IResearch search engine

Version 1.3

Table of contents

Overview

The IResearch library is meant to be treated as a standalone index that is capable of both indexing and storing individual values verbatim. Indexed data is treated on a per-version/per-revision basis, i.e. existing data version/revision is never modified and updates/removals are treated as new versions/revisions of the said data. This allows for trivial multi-threaded read/write operations on the index. The index exposes its data processing functionality via a multi-threaded ‘writer’ interface that treats each document abstraction as a collection of fields to index and/or store. The index exposes its data retrieval functionality via ‘reader’ interface that returns records from an index matching a specified query. The queries themselves are constructed query trees built directly using the query building blocks available in the API. The querying infrastructure provides the capability of ordering the result set by one or more ranking/scoring implementations. The ranking/scoring implementation logic is plugin-based and lazy-initialized during runtime as needed, allowing for addition of custom ranking/scoring logic without the need to even recompile the IResearch library.

High level architecture and main concepts

Index

An index consists of multiple independent parts, called segments and index metadata. Index metadata stores information about active index segments for the particular index version/revision. Each index segment is an index itself and consists of the following logical components:

Read/write access to the components carried via plugin-based formats. Index may contain segments created using different formats.

Document

A database record is represented as an abstraction called a document. A document is actually a collection of indexed/stored fields. In order to be processed each field should satisfy at least IndexedField or StoredField concept.

IndexedField concept

For type T to be IndexedField, the following conditions have to be satisfied for an object m of type T:

Expression Requires Effects
m.name() The output type must be convertible to irs::string_ref A value uses as a key name.
m.get_tokens() The output type must be convertible to irs::token_stream* A token stream uses for populating in invert procedure. If value is nullptr field is treated as non-indexed.
m.index_features() The output type must be implicitly convertible to irs::IndexFeatures A set of features requested for evaluation during indexing. E.g. it may contain request of processing positions and frequencies. Later the evaluated information can be used during querying and scoring.
m.features() The output type must be convertible to const irs::flags& A set of user supplied features to be associated with a field. E.g. it may contain request of storing field norms. Later the stored information can be used during querying and scoring.

StoredField concept

For type T to be StoredField, the following conditions have to be satisfied for an object m of type T:

Expression Requires Effects
m.name() The output type must be convertible to irs::string_ref A value uses as a key name.
m.write(irs::data_output& out) The output type must be convertible to bool. One may write arbitrary data to stream denoted by out in order to retrieve written value using index_reader API later. If nothing has written but returned value is true then stored value is treated as flag. If returned value is false then nothing is stored even if something has been written to out stream.

Directory

A data storage abstraction that can either store data in memory or on the filesystem depending on which implementation is instantiated. A directory stores at least all the currently in-use index data versions/revisions. For the case where there are no active users of the directory then at least the last data version/revision is stored. Unused data versions/revisions may be removed via the directory_cleaner. A single version/revision of the index is composed of one or more segments associated, and possibly shared, with the said version/revision.

Writer

A single instance per-directory object that is used for indexing data. Data may be indexed in a per-document basis or sourced from another reader for trivial directory merge functionality. Each commit() of a writer produces a new version/revision of the view of the data in the corresponding directory. Additionally the interface also provides directory defragmentation capabilities to allow compacting multiple smaller version/revision segments into larger more compact representations. A writer supports two-phase transactions via begin()/commit()/rollback() methods.

Reader

A reusable/refreshable view of an index at a given point in time. Multiple readers can use the same directory and may point to different versions/revisions of data in the said directory.

