Trie-Based Prefix Index for Java
Switch branches/tags
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
src
.gitignore
LICENSE
README.md
pom.xml

README.md

InstaTrie

Maven Central

High-performance prefix index and search word highlighter, implemented with trie data structures.

Key Features

  • Sophisticated generic data structure for efficient prefix-lookup
  • Support for custom extractor functions to retrieve keywords from POJOs
  • Ability to ignore (simple) HTML Tags
  • Highlighting of found parts in searched models
  • Order of entries is retained
  • Optional ASCII-Normalization of search keywords and queries
  • Ability to create proxies for the data structure with additional filters
  • Compatible to Java List API
  • Support for Java8 Streams API
  • Support for Java9+ Jigsaw module system
  • No further dependencies (except for the ASCII-normalization utility)

Setup

To use InstaTrie in your Maven project, you can simply add instatrie as dependency:

<dependency>
    <groupId>com.illucit</groupId>
    <artifactId>instatrie</artifactId>
    <version>2.0.0/version>
</dependency>

The only dependency of instatrie is the library lucene-utils for String normalization, which has no further dependencies itself.

Trie Data Structure

A "Trie" is a tree-like data structure, which has a character on each edge and where each node represents a prefix. It is known to be extremely efficient for storing and retrieving all words with a common prefix. For further theoretical information about tries, you can look in the Wikipedia.

The interface com.illucit.instatrie.trie.PrefixDictionary<T> and class com.illucit.instatrie.trie.Trie<T> provide an implementation of this data structure, which can hold a set of stored Strings and can retrieve all matching Strings with a given prefix. For each inserted String you can also add a payload object of the generic type <T>.

Although the Trie class was basically implemented for the Prefix Index functionality of this library, you can use it as its own. The Trie class does not implement java.util.Collection, but provides the most common methods to manipulate data. See the public interface com.illucit.instatrie.trie.PrefixDictionary<T> for a full list of supported operations.

Prefix Index Data Structure

The "Prefix Index" data structure resembles a generic collection of models of a specific type, which can be filtered by search queries.

Word Splitter

First of all, you define a function to extract a searchable String from each model, which will be indexed. This function is called a "Word Splitter" and has to be realized by a class that implements.com.illucit.instatrie.splitter.WordSplitter<T>.
E.g. if your model type is a class Person with properties firstName and lastName, your Word Splitter should take an instance of Person and return a Set containing the first name and last name of the person.
You can also use the more convenient class com.illucit.instatrie.splitter.StringWordSplitter<T>, for which you only have to prove a function get one single String from your model and the class will provide all "indexable" text fragments by a regular expression (per default all connected alphanumeric substrings) from this String representation. This also has the advantage that you don't need to provide an individual splitting into different words (e.g. if your person has multiple first names), but you only give the regular expression a String to work on.
E.g. for your example with the Person class:

WordSplitter<Person> splitter;
splitter = new StringWordSplitter<>(p -> p.getFirstName() + " " + p.getLastName());

So if you have a Person with firstName = "John Walter" and lastName = "Doe", your Splitter will provide a set with "John", "Walter" and `"Doe"``.

Note: If your model contains values with HTML tags which you want to make searchable (and therefore include in the splitter), make sure you remove the tags before (e.g. with an HTML parser utility like JSoup).

Unicode Normalization

If you actually use the StringWordSplitter class, the given String can be normalized, i.e. special unicode characters are replaced by their "basic" characters where possible, e.g. "ö" -> "o", "Â" -> "A" or "©" -> "c".

TriePrefixIndex

The interface com.illucit.instatrie.index.PrefixIndex<T> and class com.illucit.instatrie.index.TriePrefixIndex<T> provide an implementation for the "Prefix Index" collection type. The collection basically consists of an (ordered) list of models of the generic type "<T>", but provides additional methods to query the list of models for a search String and get back a (reduced, still ordered) list of models that contain this search query.

You can create and instance of the com.illucit.instatrie.index.TriePrefixIndex<T> class with an empty index by providing a Splitter for "<T>". Once the prefix index instance is created, you can add data to it by calling the TriePrefixIndex.createIndex(Collection<T> models) method with a new list of models, which will overwrite the previous index and build a new index with the given collection and the fixed word splitter.

TriePrefixIndex<Person> index = new TriePrefixIndex<>(wordSplitter);
List<Person> personList = ...
index.createIndex(personList);

PrefixSearch

The interface com.illucit.instatrie.index.PrefixSearc<T> describes the basic operations to query the search index for data. The underlying implementation provides the queried data as Java8 streams, but there are also return function as Lists.

In order to use functional composition of the search data, the PrefixSearch interface has filter and map methods, which return a dependent PrefixSearch instance coupled to the index of the original PrefixSearch but applying the given mapping- or filter-function to each results.

Querying

The "Prefix Index" collection provides two basic query operations: search and searchExact. While the operation search will retrieve all models which match all given search terms as prefix, the operation searchExact will get you all models which contain all given search terms as complete words. Searching is alsways case-insensititve, so the case of the model search terms and the query String won't matter.

First of all, the search query will itself always be splitted into individual words with the help of a StringSearchSplitter. So if you search for multiple words at once, it is always considered as "AND" search, meaning all searched terms must be found for a hit. There is currently no support for query operations, as the "Prefix Index" is built for filtering and narrowing something down, not as full-fledged search engine.

  • The query "DOE" will only search for the word "doe".
  • The queries "doe, john", "john doe" and "......(joHN]] DOe?" will all result in the search for the two terms "john" and "doe".
  • An empty query (i.e. no alphanumerical characters) will always return the full list of models as there are no filter criteria that can fail.

When one of the search methods is called, the search words are compared with the splitted words in the index for each model individually, so a model in the index can either match the criteria or not. Given the Person model mentioned above, the model has the index terms "john", "walter" and "doe".

