Nominatim/docs/develop/Tokenizers.md

Tokenizers

The tokenizer is the component of Nominatim that is responsible for analysing names of OSM objects and queries. Nominatim provides different tokenizers that use different strategies for normalisation. This page describes how tokenizers are expected to work and the public API that needs to be implemented when creating a new tokenizer. For information on how to configure a specific tokenizer for a database see the tokenizer chapter in the administration guide.

Generic Architecture

About Search Tokens

Search in Nominatim is organised around search tokens. Such a token represents string that can be part of the search query. Tokens are used so that the search index does not need to be organised around strings. Instead the database saves for each place which tokens match this place's name, address, house number etc. To be able to distinguish between these different types of information stored with the place, a search token also always has a certain type: name, house number, postcode etc.

During search an incoming query is transformed into a ordered list of such search tokens (or rather many lists, see below) and this list is then converted into a database query to find the right place.

It is the core task of the tokenizer to create, manage and assign the search tokens. The tokenizer is involved in two distinct operations:

• at import time: scanning names of OSM objects, normalizing them and building up the list of search tokens.
• at query time: scanning the query and returning the appropriate search tokens.

Importing

The indexer is responsible to enrich an OSM object (or place) with all data required for geocoding. It is split into two parts: the controller collects the places that require updating, enriches the place information as required and hands the place to Postgresql. The collector is part of the Nominatim library written in Python. Within Postgresql, the placex_update trigger is responsible to fill out all secondary tables with extra geocoding information. This part is written in PL/pgSQL.

The tokenizer is involved in both parts. When the indexer prepares a place, it hands it over to the tokenizer to inspect the names and create all the search tokens applicable for the place. This usually involves updating the tokenizer's internal token lists and creating a list of all token IDs for the specific place. This list is later needed in the PL/pgSQL part where the indexer needs to add the token IDs to the appropriate search tables. To be able to communicate the list between the Python part and the pl/pgSQL trigger, the placex table contains a special JSONB column token_info which is there for the exclusive use of the tokenizer.

The Python part of the tokenizer returns a structured information about the tokens of a place to the indexer which converts it to JSON and inserts it into the token_info column. The content of the column is then handed to the PL/pqSQL callbacks of the tokenizer which extracts the required information. Usually the tokenizer then removes all information from the token_info structure, so that no information is ever persistently saved in the table. All information that went in should have been processed after all and put into secondary tables. This is however not a hard requirement. If the tokenizer needs to store additional information about a place permanently, it may do so in the token_info column. It just may never execute searches over it and consequently not create any special indexes on it.

Querying

At query time, Nominatim builds up multiple interpretations of the search query. Each of these interpretations is tried against the database in order of the likelihood with which they match to the search query. The first interpretation that yields results wins.

The interpretations are encapsulated in the SearchDescription class. An instance of this class is created by applying a sequence of search tokens to an initially empty SearchDescription. It is the responsibility of the tokenizer to parse the search query and derive all possible sequences of search tokens. To that end the tokenizer needs to parse the search query and look up matching words in its own data structures.

Tokenizer API

The following section describes the functions that need to be implemented for a custom tokenizer implementation.

!!! warning This API is currently in early alpha status. While this API is meant to be a public API on which other tokenizers may be implemented, the API is far away from being stable at the moment.

Directory Structure

Nominatim expects two files for a tokenizer:

• nominiatim/tokenizer/<NAME>_tokenizer.py containing the Python part of the implementation
• lib-php/tokenizer/<NAME>_tokenizer.php with the PHP part of the implementation

where <NAME> is a unique name for the tokenizer consisting of only lower-case letters, digits and underscore. A tokenizer also needs to install some SQL functions. By convention, these should be placed in lib-sql/tokenizer.

If the tokenizer has a default configuration file, this should be saved in the settings/<NAME>_tokenizer.<SUFFIX>.

Configuration and Persistance

Tokenizers may define custom settings for their configuration. All settings must be prefixed with NOMINATIM_TOKENIZER_. Settings may be transient or persistent. Transient settings are loaded from the configuration file when Nominatim is started and may thus be changed at any time. Persistent settings are tied to a database installation and must only be read during installation time. If they are needed for the runtime then they must be saved into the nominatim_properties table and later loaded from there.

The Python module

The Python module is expect to export a single factory function:

def create(dsn: str, data_dir: Path) -> AbstractTokenizer

The dsn parameter contains the DSN of the Nominatim database. The data_dir is a directory in the project directory that the tokenizer may use to save database-specific data. The function must return the instance of the tokenizer class as defined below.

Python Tokenizer Class

All tokenizers must inherit from nominatim.tokenizer.base.AbstractTokenizer and implement the abstract functions defined there.

::: nominatim.tokenizer.base.AbstractTokenizer rendering: heading_level: 4

Python Analyzer Class

::: nominatim.tokenizer.base.AbstractAnalyzer rendering: heading_level: 4

PL/pgSQL Functions

The tokenizer must provide access functions for the token_info column to the indexer which extracts the necessary information for the global search tables. If the tokenizer needs additional SQL functions for private use, then these functions must be prefixed with token_ in order to ensure that there are no naming conflicts with the SQL indexer code.