Build prerequisites

CMake

v3.10 or later

Boost

v1.57.0 or later (headers only)

set environment

BOOST_ROOT=<path-to>/boost_1_57_0

Lz4

install (*nix)

make
make install

or point LZ4_ROOT at the source directory to build together with IResearch

install (win32)

If compiling IResearch with /MT add add_definitions(“/MTd”) to the end of cmake_unofficial/CMakeLists.txt since cmake will ignore the command line argument -DCMAKE_C_FLAGS=/MTd

mkdir build && cd build
cmake -DCMAKE_INSTALL_PREFIX=<install-path> -DBUILD_STATIC_LIBS=on -g "Visual studio 17" -Ax64 ../contrib/cmake_unofficial
cmake --build .
cmake --build . --target install

or point LZ4_ROOT at the source directory to build together with IResearch

set environment

LZ4_ROOT=<install-path>

win32 binaries also available in:

ICU

v53 or higher

install (*nix)

./configure --disable-samples --disable-tests --enable-static --srcdir="$(pwd)" --prefix=<install-path> --exec-prefix=<install-path>
make install

or point ICU_ROOT at the source directory to build together with IResearch or via the distributions’ package manager: libicu

install (win32)

look for link: “ICU4C Binaries”

set environment

ICU_ROOT=<path-to-icu>

Snowball

install (*nix)

the custom CMakeLists.txt is intended to be used with snowball v2.0.0 and later versions. At least it was tested to work on commit 53739a805cfa6c77ff8496dc711dc1c106d987c1

git clone https://github.com/snowballstem/snowball.git
mkdir build && cd build
cmake -DENABLE_STATIC=OFF -DNO_SHARED=OFF -g "Unix Makefiles" ..
cmake --build .
cmake -DENABLE_STATIC=OFF -DNO_SHARED=ON -g "Unix Makefiles" ..
cmake --build .

or point SNOWBALL_ROOT at the source directory to build together with IResearch or via the distributions’ package manager: libstemmer

install (win32)

the custom CMakeLists.txt was based on revision 5137019d68befd633ce8b1cd48065f41e77ed43e later versions may be used at your own risk of compilation failure

git clone https://github.com/snowballstem/snowball.git
git reset --hard adc028f3ae646623bda2f99191fe9dc3287a909b
mkdir build && cd build
set PATH=%PATH%;<path-to>/build/Debug
cmake -DENABLE_STATIC=OFF -DNO_SHARED=OFF -g "Visual studio 12" -Ax64 ..
cmake --build .
cmake -DENABLE_STATIC=OFF -DNO_SHARED=ON -g "Visual studio 12" -Ax64 ..
cmake --build .

or point SNOWBALL_ROOT at the source directory to build together with IResearch

For static builds:

  1. in MSVC open: build/snowball.sln
  2. set: stemmer -> Properties -> Configuration Properties -> C/C++ -> Code Generation -> Runtime Library = /MTd
  3. BUILD -> Build Solution

set environment

SNOWBALL_ROOT=<path-to-snowball>

VelocyPack

point VPACK_ROOT at the source directory to build together with IResearch

Gooogle test

install (*nix)

mkdir build && cd build
cmake ..
make

or point GTEST_ROOT at the source directory to build together with IResearch

install (win32)

mkdir build && cd build
cmake -g "Visual studio 12" -Ax64 -Dgtest_force_shared_crt=ON -DCMAKE_DEBUG_POSTFIX="" ..
cmake --build .
mv Debug ../lib

or point GTEST_ROOT at the source directory to build together with IResearch

set environment

GTEST_ROOT=<path-to-gtest>

Stopword list (for use with analysis::text_analyzer)

download any number of lists of stopwords, e.g. from: https://github.com/snowballstem/snowball-website/tree/master/algorithms/*/stop.txt https://code.google.com/p/stop-words/

install

  1. mkdir
  2. for each language, (e.g. “c”, “en”, “es”, “ru”), create a corresponding subdirectory (a directory name has 2 letters except the default locale “c” which has 1 letter)
  3. place the files with stopwords, (utf8 encoded with one word per line, any text after the first whitespace is ignored), in the directory corresponding to its language (multiple files per language are supported and will be interpreted as a single list)

set environment

IRESEARCH_TEXT_STOPWORD_PATH=<path-to-stopword-lists>

If the variable IRESEARCH_TEXT_STOPWORD_PATH is left unset then locale specific stopword-list subdirectories are deemed to be located in the current working directory

Build

git clone <IResearch code repository>/iresearch.git iresearch
cd iresearch
mkdir build && cd build

generate build file <*nix>:

cmake -DCMAKE_BUILD_TYPE=[Debug|Release|Coverage] -g "Unix Makefiles" ..
  1. if some libraries are not found by the build then set the needed environment

    variables (e.g. BOOST_ROOT, BOOST_LIBRARYDIR, LZ4_ROOT, OPENFST_ROOT, GTEST_ROOT)

  2. if ICU or Snowball from the distribution paths are not found, the following additional

    environment variables might be required: ICU_ROOT_SUFFIX=x86_64-linux-gnu SNOWBALL_ROOT_SUFFIX=x86_64-linux-gnu

generate build file (win32):

cmake -g "Visual studio 12" -Ax64 ..

