "org.apache.lucene.analysis.core.KeywordAnalyzer"
            }

With lowercaseExpandedTerms set to false, these two queries will return different results:

SELECT from Person WHERE SEARCH_CLASS("NAME") = true

SELECT from Person WHERE WHERE SEARCH_CLASS("name") = true

Querying Lucene FullText Indexes

OrientDB 3.0.x introduced search functions: SEARCH_CLASS, SEARCH_FIELDS, SEARCH_INDEX, SEARCH_MORE Every function accepts as last, optional, parameter a JSON with additional configuration.

SEARCH_CLASS

The best way to use the search capabilities of OrientDB is to define a single multi-fields index and use the SEARCH_CLASS function. In case more than one full-text index are defined over a class, an error is raised in case of SEARCH_CLASSI invocation.

Suppose to have this index

CREATE INDEX City.fulltex ON City(name, description) FULLTEXT ENGINE LUCENE

A query that retrieve cities with the name starting with cas and description containing the word beautiful:

SELECT FROM City WHERE SEARCH_CLASS("+name:cas*  +description:beautiful") = true

The function accepts metadata JSON as second parameter:

SELECT FROM City WHERE SEARCH_CLASS("+name:cas*  +description:beautiful", {
    "allowLeadingWildcard": true ,
    "lowercaseExpandedTerms": false,
    "boost": {
        "name": 2
    },
    "highlight": {
        "fields": ["name"],
        "start": "",
        "end": ""
    }
}) = true

The query shows query parser's configuration overrides, boost of field name with highlight. Highlight and boost will be explained later.

SEARCH_MORE

OrientDB exposes the Lucene's more like this capability with a dedicated function.

The first parameter is the array of RID of elements to be used to calculate similarity, the second parameter the usual metadata JSON used to tune the query behaviour.

SELECT FROM City WHERE SEARCH_MORE([#25:2, #25:3],{'minTermFreq':1, 'minDocFreq':1} ) = true

It is possible to use a query to gather RID of documents to be used to calculate similarity:

SELECT FROM City
    let $a=(SELECT @rid FROM City WHERE name = 'Rome')
    WHERE SEARCH_MORE( $a, { 'minTermFreq':1, 'minDocFreq':1} ) = true

Lucene's MLT has a lot of parameter, and all these are exposed through the metadata JSON: http://lucene.apache.org/core/6_6_0/queries/org/apache/lucene/queries/mlt/MoreLikeThis.html

fieldNames: array of field's names to be used to extract content
maxQueryTerms
minDocFreq
maxDocFreq
minTermFreq
boost
boostFactor
maxWordLen
minWordLen
maxNumTokensParsed
stopWords

Query parser's runtime configuration

It is possible to override the query parser's configuration given at creation index time at runtime passing a json:

SELECT from Person WHERE SEARCH_CLASS("bob",{
        "allowLeadingWildcard": true ,
        "lowercaseExpandedTerms": false
    } ) = true

The same can be done for query analyzer, overriding the configuration given at index creation's time:

SELECT from Person WHERE SEARCH_CLASS("bob",{
        "customAnalysis": true ,
        "query": "org.apache.lucene.analysis.standard.StandardAnalyzer",
        "name_query": "org.apache.lucene.analysis.en.EnglishAnalyzer"
    } ) = true

The customAnalysis flag is mandatory to enable the runtime configuration of query analyzers. The runtime configuration is per query and it isn't stored nor reused for a subsequent query. The custom configuration can be used with all the functions.

SEARCH_INDEX

The SEARCH_INDEX function allows to execute the query on a single index. It is useful if more than one index are defined over a class.

SELECT FROM City WHERE SEARCH_INDEX("City.name", "cas*") = true

The function accepts a JSON as third parameter, as for SEARCH_CLASS.

SEARCH_FIELDS

The SEARCH_FIELDS function allows to execute query over the index that is defined over one ormore fields:

SELECT FROM City WHERE SEARCH_FIELDS(["name", "description"], "name:cas* description:beautiful") = true

The function accepts a JSON as third parameter, as for SEARCH_CLASS.

