Whoosh 2.x release notes¶
- Removed on-disk word graph implementation of spell checking in favor of much simpler and faster FSA implementation over the term file.
- Many bug fixes.
- Removed backwards compatibility with indexes created by versions prior to 2.5. You may need to re-index if you are using an old index that hasn’t been updated.
- This is the last 2.x release before a major overhaul that will break backwards compatibility.
Whoosh 2.5 will read existing indexes, but segments created by 2.5 will not be readable by older versions of Whoosh.
As a replacement for field caches to speed up sorting, Whoosh now supports adding a
sortable=Truekeyword argument to fields. This makes Whoosh store a sortable representation of the field’s values in a “column” format (which associates a “key” value with each document). This is more robust, efficient, and customizable than the old behavior. You should now specify
sortable=Trueon fields that you plan on using to sort or group search results.
(You can still sort/group on fields that don’t have
sortable=True, however it will use more RAM and be slower as Whoosh caches the field values in memory.)
Fields that use
sortable=Truecan avoid specifying
stored=True. The field’s value will still be available on
Hitobjects (the value will be retrieved from the column instead of from the stored fields). This may actually be faster for certain types of values.
Whoosh will now detect common types of OR queries and use optimized read-ahead matchers to speed them up by several times.
Whoosh now includes pure-Python implementations of the Snowball stemmers and stop word lists for various languages adapted from NLTK. These are available through the
whoosh.analysis.LanguageAnalyzeranalyzer or through the
lang=keyword argument to the
You can now use the
whoosh.filedb.filestore.Storage.destory()methods as a consistent API to set up and tear down different types of storage.
Many bug fixes and speed improvements.
Switched unit tests to use
- By default, Whoosh now assembles the individual files of a segment into a single file when committing. This has a small performance penalty but solves a problem where Whoosh can keep too many files open. Whoosh is also now smarter about using mmap.
- Added functionality to index and search hierarchical documents. See Indexing and searching document hierarchies.
- Rewrote the Directed Acyclic Word Graph implementation (used in spell checking) to be faster and more space-efficient. Word graph files created by previous versions will be ignored, meaning that spell checking may become slower unless/until you replace the old segments (for example, by optimizing).
- Rewrote multiprocessing indexing to be faster and simpler. You can now
myindex.writer(procs=n)to get a multiprocessing writer, or
myindex.writer(procs=n, multisegment=True)to get a multiprocessing writer that leaves behind multiple segments, like the old MultiSegmentWriter. (
MultiSegmentWriteris still available as a function that returns the new class.)
- When creating
Termquery objects for special fields (e.g. NUMERIC or BOOLEAN), you can now use the field’s literal type instead of a string as the second argument, for example
Term("bool", True). (This change may cause problems interacting with functions that expect query objects to be pure textual, such as spell checking.)
- All writing to and reading from on-disk indexes is now done through “codec” objects. This architecture should make it easier to add optional or experimental features, and maintain backwards compatibility.
- Fixes issues #75, #137, #206, #213, #215, #219, #223, #226, #230, #233, #238, #239, #240, #241, #243, #244, #245, #252, #253, and other bugs. Thanks to Thomas Waldmann and Alexei Gousev for the help!
- Fixes bug in BM25F scoring function, leading to increased precision in search results.
- Fixes issues #203, #205, #206, #208, #209, #212.
- Fixes issue #200.
- Added a
whoosh.query.Regexterm query type, similar to
whoosh.query.Wildcard. The parser does not allow regex term queries by default. You need to add the
whoosh.qparser.RegexPluginplugin. After you add the plugin, you can use
r"expression"query syntax for regular expression term queries. For example,
- Added the
whoosh.qparser.PseudoFieldPluginparser plugin. This plugin lets you create “pseudo-fields” that run a transform function on whatever query syntax the user applies the field to. This is fairly advanced functionality right now; I’m trying to think of ways to make its power easier to access.
- The documents in the lists in the dictionary returned by
Results.groups()by default are now in the same relative order as in the results. This makes it much easier to display the “top N” results in each category, for example.
groupidskeyword argument to
Searcher.searchhas been removed. Instead you can now pass a
whoosh.sorting.FacetMapobject to the
maptypeargument to control how faceted documents are grouped, and/or set the
maptypeargument on individual
whoosh.sorting.FacetType`objects to set custom grouping per facet. See Sorting and faceting for more information.
Searcher.document_numbers()with no arguments now yields all documents/numbers.
Writer.update_document()with no unique fields is now equivalent to calling
Writer.add_document()with the same arguments.
- Fixed a problem with keyword expansion where the code was building a cache that was fast on small indexes, but unacceptably slow on large indexes.
