About analyzers


An analyzer is a function or callable class (a class with a __call__ method) that takes a unicode string and returns a generator of tokens. Usually a “token” is a word, for example the string “Mary had a little lamb” might yield the tokens “Mary”, “had”, “a”, “little”, and “lamb”. However, tokens do not necessarily correspond to words. For example, you might tokenize Chinese text into individual characters or bi-grams. Tokens are the units of indexing, that is, they are what you are able to look up in the index.

An analyzer is basically just a wrapper for a tokenizer and zero or more filters. The analyzer’s __call__ method will pass its parameters to a tokenizer, and the tokenizer will usually be wrapped in a few filters.

A tokenizer is a callable that takes a unicode string and yields a series of analysis.Token objects.

For example, the provided whoosh.analysis.RegexTokenizer class implements a customizable, regular-expression-based tokenizer that extracts words and ignores whitespace and punctuation.

>>> from whoosh.analysis import RegexTokenizer
>>> tokenizer = RegexTokenizer()
>>> for token in tokenizer(u"Hello there my friend!"):
...   print repr(token.text)

A filter is a callable that takes a generator of Tokens (either a tokenizer or another filter) and in turn yields a series of Tokens.

For example, the provided whoosh.analysis.LowercaseFilter() filters tokens by converting their text to lowercase. The implementation is very simple:

def LowercaseFilter(tokens):
    """Uses lower() to lowercase token text. For example, tokens
    "This","is","a","TEST" become "this","is","a","test".

    for t in tokens:
        t.text = t.text.lower()
        yield t

You can wrap the filter around a tokenizer to see it in operation:

>>> from whoosh.analysis import LowercaseFilter
>>> for token in LowercaseFilter(tokenizer(u"These ARE the things I want!")):
...   print repr(token.text)

An analyzer is just a means of combining a tokenizer and some filters into a single package.

You can implement an analyzer as a custom class or function, or compose tokenizers and filters together using the | character:

my_analyzer = RegexTokenizer() | LowercaseFilter() | StopFilter()

The first item must be a tokenizer and the rest must be filters (you can’t put a filter first or a tokenizer after the first item). Note that this only works if at least the tokenizer is a subclass of whoosh.analysis.Composable, as all the tokenizers and filters that ship with Whoosh are.

See the whoosh.analysis module for information on the available analyzers, tokenizers, and filters shipped with Whoosh.

Using analyzers

When you create a field in a schema, you can specify your analyzer as a keyword argument to the field object:

schema = Schema(content=TEXT(analyzer=StemmingAnalyzer()))

Advanced Analysis

Token objects

The Token class has no methods. It is merely a place to record certain attributes. A Token object actually has two kinds of attributes: settings that record what kind of information the Token object does or should contain, and information about the current token.

Token setting attributes

A Token object should always have the following attributes. A tokenizer or filter can check these attributes to see what kind of information is available and/or what kind of information they should be setting on the Token object.

These attributes are set by the tokenizer when it creates the Token(s), based on the parameters passed to it from the Analyzer.

Filters should not change the values of these attributes.

Type Attribute name Description Default
str mode The mode in which the analyzer is being called, e.g. ‘index’ during indexing or ‘query’ during query parsing ‘’
bool positions Whether term positions are recorded in the token False
bool chars Whether term start and end character indices are recorded in the token False
bool boosts Whether per-term boosts are recorded in the token False
bool removestops Whether stop-words should be removed from the token stream True

Token information attributes

A Token object may have any of the following attributes. The text attribute should always be present. The original attribute may be set by a tokenizer. All other attributes should only be accessed or set based on the values of the “settings” attributes above.

Type Name Description
unicode text The text of the token (this should always be present)
unicode original The original (pre-filtered) text of the token. The tokenizer may record this, and filters are expected not to modify it.
int pos The position of the token in the stream, starting at 0 (only set if positions is True)
int startchar The character index of the start of the token in the original string (only set if chars is True)
int endchar The character index of the end of the token in the original string (only set if chars is True)
float boost The boost for this token (only set if boosts is True)
bool stopped Whether this token is a “stop” word (only set if removestops is False)

So why are most of the information attributes optional? Different field formats require different levels of information about each token. For example, the Frequency format only needs the token text. The Positions format records term positions, so it needs them on the Token. The Characters format records term positions and the start and end character indices of each term, so it needs them on the token, and so on.

The Format object that represents the format of each field calls the analyzer for the field, and passes it parameters corresponding to the types of information it needs, e.g.:

analyzer(unicode_string, positions=True)

The analyzer can then pass that information to a tokenizer so the tokenizer initializes the required attributes on the Token object(s) it produces.

