Whoosh 1.x release notes

Whoosh 1.8.3

Whoosh 1.8.3 contains important bugfixes and new functionality. Thanks to all the mailing list and BitBucket users who helped with the fixes!

Fixed a bad Collector bug where the docset of a Results object did not match the actual results.

You can now pass a sequence of objects to a keyword argument in add_document and update_document (currently this will not work for unique fields in update_document). This is useful for non-text fields such as DATETIME and NUMERIC, allowing you to index multiple dates/numbers for a document:

writer.add_document(shoe=u"Saucony Kinvara", sizes=[10.0, 9.5, 12])

This version reverts to using the CDB hash function for hash files instead of Python’s hash() because the latter is not meant to be stored externally. This change maintains backwards compatibility with old files.

The Searcher.search method now takes a mask keyword argument. This is the opposite of the filter argument. Where the filter specifies the set of documents that can appear in the results, the mask specifies a set of documents that must not appear in the results.

Fixed performance problems in Searcher.more_like. This method now also takes a filter keyword argument like Searcher.search.

Improved documentation.

Whoosh 1.8.2

Whoosh 1.8.2 fixes some bugs, including a mistyped signature in Searcher.more_like and a bad bug in Collector that could screw up the ordering of results given certain parameters.

Whoosh 1.8.1

Whoosh 1.8.1 includes a few recent bugfixes/improvements:

  • ListMatcher.skip_to_quality() wasn’t returning an integer, resulting in a “None + int” error.
  • Fixed locking and memcache sync bugs in the Google App Engine storage object.
  • MultifieldPlugin wasn’t working correctly with groups.
    • The binary matcher trees of Or and And are now generated using a Huffman-like algorithm instead perfectly balanced. This gives a noticeable speed improvement because less information has to be passed up/down the tree.

Whoosh 1.8

This release relicensed the Whoosh source code under the Simplified BSD (A.K.A. “two-clause” or “FreeBSD”) license. See LICENSE.txt for more information.

Whoosh 1.7.7

Setting a TEXT field to store term vectors is now much easier. Instead of having to pass an instantiated whoosh.formats.Format object to the vector= keyword argument, you can pass True to automatically use the same format and analyzer as the inverted index. Alternatively, you can pass a Format subclass and Whoosh will instantiate it for you.

For example, to store term vectors using the same settings as the inverted index (Positions format and StandardAnalyzer):

from whoosh.fields import Schema, TEXT

schema = Schema(content=TEXT(vector=True))

To store term vectors that use the same analyzer as the inverted index (StandardAnalyzer by default) but only store term frequency:

from whoosh.formats import Frequency

schema = Schema(content=TEXT(vector=Frequency))

Note that currently the only place term vectors are used in Whoosh is keyword extraction/more like this, but they can be useful for expert users with custom code.

Added whoosh.searching.Searcher.more_like() and whoosh.searching.Hit.more_like_this() methods, as shortcuts for doing keyword extraction yourself. Return a Results object.

“python setup.py test” works again, as long as you have nose installed.

The whoosh.searching.Searcher.sort_query_using() method lets you sort documents matching a given query using an arbitrary function. Note that like “complex” searching with the Sorter object, this can be slow on large multi-segment indexes.

Whoosh 1.7

You can once again perform complex sorting of search results (that is, a sort with some fields ascending and some fields descending).

You can still use the sortedby keyword argument to whoosh.searching.Searcher.search() to do a simple sort (where all fields are sorted in the same direction), or you can use the new Sorter class to do a simple or complex sort:

searcher = myindex.searcher()
sorter = searcher.sorter()
# Sort first by the group field, ascending
# Then by the price field, descending
sorter.add_field("price", reverse=True)
# Get the Results
results = sorter.sort_query(myquery)

See the documentation for the Sorter class for more information. Bear in mind that complex sorts will be much slower on large indexes because they can’t use the per-segment field caches.

You can now get highlighted snippets for a hit automatically using whoosh.searching.Hit.highlights():

results = searcher.search(myquery, limit=20)
for hit in results:
    print hit["title"]
    print hit.highlights("content")

See whoosh.searching.Hit.highlights() for more information.

Added the ability to filter search results so that only hits in a Results set, a set of docnums, or matching a query are returned. The filter is cached on the searcher.

# Search within previous results newresults = searcher.search(newquery, filter=oldresults)

# Search within the “basics” chapter results = searcher.search(userquery, filter=query.Term(“chapter”, “basics”))

You can now specify a time limit for a search. If the search does not finish in the given time, a whoosh.searching.TimeLimit exception is raised, but you can still retrieve the partial results from the collector. See the timelimit and greedy arguments in the whoosh.searching.Collector documentation.

Added back the ability to set whoosh.analysis.StemFilter to use an unlimited cache. This is useful for one-shot batch indexing (see Tips for speeding up batch indexing).