The following functions are expected:

FUNCTION token_get_name_search_tokens(info JSONB) RETURNS INTEGER[]

Return an array of token IDs of search terms that should match the name(s) for the given place. These tokens are used to look up the place by name and, where the place functions as part of an address for another place, by address. Must return NULL when the place has no name.

FUNCTION token_get_name_match_tokens(info JSONB) RETURNS INTEGER[]

Return an array of token IDs of full names of the place that should be used to match addresses. The list of match tokens is usually more strict than search tokens as it is used to find a match between two OSM tag values which are expected to contain matching full names. Partial terms should not be used for match tokens. Must return NULL when the place has no name.

FUNCTION token_get_housenumber_search_tokens(info JSONB) RETURNS INTEGER[]

Return an array of token IDs of house number tokens that apply to the place. Note that a place may have multiple house numbers, for example when apartments each have their own number. Must be NULL when the place has no house numbers.

FUNCTION token_normalized_housenumber(info JSONB) RETURNS TEXT

Return the house number(s) in the normalized form that can be matched against a house number token text. If a place has multiple house numbers they must be listed with a semicolon as delimiter. Must be NULL when the place has no house numbers.

FUNCTION token_addr_street_match_tokens(info JSONB) RETURNS INTEGER[]

Return the match token IDs by which to search a matching street from the addr:street tag. These IDs will be matched against the IDs supplied by token_get_name_match_tokens. Must be NULL when the place has no addr:street tag.

FUNCTION token_addr_place_match_tokens(info JSONB) RETURNS INTEGER[]

Return the match token IDs by which to search a matching place from the addr:place tag. These IDs will be matched against the IDs supplied by token_get_name_match_tokens. Must be NULL when the place has no addr:place tag.

FUNCTION token_addr_place_search_tokens(info JSONB) RETURNS INTEGER[]

Return the search token IDs extracted from the addr:place tag. These tokens are used for searches by address when no matching place can be found in the database. Must be NULL when the place has no addr:place tag.

CREATE TYPE token_addresstoken AS (
  key TEXT,
  match_tokens INT[],
  search_tokens INT[]
);

FUNCTION token_get_address_tokens(info JSONB) RETURNS SETOF token_addresstoken

Return the match and search token IDs for explicit addr:* tags for the place other than addr:street and addr:place. For each address item there are three pieces of information returned:

• key contains the type of address item (city, county, etc.). This is the key handed in with the address dictionary.
• match_tokens is the list of token IDs used to find the corresponding place object for the address part. The list is matched against the IDs from token_get_name_match_tokens.
• search_tokens is the list of token IDs under which to search the address item. It is used when no corresponding place object was found.

FUNCTION token_normalized_postcode(postcode TEXT) RETURNS TEXT

Return the normalized version of the given postcode. This function must return the same value as the Python function AbstractAnalyzer->normalize_postcode().

FUNCTION token_strip_info(info JSONB) RETURNS JSONB

Return the part of the token_info field that should be stored in the database permanently. The indexer calls this function when all processing is done and replaces the content of the token_info column with the returned value before the trigger stores the information in the database. May return NULL if no information should be stored permanently.

PHP Tokenizer class

The PHP tokenizer class is instantiated once per request and responsible for analyzing the incoming query. Multiple requests may be in flight in parallel.

The class is expected to be found under the name of \Nominatim\Tokenizer. To find the class the PHP code includes the file tokenizer/tokenizer.php in the project directory. This file must be created when the tokenizer is first set up on import. The file should initialize any configuration variables by setting PHP constants and then require the file with the actual implementation of the tokenizer.

The tokenizer class must implement the following functions:

public function __construct(object &$oDB)

The constructor of the class receives a database connection that can be used to query persistent data in the database.

public function checkStatus()

Check that the tokenizer can access its persistent data structures. If there is an issue, throw an \Exception.

public function normalizeString(string $sTerm) : string

Normalize string to a form to be used for comparisons when reordering results. Nominatim reweighs results how well the final display string matches the actual query. Before comparing result and query, names and query are normalised against this function. The tokenizer can thus remove all properties that should not be taken into account for reweighing, e.g. special characters or case.

public function tokensForSpecialTerm(string $sTerm) : array

Return the list of special term tokens that match the given term.

public function extractTokensFromPhrases(array &$aPhrases) : TokenList

Parse the given phrases, splitting them into word lists and retrieve the matching tokens.

The phrase array may take on two forms. In unstructured searches (using q= parameter) the search query is split at the commas and the elements are put into a sorted list. For structured searches the phrase array is an associative array where the key designates the type of the term (street, city, county etc.) The tokenizer may ignore the phrase type at this stage in parsing. Matching phrase type and appropriate search token type will be done later when the SearchDescription is built.

For each phrase in the list of phrases, the function must analyse the phrase string and then call setWordSets() to communicate the result of the analysis. A word set is a list of strings, where each string refers to a search token. A phrase may have multiple interpretations. Therefore a list of word sets is usually attached to the phrase. The search tokens themselves are returned by the function in an associative array, where the key corresponds to the strings given in the word sets. The value is a list of search tokens. Thus a single string in the list of word sets may refer to multiple search tokens.