3 The tokenizer is the component of Nominatim that is responsible for
4 analysing names of OSM objects and queries. Nominatim provides different
5 tokenizers that use different strategies for normalisation. This page describes
6 how tokenizers are expected to work and the public API that needs to be
7 implemented when creating a new tokenizer. For information on how to configure
8 a specific tokenizer for a database see the
9 [tokenizer chapter in the Customization Guide](../customize/Tokenizers.md).
11 ## Generic Architecture
13 ### About Search Tokens
15 Search in Nominatim is organised around search tokens. Such a token represents
16 string that can be part of the search query. Tokens are used so that the search
17 index does not need to be organised around strings. Instead the database saves
18 for each place which tokens match this place's name, address, house number etc.
19 To be able to distinguish between these different types of information stored
20 with the place, a search token also always has a certain type: name, house number,
23 During search an incoming query is transformed into a ordered list of such
24 search tokens (or rather many lists, see below) and this list is then converted
25 into a database query to find the right place.
27 It is the core task of the tokenizer to create, manage and assign the search
28 tokens. The tokenizer is involved in two distinct operations:
30 * __at import time__: scanning names of OSM objects, normalizing them and
31 building up the list of search tokens.
32 * __at query time__: scanning the query and returning the appropriate search
38 The indexer is responsible to enrich an OSM object (or place) with all data
39 required for geocoding. It is split into two parts: the controller collects
40 the places that require updating, enriches the place information as required
41 and hands the place to Postgresql. The collector is part of the Nominatim
42 library written in Python. Within Postgresql, the `placex_update`
43 trigger is responsible to fill out all secondary tables with extra geocoding
44 information. This part is written in PL/pgSQL.
46 The tokenizer is involved in both parts. When the indexer prepares a place,
47 it hands it over to the tokenizer to inspect the names and create all the
48 search tokens applicable for the place. This usually involves updating the
49 tokenizer's internal token lists and creating a list of all token IDs for
50 the specific place. This list is later needed in the PL/pgSQL part where the
51 indexer needs to add the token IDs to the appropriate search tables. To be
52 able to communicate the list between the Python part and the pl/pgSQL trigger,
53 the `placex` table contains a special JSONB column `token_info` which is there
54 for the exclusive use of the tokenizer.
56 The Python part of the tokenizer returns a structured information about the
57 tokens of a place to the indexer which converts it to JSON and inserts it into
58 the `token_info` column. The content of the column is then handed to the PL/pqSQL
59 callbacks of the tokenizer which extracts the required information. Usually
60 the tokenizer then removes all information from the `token_info` structure,
61 so that no information is ever persistently saved in the table. All information
62 that went in should have been processed after all and put into secondary tables.
63 This is however not a hard requirement. If the tokenizer needs to store
64 additional information about a place permanently, it may do so in the
65 `token_info` column. It just may never execute searches over it and
66 consequently not create any special indexes on it.
70 At query time, Nominatim builds up multiple _interpretations_ of the search
71 query. Each of these interpretations is tried against the database in order
72 of the likelihood with which they match to the search query. The first
73 interpretation that yields results wins.
75 The interpretations are encapsulated in the `SearchDescription` class. An
76 instance of this class is created by applying a sequence of
77 _search tokens_ to an initially empty SearchDescription. It is the
78 responsibility of the tokenizer to parse the search query and derive all
79 possible sequences of search tokens. To that end the tokenizer needs to parse
80 the search query and look up matching words in its own data structures.
84 The following section describes the functions that need to be implemented
85 for a custom tokenizer implementation.
88 This API is currently in early alpha status. While this API is meant to
89 be a public API on which other tokenizers may be implemented, the API is
90 far away from being stable at the moment.
92 ### Directory Structure
94 Nominatim expects a single file `src/nominatim_db/tokenizer/<NAME>_tokenizer.py`
95 containing the Python part of the implementation.
96 `<NAME>` is a unique name for the tokenizer consisting of only lower-case
97 letters, digits and underscore. A tokenizer also needs to install some SQL
98 functions. By convention, these should be placed in `lib-sql/tokenizer`.
100 If the tokenizer has a default configuration file, this should be saved in
101 the `settings/<NAME>_tokenizer.<SUFFIX>`.
103 ### Configuration and Persistence
105 Tokenizers may define custom settings for their configuration. All settings
106 must be prefixed with `NOMINATIM_TOKENIZER_`. Settings may be transient or
107 persistent. Transient settings are loaded from the configuration file when
108 Nominatim is started and may thus be changed at any time. Persistent settings
109 are tied to a database installation and must only be read during installation
110 time. If they are needed for the runtime then they must be saved into the
111 `nominatim_properties` table and later loaded from there.
113 ### The Python module
115 The Python module is expect to export a single factory function:
118 def create(dsn: str, data_dir: Path) -> AbstractTokenizer
121 The `dsn` parameter contains the DSN of the Nominatim database. The `data_dir`
122 is a directory in the project directory that the tokenizer may use to save
123 database-specific data. The function must return the instance of the tokenizer
124 class as defined below.