The normalize() method of the And and Or queries now merges overlapping range queries for more efficient queries.

Query objects now have __hash__ methods allowing them to be used as dictionary keys.

The API of the highlight module has changed slightly. Most of the functions in the module have been converted to classes. However, most old code should still work. The NullFragmeter is now called WholeFragmenter, but the old name is still available as an alias.

Fixed MultiPool so it won’t fill up the temp directory with job files.

Fixed a bug where Phrase query objects did not use their boost factor.

Fixed a bug where a fieldname after an open parenthesis wasn’t parsed correctly. The change alters the semantics of certain parsing “corner cases” (such as a:b:c:d).

Whoosh 1.6

The whoosh.writing.BatchWriter class is now called whoosh.writing.BufferedWriter. It is similar to the old BatchWriter class but allows you to search and update the buffered documents as well as the documents that have been flushed to disk:

writer = writing.BufferedWriter(myindex)

# You can update (replace) documents in RAM without having to commit them
# to disk
writer.add_document(path="/a", text="Hi there")
writer.update_document(path="/a", text="Hello there")

# Search committed and uncommited documents by getting a searcher from the
# writer instead of the index
searcher = writer.searcher()

(BatchWriter is still available as an alias for backwards compatibility.)

The whoosh.qparser.QueryParser initialization method now requires a schema as the second argument. Previously the default was to create a QueryParser without a schema, which was confusing:

qp = qparser.QueryParser("content", myindex.schema)

The whoosh.searching.Searcher.search() method now takes a scored keyword. If you search with scored=False, the results will be in “natural” order (the order the documents were added to the index). This is useful when you don’t need scored results but want the convenience of the Results object.

Added the whoosh.qparser.GtLtPlugin parser plugin to allow greater than/less as an alternative syntax for ranges:

count:>100 tag:<=zebra date:>='29 march 2001'

Added the ability to define schemas declaratively, similar to Django models:

from whoosh import index
from whoosh.fields import SchemaClass, ID, KEYWORD, STORED, TEXT

class MySchema(SchemaClass):
    uuid = ID(stored=True, unique=True)
    path = STORED
    tags = KEYWORD(stored=True)
    content = TEXT

index.create_in("indexdir", MySchema)

Whoosh 1.6.2: Added whoosh.searching.TermTrackingCollector which tracks which part of the query matched which documents in the final results.

Replaced the unbounded cache in whoosh.analysis.StemFilter with a bounded LRU (least recently used) cache. This will make stemming analysis slightly slower but prevent it from eating up too much memory over time.

Added a simple whoosh.analysis.PyStemmerFilter that works when the py-stemmer library is installed:

ana = RegexTokenizer() | PyStemmerFilter("spanish")

The estimation of memory usage for the limitmb keyword argument to FileIndex.writer() is more accurate, which should help keep memory usage memory usage by the sorting pool closer to the limit.

The whoosh.ramdb package was removed and replaced with a single whoosh.ramindex module.

Miscellaneous bug fixes.

Whoosh 1.5


Whoosh 1.5 is incompatible with previous indexes. You must recreate existing indexes with Whoosh 1.5.

Fixed a bug where postings were not portable across different endian platforms.

New generalized field cache system, using per-reader caches, for much faster sorting and faceting of search results, as well as much faster multi-term (e.g. prefix and wildcard) and range queries, especially for large indexes and/or indexes with multiple segments.

Changed the faceting API. See Sorting and faceting.

Faster storage and retrieval of posting values.

Added per-field multitoken_query attribute to control how the query parser deals with a “term” that when analyzed generates multiple tokens. The default value is “first” which throws away all but the first token (the previous behavior). Other possible values are “and”, “or”, or “phrase”.

Added whoosh.analysis.DoubleMetaphoneFilter, whoosh.analysis.SubstitutionFilter, and whoosh.analysis.ShingleFilter.

Added whoosh.qparser.CopyFieldPlugin.

Added whoosh.query.Otherwise.

Generalized parsing of operators (such as OR, AND, NOT, etc.) in the query parser to make it easier to add new operators. In intend to add a better API for this in a future release.

Switched NUMERIC and DATETIME fields to use more compact on-disk representations of numbers.

Fixed a bug in the porter2 stemmer when stemming the string “y”.

Added methods to whoosh.searching.Hit to make it more like a dict.

Short posting lists (by default, single postings) are inline in the term file instead of written to the posting file for faster retrieval and a small saving in disk space.

Whoosh 1.3

Whoosh 1.3 adds a more efficient DATETIME field based on the new tiered NUMERIC field, and the DateParserPlugin. See Indexing and parsing dates/times.

Whoosh 1.2

Whoosh 1.2 adds tiered indexing for NUMERIC fields, resulting in much faster range queries on numeric fields.

Whoosh 1.0

Whoosh 1.0 is a major milestone release with vastly improved performance and several useful new features.