  • If you make an exact query for "doe" the model will match, as the word "doe" is fully contained in the set of index terms
  • If you make an exact query for "walt" the model will not match, as the word "walt" is only a prefix of the index term "walter", but not fully contained in the index terms
  • If you make a (non-exact) query for "walt" the model will match, because "walt" is a prefix of the index term "walter"
  • If you make a query for "jane" and "doe" the model will not match, because "jane" is not a prefix of an index term (although "doe" is)

Note:If a model matches an exact query, it always also matches the non-exact query, because each word is considered a prefix of itself

List Interface

The PrefixSearch interface does not directly extend a class from the Java Collection API. It provides, however, a method decorateAsList which will return a proxy object which implements java.util.List<T> and is still connected to the original PrefixSearch instance. All write-operations of the List-proxy will result in UnsupportedOperationException, and all read-operations will be delegated to the internal list of models contained in the current index of the PrefixSearch.

The returned proxy is permanently connected to the original TriePrefixIndex instance, so when you rebuild the index on one of the instances, it will also change the index of the other instance. All Trie query operations on both instances will return exactly the same result.

Concurrency

The class TriePrefixIndex is designed for concurrent access. An instance of TriePrefixIndex always takes the most recent available index data. If you call the createIndex method, until the method is finished all method call querying the prefix index will still be going to the current index. Once the index calculation is done, it will be set as most recent index for the TriePrefixIndex in an atomic operation and all future query calls will use the new index.

You can use a TriePrefixIndex instance from multiple threads simultaniously without gettings "concurrent modification" problems, as each query call will use the index data from the beginning of the call during the whole execution of the query.

If multiple Threads share an instance of TriePrefixIndex and one thread refreshes the index, all Threads will get the new index from the moment it is calculated completely.

Java8 Features

In Java8 it is possible to use "Lambda Expressions" to provide implementations of interfaces with only 1 unimplemented method (like java.util.Function). Instead of providing a WordSplitter object when creating a TriePrefixIndex, you can also directly pass a lambda expression for a StringWordSplitter in the constructor:

 TriePrefixIndex<Person> index = new TriePrefixIndex<>(p -> p.getFirstName() + " " + p.getLastName());

The PrefixIndex also provides a method searchStream which behaves exactly like the search method, but instead of returning a List of models, it will return a Java8 java.util.stream.Stream<T> containing the models.

Highlighting

When you use the PrefixIndex to filter down elements in a graphical user interface (GUI), especially on a web page, it is helpful to see where a model matched the search terms. For this, the InstaTrie library also provides String highlighting capabilities.

The interface com.illucit.instatrie.highlight.SubwordHighlighter contains methods to generate a String segmentation of a given model String property, in which the highlighted parts can be identified. The default implementation of SubwordHighlighter is the class StringWordSplitter, also used in splitting. The class TriePrefixIndex also implements the interface SubwordHighlighter and delegates all its calls to the StringWordSplitter used for splitting the search queries.

There are two different modi in which the highlighting can run: "HTML mode" and "Text mode". The "HTML mode" is designed for the cases in which the searchable model properties can contain simple HTML markup, e.g. bold, underlined, subscript etc. Because InstaTrie has no bundled HTML parser, only simple tags without attributes and with distinct closing tags (i.e. no "empty" tags) are supported. These tags are ignored when calculating the search highlights, even if the highlights span across multiple tags. In "Text mode" the given model String is taken "as-is". In both cases the matching is ignoring case and also ignoring unicode specialities (unless deactivated in the StringWordSplitter), so that e.g. "DØĚ" will match "doe".

To use the highlighting, you can call the methods highlightSubwordPrefixes(value, queryWords) or highlightSubwordPrefixesWithHtml(value, queryWords) on a SubwordHighlighter instance. The first parameter is the model String value that you want to render with highlighting and the second parameter is a Set with the search termn that were used for search, as returned by the WordSplitter. The TriePrefixIndex class also provides the methods getHighlighted(modelValue, query) and getHighlightedHtml(modelValue, query) which do the splitting of the query String query into the queryWords set for you with the internal query splitter.

The returned value is an instance of the class com.illucit.instatrie.highlight.HighlightedString which contains a method getSegments() returning a list of all highlighted and non-highligted segmenty, you can use for rendering.

E.g.

Person john = new Person("John Walter", "Doe");
Person jane = new Person("Jane", "Doe");
TriePrefixIndex<Person> index = ...
String search = "walt do";
HighlightedString highlighted;

highlighted = index.getHighlighted(john.getFirstName(), search);
for (HighlightSegment segment : highlighted.getSegments()) {
    if (segment.isHighlighted()) {
        System.out.print("[[" + segment.getValue() + "]]");
    } else {
        System.out.print(segment.getValue());
    }
}
System.out.println();
// Prints: John [[Walt]]er

highlighted = index.getHighlighted(john.getLastName(), search);
for (HighlightSegment segment : highlighted.getSegments()) {
    if (segment.isHighlighted()) {
        System.out.print("[[" + segment.getValue() + "]]");
    } else {
        System.out.print(segment.getValue());
    }
}
System.out.println();
// Prints: [[Do]]e

highlighted = index.getHighlighted(jane.getFirstName(), search);
for (HighlightSegment segment : highlighted.getSegments()) {
    if (segment.isHighlighted()) {
        System.out.print("[[" + segment.getValue() + "]]");
    } else {
        System.out.print(segment.getValue());
    }
}
System.out.println();
// Prints: Jane

Future Planning

There are some feature ideas which didn't make it into the current implementation:

  • Currently only the whole index can be refreshed. If you want to add a model to the index, you have to combine the current list of models with the new model in a temporary list, and create a new index from the whole new list.