If some libraries are not found by the build then set the needed environment variables (e.g. BOOST_ROOT, BOOST_LIBRARYDIR, LZ4_ROOT, OPENFST_ROOT, GTEST_ROOT)

set Build Identifier for this build (optional)

echo "<build_identifier>" > BUILD_IDENTIFIER

build library:

cmake --build .

test library:

cmake --build . --target iresearch-check

install library:

cmake --build . --target install

code coverage:

cmake --build . --target iresearch-coverage

Pyresearch

There is Python wrapper for IResearch. Wrapper gives access to directory reader object. For usage example see /python/scripts

Build

To build Pyresearch SWIG generator should be available. Add -DUSE_PYRESEARCH=ON to cmake command-line to generate Pyresearch targets

Install

Run target pyresearch-install

win32 install notes:

Some version of ICU installers seems to fail to make available all icu dlls through PATH enviroment variable, manual adjustment may be needed.

(*nix) install notes:

Shared version of libiresearch is used. Install IResearch before running Pyresearch.

External 3rd party dependencies

External 3rd party dependencies must be made available to the IResearch library separately. They may either be installed through the distribution package management system or build from source and the appropriate environment variables set accordingly.

Boost

v1.57.0 or later (locale system thread) used for functionality not available in the STL (excluding functionality available in ICU)

Lz4

used for compression/decompression of byte/string data

ICU

used by analyzers for parsing, transforming and tokenising string data

Snowball

used by analyzers for computing word stems (i.e. roots) for more flexible matching matching of words from languages not supported by ‘snowball’ are done verbatim

Google Test

used for writing tests for the IResearch library

VelocyPack

used for JSON serialization/deserialization

Stopword list

used by analysis::text_analyzer for filtering out noise words that should not impact text ranging e.g. for ‘en’ these are usualy ‘a’, ‘the’, etc… download any number of lists of stopwords, e.g. from: https://github.com/snowballstem/snowball-website/tree/master/algorithms/*/stop.txt https://code.google.com/p/stop-words/ or create a custom language-specific list of stopwords place the files with stopwords, (utf8 encoded with one word per line, any text after the first whitespace is ignored), in the directory corresponding to its language (multiple files per language are supported and will be interpreted as a single list)

Query filter building blocks

Filter Description
irs::by_edit_distance for filtering of values based on Levenshtein distance
irs::by_granular_range for faster filtering of numeric values within a given range, with the possibility of specifying open/closed ranges
irs::by_ngram_similarity for filtering of values based on NGram model
irs::by_phrase for word-position-sensitive filtering of values, with the possibility of skipping selected positions
irs::by_prefix for filtering of exact value prefixes
irs::by_range for filtering of values within a given range, with the possibility of specifying open/closed ranges
irs::by_same_position for term-insertion-order sensitive filtering of exact values
irs::by_term for filtering of exact values
irs::by_terms for filtering of exact values by a set of specified terms
irs::by_wildcard for filtering of values based on matching pattern
irs::ByNestedFilter for filtering of documents based on matching pattern on its sub-documents
irs::And boolean conjunction of multiple filters, influencing document ranks/scores as appropriate
irs::Or boolean disjunction of multiple filters, influencing document ranks/scores as appropriate (including “minimum match” functionality)
irs::Not boolean negation of multiple filters

Supported compilers

License

Copyright (c) 2017-2023 ArangoDB GmbH

Copyright (c) 2016-2017 EMC Corporation

This software is provided under the Apache 2.0 Software license provided in the LICENSE.md file. Licensing information for third-party products used by IResearch search engine can be found in THIRD_PARTY_README.md