Performing different analysis for indexing and query parsing

Whoosh sets the mode setting attribute to indicate whether the analyzer is being called by the indexer (mode='index') or the query parser (mode='query'). This is useful if there’s a transformation that you only want to apply at indexing or query parsing:

class MyFilter(Filter):
    def __call__(self, tokens):
        for t in tokens:
            if t.mode == 'query':

The whoosh.analysis.MultiFilter filter class lets you specify different filters to use based on the mode setting:

intraword = MultiFilter(index=IntraWordFilter(mergewords=True, mergenums=True),
                        query=IntraWordFilter(mergewords=False, mergenums=False))

Stop words

“Stop” words are words that are so common it’s often counter-productive to index them, such as “and”, “or”, “if”, etc. The provided analysis.StopFilter lets you filter out stop words, and includes a default list of common stop words.

>>> from whoosh.analysis import StopFilter
>>> stopper = StopFilter()
>>> for token in stopper(LowercaseFilter(tokenizer(u"These ARE the things I want!"))):
...   print repr(token.text)

However, this seemingly simple filter idea raises a couple of minor but slightly thorny issues: renumbering term positions and keeping or removing stopped words.

Renumbering term positions

Remember that analyzers are sometimes asked to record the position of each token in the token stream:

Token.text u’Mary’ u’had’ u’a’ u’lamb’
Token.pos 0 1 2 3

So what happens to the pos attribute of the tokens if StopFilter removes the words had and a from the stream? Should it renumber the positions to pretend the “stopped” words never existed? I.e.:

Token.text u’Mary’ u’lamb’
Token.pos 0 1

or should it preserve the original positions of the words? I.e:

Token.text u’Mary’ u’lamb’
Token.pos 0 3

It turns out that different situations call for different solutions, so the provided StopFilter class supports both of the above behaviors. Renumbering is the default, since that is usually the most useful and is necessary to support phrase searching. However, you can set a parameter in StopFilter’s constructor to tell it not to renumber positions:

stopper = StopFilter(renumber=False)

Removing or leaving stop words

The point of using StopFilter is to remove stop words, right? Well, there are actually some situations where you might want to mark tokens as “stopped” but not remove them from the token stream.

For example, if you were writing your own query parser, you could run the user’s query through a field’s analyzer to break it into tokens. In that case, you might want to know which words were “stopped” so you can provide helpful feedback to the end user (e.g. “The following words are too common to search for:”).

In other cases, you might want to leave stopped words in the stream for certain filtering steps (for example, you might have a step that looks at previous tokens, and want the stopped tokens to be part of the process), but then remove them later.

The analysis module provides a couple of tools for keeping and removing stop-words in the stream.

The removestops parameter passed to the analyzer’s __call__ method (and copied to the Token object as an attribute) specifies whether stop words should be removed from the stream or left in.

>>> from whoosh.analysis import StandardAnalyzer
>>> analyzer = StandardAnalyzer()
>>> [(t.text, t.stopped) for t in analyzer(u"This is a test")]
[(u'test', False)]
>>> [(t.text, t.stopped) for t in analyzer(u"This is a test", removestops=False)]
[(u'this', True), (u'is', True), (u'a', True), (u'test', False)]

The analysis.unstopped() filter function takes a token generator and yields only the tokens whose stopped attribute is False.


Even if you leave stopped words in the stream in an analyzer you use for indexing, the indexer will ignore any tokens where the stopped attribute is True.

Implementation notes

Because object creation is slow in Python, the stock tokenizers do not create a new analysis.Token object for each token. Instead, they create one Token object and yield it over and over. This is a nice performance shortcut but can lead to strange behavior if your code tries to remember tokens between loops of the generator.

Because the analyzer only has one Token object, of which it keeps changing the attributes, if you keep a copy of the Token you get from a loop of the generator, it will be changed from under you. For example:

>>> list(tokenizer(u"Hello there my friend"))
[Token(u"friend"), Token(u"friend"), Token(u"friend"), Token(u"friend")]

Instead, do this:

>>> [t.text for t in tokenizer(u"Hello there my friend")]

That is, save the attributes, not the token object itself.

If you implement your own tokenizer, filter, or analyzer as a class, you should implement an __eq__ method. This is important to allow comparison of Schema objects.

The mixing of persistent “setting” and transient “information” attributes on the Token object is not especially elegant. If I ever have a better idea I might change it. ;) Nothing requires that an Analyzer be implemented by calling a tokenizer and filters. Tokenizers and filters are simply a convenient way to structure the code. You’re free to write an analyzer any way you want, as long as it implements __call__.