The index format of this version is not compatibile with indexes created by previous versions of Whoosh. You will need to reindex your data to use this version.

Orders of magnitude faster searches for common terms. Whoosh now uses optimizations similar to those in Xapian to skip reading low-scoring postings.

Faster indexing and ability to use multiple processors (via multiprocessing module) to speed up indexing.

Flexible Schema: you can now add and remove fields in an index with the whoosh.writing.IndexWriter.add_field() and whoosh.writing.IndexWriter.remove_field() methods.

New hand-written query parser based on plug-ins. Less brittle, more robust, more flexible, and easier to fix/improve than the old pyparsing-based parser.

On-disk formats now use 64-bit disk pointers allowing files larger than 4 GB.

New whoosh.searching.Facets class efficiently sorts results into facets based on any criteria that can be expressed as queries, for example tags or price ranges.

New whoosh.writing.BatchWriter class automatically batches up individual add_document and/or delete_document calls until a certain number of calls or a certain amount of time passes, then commits them all at once.

New whoosh.analysis.BiWordFilter lets you create bi-word indexed fields a possible alternative to phrase searching.

Fixed bug where files could be deleted before a reader could open them in threaded situations.

New whoosh.analysis.NgramFilter filter, whoosh.analysis.NgramWordAnalyzer analyzer, and whoosh.fields.NGRAMWORDS field type allow producing n-grams from tokenized text.

Errors in query parsing now raise a specific whoosh.qparse.QueryParserError exception instead of a generic exception.

Previously, the query string * was optimized to a whoosh.query.Every query which matched every document. Now the Every query only matches documents that actually have an indexed term from the given field, to better match the intuitive sense of what a query string like tag:* should do.

New whoosh.searching.Searcher.key_terms_from_text() method lets you extract key words from arbitrary text instead of documents in the index.

Previously the whoosh.searching.Searcher.key_terms() and whoosh.searching.Results.key_terms() methods required that the given field store term vectors. They now also work if the given field is stored instead. They will analyze the stored string into a term vector on-the-fly. The field must still be indexed.

User API changes

The default for the limit keyword argument to whoosh.searching.Searcher.search() is now 10. To return all results in a single Results object, use limit=None.

The Index object no longer represents a snapshot of the index at the time the object was instantiated. Instead it always represents the index in the abstract. Searcher and IndexReader objects obtained from the Index object still represent the index as it was at the time they were created.

Because the Index object no longer represents the index at a specific version, several methods such as up_to_date and refresh were removed from its interface. The Searcher object now has last_modified(), up_to_date(), and refresh() methods similar to those that used to be on Index.

The document deletion and field add/remove methods on the Index object now create a writer behind the scenes to accomplish each call. This means they write to the index immediately, so you don’t need to call commit on the Index. Also, it will be much faster if you need to call them multiple times to create your own writer instead:

# Don't do this
for id in my_list_of_ids_to_delete:
    myindex.delete_by_term("id", id)

# Instead do this
writer = myindex.writer()
for id in my_list_of_ids_to_delete:
    writer.delete_by_term("id", id)

The postlimit argument to Index.writer() has been changed to postlimitmb and is now expressed in megabytes instead of bytes:

writer = myindex.writer(postlimitmb=128)

Instead of having to import whoosh.filedb.filewriting.NO_MERGE or whoosh.filedb.filewriting.OPTIMIZE to use as arguments to commit(), you can now simply do the following:

# Do not merge segments

# or

# Merge all segments

The whoosh.postings module is gone. The whoosh.matching module contains classes for posting list readers.

Whoosh no longer maps field names to numbers for internal use or writing to disk. Any low-level method that accepted field numbers now accept field names instead.

Custom Weighting implementations that use the final() method must now set the use_final attribute to True:

from whoosh.scoring import BM25F

class MyWeighting(BM25F):
    use_final = True

    def final(searcher, docnum, score):
        return score + docnum * 10

This disables the new optimizations, forcing Whoosh to score every matching document.

whoosh.writing.AsyncWriter now takes an whoosh.index.Index object as its first argument, not a callable. Also, the keyword arguments to pass to the index’s writer() method should now be passed as a dictionary using the writerargs keyword argument.

Whoosh now stores per-document field length using an approximation rather than exactly. For low numbers the approximation is perfectly accurate, while high numbers will be approximated less accurately.

The doc_field_length method on searchers and readers now takes a second argument representing the default to return if the given document and field do not have a length (i.e. the field is not scored or the field was not provided for the given document).

The whoosh.analysis.StopFilter now has a maxsize argument as well as a minsize argument to its initializer. Analyzers that use the StopFilter have the maxsize argument in their initializers now also.

The interface of whoosh.writing.AsyncWriter has changed.


  • Because the file backend now writes 64-bit disk pointers and field names instead of numbers, the size of an index on disk will grow compared to previous versions.
  • Unit tests should no longer leave directories and files behind.