Numeric and date range queries

If the index is defined over a numeric field (INTEGER, LONG, DOUBLE) or a date field (DATE, DATETIME), the engine supports range queries Suppose to have a City class witha multi-field Lucene index defined:

CREATE CLASS CITY EXTENDS V
CREATE PROPERTY CITY.name STRING
CREATE PROPERTY CITY.size INTEGER
CREATE INDEX City.name ON City(name,size) FULLTEXT ENGINE LUCENE

Then query using ranges:

SELECT FROM City WHERE SEARCH_CLASS('name:cas* AND size:[15000 TO 20000]') = true

Ranges can be applied to DATE/DATETIME field as well. Create a Lucene index over a property:

CREATE CLASS Article EXTENDS V
CREATE PROPERTY Article.createdAt DATETIME
CREATE INDEX Article.createdAt  ON Article(createdAt) FULLTEXT ENGINE LUCENE

Then query to retrieve articles published only in a given time range:

SELECT FROM Article WHERE SEARCH_CLASS('[201612221000 TO 201612221100]') =true

Retrieve the Score

When the lucene index is used in a query, the results set carries a context variable for each record representing the score. To display the score add $score in projections.

SELECT *,$score FROM V WHERE name LUCENE "test*"

Highlighting

OrientDB uses the Lucene's highlighter. Highlighting can be configured using the metadata JSON. The highlighted content of a field is returned in a dedicated field suffixed with _hl:

SELECT name, $name_hl, description, $description_hl FROM City
WHERE SEARCH_CLASS("+name:cas*  +description:beautiful", {
    "highlight": {
        "fields": ["name", "description"],
        "start": "",
        "end": ""
    }
}) = true

Parameters fields: array of field names to be highlighted start: start delimiter for highlighted text (default \) end: end delimiter for highlighted text (default \) maxNumFragments: maximum number of text's fragments to highlight

Sorting

Documents retrieved by a search call are ordered by their score. It is possible to configure the way the document are sorted. Read carefully the official documentation about sorting : https://lucene.apache.org/core/6_6_1/core/org/apache/lucene/search/Sort.html

SELECT name, description, size FROM City
WHERE SEARCH_CLASS("+name:cas*  +description:beautiful", {
    "sort": [ { 'field': 'size', reverse:true, type:'INT' }]
}) = true

Sort over multiple fields is possible:

SELECT name, description, size FROM City
WHERE SEARCH_CLASS("+name:cas*  +description:beautiful", {
    "sort": [
        { 'field': 'size', reverse:true, type:'INT' },
        { 'field': 'name', reverse:false, type:'STRING' },
        { reverse:false, type:'DOC' },
        ]
}) = true

Sort configuraton: field: is the field name. Could be absent only if the sort type is DOC or INDEX reverse: if set to true, will sort for the given field in reverse order type*: look to https://lucene.apache.org/core/6_6_1/core/org/apache/lucene/search/SortField.Type.html

CUSTOM type is not supported

Cross class search (Enterprise Edition)

Bundled with the enterprise edition there's the SEARH_CROSS function that is able to search over all the Lucene indexes defined on a database

Suppose to define two indexes:

CREATE INDEX Song.title ON Song (title,author) FULLTEXT ENGINE LUCENE METADATA
CREATE INDEX Author.name on Author(name,score) FULLTEXT ENGINE LUCENE METADATA

Searching for a term on each class implies a lot of different queries to be aggregated.

The SEARCH_CLASS function automatically performs the given query on each full-text index configured inside the database.

SELECT  EXPAND(SEARCH_CROSS('beautiful'))

The query will be execute over all the indexes configured on each field. It is possible to search over a given field of a certain class, just qualify the field names with their class name:

SELECT  EXPAND(SEARCH_CROSS('Song.title:beautiful  Author.name:bob'))

Another way is to use the metadata field _CLASS present in every index:

SELECT expand(SEARCH_CROSS('(+_CLASS:Song +title:beautiful) (+_CLASS:Author +name:bob)') )

All the options of a Lucene's query are allowed: inline boosting, phrase queries, proximity etc.

The function accepts a metadata JSON as second parameter:

SELECT  EXPAND(SEARCH_CROSS('Author.name:bob Song.title:*tain', {"
   "allowLeadingWildcard" : true,
   "boost": {
        "Author.name": 2.0
        }
   }
)

Highlight isn't supported yet.