- Added the hyphen (
-) to the list of characters that match a “wildcard” token, to make parsing slightly more predictable. A true fix will have to wait for another parser rewrite.
- Fixed an unused
__future__import and use of
float("nan")which were breaking under Python 2.5.
- Fixed a bug where vectored fields with only one term stored an empty term vector.
- Various other bug fixes.
- Fixes several bugs, including a bad bug in BM25F scoring.
add_document(), You can now pass query-like strings for BOOLEAN and DATETIME fields (e.g
dtfield="20101131-16:01") as an alternative to actual
datetimeobjects. The implementation of this is incomplete: it only works in the default
filedbbackend, and if the field is stored, the stored value will be the string, not the parsed object.
This release fixes several bugs, and contains speed improvments to highlighting. See How to create highlighted search result excerpts for more information.
Whoosh is now compatible with Python 3 (tested with Python 3.2). Special thanks to Vinay Sajip who did the work, and also Jordan Sherer who helped fix later issues.
Sorting and grouping (faceting) now use a new system of “facet” objects which are much more flexible than the previous field-based system.
For example, to sort by first name and then score:
from whoosh import sorting mf = sorting.MultiFacet([sorting.FieldFacet("firstname"), sorting.ScoreFacet()]) results = searcher.search(myquery, sortedby=mf)
In addition to the previously supported sorting/grouping by field contents and/or query results, you can now use numeric ranges, date ranges, score, and more. The new faceting system also supports overlapping groups.
(The old “Sorter” API still works but is deprecated and may be removed in a future version.)
See Sorting and faceting for more information.
Completely revamped spell-checking to make it much faster, easier, and more flexible. You can enable generation of the graph files use by spell checking using the
spelling=Trueargument to a field type:
schema = fields.Schema(text=fields.TEXT(spelling=True))
(Spelling suggestion methods will work on fields without
spelling=Truebut will slower.) The spelling graph will be updated automatically as new documents are added – it is no longer necessary to maintain a separate “spelling index”.
You can get suggestions for individual words using
suglist = searcher.suggest("content", "werd", limit=3)
Whoosh now includes convenience methods to spell-check and correct user queries, with optional highlighting of corrections using the
from whoosh import highlight, qparser # User query string qstring = request.get("q") # Parse into query object parser = qparser.QueryParser("content", myindex.schema) qobject = parser.parse(qstring) results = searcher.search(qobject) if not results: correction = searcher.correct_query(gobject, gstring) # correction.query = corrected query object # correction.string = corrected query string # Format the corrected query string with HTML highlighting cstring = correction.format_string(highlight.HtmlFormatter())
Spelling suggestions can come from field contents and/or lists of words. For stemmed fields the spelling suggestions automatically use the unstemmed forms of the words.
There are APIs for spelling suggestions and query correction, so highly motivated users could conceivably replace the defaults with more sophisticated behaviors (for example, to take context into account).
See “Did you mean… ?” Correcting errors in user queries for more information.
whoosh.query.FuzzyTermnow uses the new word graph feature as well and so is much faster.
You can now set a boost factor for individual documents as you index them, to increase the score of terms in those documents in searches. See the documentation for the
add_document()for more information.
Added built-in recording of which terms matched in which documents. Use the
whoosh.searching.Hit.contains_term()to check matched terms.
Whoosh now supports whole-term quality optimizations, so for example if the system knows that a UnionMatcher cannot possibly contribute to the “top N” results unless both sub-matchers match, it will replace the UnionMatcher with an IntersectionMatcher which is faster to compute. The performance improvement is not as dramatic as from block quality optimizations, but it can be noticeable.
Fixed a bug that prevented block quality optimizations in queries with words not in the index, which could severely degrade performance.
Block quality optimizations now use the actual scoring algorithm to calculate block quality instead of an approximation, which fixes issues where ordering of results could be different for searches with and without the optimizations.
the BOOLEAN field type now supports field boosts.
Re-architected the query parser to make the code easier to understand. Custom parser plugins from previous versions will probably break in Whoosh 2.0.
Various bug-fixes and performance improvements.
Removed the “read lock”, which caused more problems than it solved. Now when opening a reader, if segments are deleted out from under the reader as it is opened, the code simply retries.
The term quality optimizations required changes to the on-disk formats. Whoosh 2.0 if backwards-compatible with the old format. As you rewrite an index using Whoosh 2.0, by default it will use the new formats for new segments, making the index incompatible with older versions.
To upgrade an existing index to use the new formats immediately, use
Removed the experimental
TermTrackingCollectorsince it is replaced by the new built-in term recording functionality.
Removed the experimental
Searcher.define_facetsfeature until a future release when it will be replaced by a more robust and useful feature.
Reader iteration methods (
iter_field, etc.) now yield
The arguments to