126 ### Python Tokenizer Class
128 All tokenizers must inherit from `nominatim_db.tokenizer.base.AbstractTokenizer`
129 and implement the abstract functions defined there.
131 ::: nominatim_db.tokenizer.base.AbstractTokenizer
135 ### Python Analyzer Class
137 ::: nominatim_db.tokenizer.base.AbstractAnalyzer
141 ### PL/pgSQL Functions
143 The tokenizer must provide access functions for the `token_info` column
144 to the indexer which extracts the necessary information for the global
145 search tables. If the tokenizer needs additional SQL functions for private
146 use, then these functions must be prefixed with `token_` in order to ensure
147 that there are no naming conflicts with the SQL indexer code.
149 The following functions are expected:
152 FUNCTION token_get_name_search_tokens(info JSONB) RETURNS INTEGER[]
155 Return an array of token IDs of search terms that should match
156 the name(s) for the given place. These tokens are used to look up the place
157 by name and, where the place functions as part of an address for another place,
158 by address. Must return NULL when the place has no name.
161 FUNCTION token_get_name_match_tokens(info JSONB) RETURNS INTEGER[]
164 Return an array of token IDs of full names of the place that should be used
165 to match addresses. The list of match tokens is usually more strict than
166 search tokens as it is used to find a match between two OSM tag values which
167 are expected to contain matching full names. Partial terms should not be
168 used for match tokens. Must return NULL when the place has no name.
171 FUNCTION token_get_housenumber_search_tokens(info JSONB) RETURNS INTEGER[]
174 Return an array of token IDs of house number tokens that apply to the place.
175 Note that a place may have multiple house numbers, for example when apartments
176 each have their own number. Must be NULL when the place has no house numbers.
179 FUNCTION token_normalized_housenumber(info JSONB) RETURNS TEXT
182 Return the house number(s) in the normalized form that can be matched against
183 a house number token text. If a place has multiple house numbers they must
184 be listed with a semicolon as delimiter. Must be NULL when the place has no
188 FUNCTION token_is_street_address(info JSONB) RETURNS BOOLEAN
191 Return true if this is an object that should be parented against a street.
192 Only relevant for objects with address rank 30.
195 FUNCTION token_has_addr_street(info JSONB) RETURNS BOOLEAN
198 Return true if there are street names to match against for finding the
199 parent of the object.
203 FUNCTION token_has_addr_place(info JSONB) RETURNS BOOLEAN
206 Return true if there are place names to match against for finding the
207 parent of the object.
210 FUNCTION token_matches_street(info JSONB, street_tokens INTEGER[]) RETURNS BOOLEAN
213 Check if the given tokens (previously saved from `token_get_name_match_tokens()`)
214 match against the `addr:street` tag name. Must return either NULL or FALSE
215 when the place has no `addr:street` tag.
218 FUNCTION token_matches_place(info JSONB, place_tokens INTEGER[]) RETURNS BOOLEAN
221 Check if the given tokens (previously saved from `token_get_name_match_tokens()`)
222 match against the `addr:place` tag name. Must return either NULL or FALSE
223 when the place has no `addr:place` tag.
227 FUNCTION token_addr_place_search_tokens(info JSONB) RETURNS INTEGER[]
230 Return the search token IDs extracted from the `addr:place` tag. These tokens
231 are used for searches by address when no matching place can be found in the
232 database. Must be NULL when the place has no `addr:place` tag.
235 FUNCTION token_get_address_keys(info JSONB) RETURNS SETOF TEXT
238 Return the set of keys for which address information is provided. This
239 should correspond to the list of (relevant) `addr:*` tags with the `addr:`
240 prefix removed or the keys used in the `address` dictionary of the place info.
243 FUNCTION token_get_address_search_tokens(info JSONB, key TEXT) RETURNS INTEGER[]
246 Return the array of search tokens for the given address part. `key` can be
247 expected to be one of those returned with `token_get_address_keys()`. The
248 search tokens are added to the address search vector of the place, when no
249 corresponding OSM object could be found for the given address part from which
250 to copy the name information.
253 FUNCTION token_matches_address(info JSONB, key TEXT, tokens INTEGER[])
256 Check if the given tokens match against the address part `key`.
258 __Warning:__ the tokens that are handed in are the lists previously saved
259 from `token_get_name_search_tokens()`, _not_ from the match token list. This
260 is an historical oddity which will be fixed at some point in the future.
261 Currently, tokenizers are encouraged to make sure that matching works against
262 both the search token list and the match token list.
265 FUNCTION token_get_postcode(info JSONB) RETURNS TEXT
268 Return the postcode for the object, if any exists. The postcode must be in
269 the form that should also be presented to the end-user.
272 FUNCTION token_strip_info(info JSONB) RETURNS JSONB
275 Return the part of the `token_info` field that should be stored in the database
276 permanently. The indexer calls this function when all processing is done and
277 replaces the content of the `token_info` column with the returned value before
278 the trigger stores the information in the database. May return NULL if no
279 information should be stored permanently.