Lucene Writer fine tuning (expert)

It is possible to fine tune the behaviour of the underlying Lucene's IndexWriter

CREATE INDEX City.name ON City(name)
    FULLTEXT ENGINE LUCENE METADATA {
        "directory_type": "nio",
        "use_compound_file": false,
        "ram_buffer_MB": "16",
        "max_buffered_docs": "-1",
        "max_buffered_delete_terms": "-1",
        "ram_per_thread_MB": "1024",
        "default": "org.apache.lucene.analysis.standard.StandardAnalyzer"
    }

directory_type: configure the acces type to the Lucene's index
- nio (default): the index is opened with NIOFSDirectory
- mmap: the index is opened with MMapDirectory
- ram: index will be created in memory with RAMDirectory
use_compound_file: default is false
ram_buffer_MB: size of the document's buffer in MB, default value is 16 MB (which means flush when buffered docs consume approximately 16 MB RAM)
max_buffered_docs: size of the document's buffer in number of docs, disabled by default (because IndexWriter flushes by RAM usage by default)
max_buffered_delete_terms: disabled by default (because IndexWriter flushes by RAM usage by default).
ram_per_thread_MB: default value is 1945

For a detailed explanation of config parameters and IndexWriter behaviour

indexWriterConfig : https://lucene.apache.org/core/6_6_0/core/org/apache/lucene/index/IndexWriterConfig.html
indexWriter: https://lucene.apache.org/core/6_6_0/core/org/apache/lucene/index/IndexWriter.html

Index lifecycle

Lucene indexes are lazy. If the index is in idle mode, no reads and no writes, it will be closed. Intervals are fully configurable.

flushIndexInterval: flushing index interval in milliseconds, default to 20000 (10s)
closeAfterInterval: closing index interval in milliseconds, default to 120000 (12m)
firstFlushAfter: first flush time in milliseconds, default to 10000 (10s)

To configure the index lifecycle, just pass the parameters in the JSON of metadata:

CREATE INDEX City.name ON City(name) FULLTEXT ENGINE LUCENE METADATA
{
  "flushIndexInterval": 200000,
  "closeAfterInterval": 200000,
  "firstFlushAfter": 20000
}

Create index using the Java API

The FullText Index with the Lucene Engine is configurable through the Java API.

OSchema schema = databaseDocumentTx.getMetadata().getSchema();
    OClass oClass = schema.createClass("Foo");
    oClass.createProperty("name", OType.STRING);
    oClass.createIndex("City.name", "FULLTEXT", null, null, "LUCENE", new String[] { "name"});

The LUCENE operator (deprecated)

NOTE: LUCENE operator is translated to SEARCH_FIELDS function, but it doesn't support the metadata JSON

You can query the Lucene FullText Index using the custom operator LUCENE with the Query Parser Syntax from the Lucene Engine.

SELECT FROM V WHERE name LUCENE "test*"

This query searches for test, tests, tester, and so on from the property name of the class V. The query can use proximity operator ~, the required (+) and prohibit (-) operators, phrase queries, regexp queries:

SELECT FROM Article WHERE content LUCENE "(+graph -rdbms) AND +cloud"

Working with multiple fields (deprecated)

NOTE: define a single Lucene index on the class and use SEARCH_CLASS function

In addition to the standard Lucene query above, you can also query multiple fields. For example,

SELECT FROM Class WHERE [prop1, prop2] LUCENE "query"

In this case, if the word query is a plain string, the engine parses the query using MultiFieldQueryParser on each indexed field.

To execute a more complex query on each field, surround your query with parentheses, which causes the query to address specific fields.

SELECT FROM Article WHERE [content, author] LUCENE "(content:graph AND author:john)"

Here, the engine parses the query using the QueryParser

Creating a Manual Lucene Index (deprecated)

NOTE: avoid manual Lucene index

The Lucene Engine supports index creation without the need for a class.

Syntax:

CREATE INDEX  FULLTEXT ENGINE LUCENE  [] [METADATA {}]

For example, create a manual index using the CREATE INDEX command:

CREATE INDEX Manual FULLTEXT ENGINE LUCENE STRING, STRING

Once you have created the index Manual, you can insert values in index using the INSERT INTO INDEX:... command.

INSERT INTO INDEX:Manual (key, rid) VALUES(['Enrico', 'Rome'], #5:0)

You can then query the index through SELECT...FROM INDEX::

SELECT FROM INDEX:Manual WHERE key LUCENE "Enrico"

Manual indexes could be created programmatically using the Java API

ODocument meta = new ODocument().field("analyzer", StandardAnalyzer.class.getName());
OIndex index = databaseDocumentTx.getMetadata().getIndexManager()
     .createIndex("apiManual", OClass.INDEX_TYPE.FULLTEXT.toString(),
         new OSimpleKeyIndexDefinition(1, OType.STRING, OType.STRING), null, null, meta, OLuceneIndexFactory.LUCENE_ALGORITHM);