Sorting and faceting

Note

The API for sorting and faceting changed in Whoosh 3.0.

Overview

Sorting and faceting search results in Whoosh is based on facets. Each facet associates a value with each document in the search results, allowing you to sort by the keys or use them to group the documents. Whoosh includes a variety of facet types you can use for sorting and grouping (see below).

Sorting

By default, the results of a search are sorted with the highest-scoring documents first. You can use the sortedby keyword argument to order the results by some other criteria instead, such as the value of a field.

Making fields sortable

In order to sort on a field, you should create the field using the sortable=True keyword argument:

schema = fields.Schema(title=fields.TEXT(sortable=True),
                       content=fields.TEXT,
                       modified=fields.DATETIME(sortable=True)
                       )

It’s possible to sort on a field that doesn’t have sortable=True, but this requires Whoosh to load the unique terms in the field into memory. Using sortable is much more efficient.

About column types

When you create a field using sortable=True, you are telling Whoosh to store per-document values for that field in a column. A column object specifies the format to use to store the per-document values on disk.

The whoosh.columns module contains several different column object implementations. Each field type specifies a reasonable default column type (for example, the default for text fields is whoosh.columns.VarBytesColumn, the default for numeric fields is whoosh.columns.NumericColumn). However, if you want maximum efficiency you may want to use a different column type for a field.

For example, if all document values in a field are a fixed length, you can use a whoosh.columns.FixedBytesColumn. If you have a field where many documents share a relatively small number of possible values (an example might be a “category” field, or “month” or other enumeration type fields), you might want to use whoosh.columns.RefBytesColumn (which can handle both variable and fixed-length values). There are column types for storing per-document bit values, structs, pickled objects, and compressed byte values.

To specify a custom column object for a field, pass it as the sortable keyword argument instead of True:

from whoosh import columns, fields

category_col = columns.RefBytesColumn()
schema = fields.Schema(title=fields.TEXT(sortable=True),
                       category=fields.KEYWORD(sortable=category_col)

Using a COLUMN field for custom sort keys

When you add a document with a sortable field, Whoosh uses the value you pass for the field as the sortable value. For example, if “title” is a sortable field, and you add this document:

writer.add_document(title="Mr. Palomar")

...then Mr. Palomar is stored in the field column as the sorting key for the document.

This is usually good, but sometimes you need to “massage” the sortable key so it’s different from the value the user searches and/or sees in the interface. For example, if you allow the user to sort by title, you might want to use different values for the visible title and the value used for sorting:

# Visible title
title = "The Unbearable Lightness of Being"

# Sortable title: converted to lowercase (to prevent different ordering
# depending on uppercase/lowercase), with initial article moved to the end
sort_title = "unbearable lightness of being, the"

The best way to do this is to use an additional field just for sorting. You can use the whoosh.fields.COLUMN field type to create a field that is not indexed or stored, it only holds per-document column values:

schema = fields.Schema(title=fields.TEXT(stored=True),
                       sort_title=fields.COLUMN(columns.VarBytesColumn())
                       )

The single argument to the whoosh.fields.COLUMN initializer is a whoosh.columns.ColumnType object. You can use any of the various column types in the whoosh.columns module.

As another example, say you are indexing documents that have a custom sorting order associated with each document, such as a “priority” number:

name=Big Wheel
price=100
priority=1

name=Toss Across
price=40
priority=3

name=Slinky
price=25
priority=2
...

You can use a column field with a numeric column object to hold the “priority” and use it for sorting:

schema = fields.Schema(name=fields.TEXT(stored=True),
                       price=fields.NUMERIC(stored=True),
                       priority=fields.COLUMN(columns.NumericColumn("i"),
                       )

(Note that columns.NumericColumn takes a type code character like the codes used by Python’s struct and array modules.)

Making existing fields sortable

If you have an existing index from before the sortable argument was added in Whoosh 3.0, or you didn’t think you needed a field to be sortable but now you find that you need to sort it, you can add “sortability” to an existing index using the whoosh.sorting.add_sortable() utility function:

from whoosh import columns, fields, index, sorting

# Say we have an existing index with this schema
schema = fields.Schema(title=fields.TEXT,
                       price=fields.NUMERIC)

# To use add_sortable, first open a writer for the index
ix = index.open_dir("indexdir")
with ix.writer() as w:
    # Add sortable=True to the "price" field using field terms as the
    # sortable values
    sorting.add_sortable(w, "price", sorting.FieldFacet("price"))

    # Add sortable=True to the "title" field using the
    # stored field values as the sortable value
    sorting.add_sortable(w, "title", sorting.StoredFieldFacet("title"))

You can specify a custom column type when you call add_sortable using the column keyword argument:

add_sortable(w, "chapter", sorting.FieldFacet("chapter"),
             column=columns.RefBytesColumn())

See the documentation for add_sortable() for more information.

Sorting search results

When you tell Whoosh to sort by a field (or fields), it uses the per-document values in the field’s column as sorting keys for the documents.

Normally search results are sorted by descending relevance score. You can tell Whoosh to use a different ordering by passing the sortedby keyword argument to the search() method:

from whoosh import fields, index, qparser

schema = fields.Schema(title=fields.TEXT(stored=True),
                       price=fields.NUMERIC(sortable=True))
ix = index.create_in("indexdir", schema)

with ix.writer() as w:
    w.add_document(title="Big Deal", price=20)
    w.add_document(title="Mr. Big", price=10)
    w.add_document(title="Big Top", price=15)

with ix.searcher() as s:
    qp = qparser.QueryParser("big", ix.schema)
    q = qp.parse(user_query_string)

    # Sort search results from lowest to highest price
    results = s.search(q, sortedby="price")
    for hit in results:
        print(hit["title"])

You can use any of the following objects as sortedby values:

A FacetType object
Uses this object to sort the documents. See below for the available facet types.
A field name string
Converts the field name into a FieldFacet (see below) and uses it to sort the documents.
A list of FacetType objects and/or field name strings
Bundles the facets together into a MultiFacet so you can sort by multiple keys. Note that this shortcut does not allow you to reverse the sort direction of individual facets. To do that, you need to construct the MultiFacet object yourself.

Note

You can use the reverse=True keyword argument to the Searcher.search() method to reverse the overall sort direction. This is more efficient than reversing each individual facet.

Examples

Sort by the value of the size field:

results = searcher.search(myquery, sortedby="size")

Sort by the reverse (highest-to-lowest) order of the “price” field:

facet = sorting.FieldFacet("price", reverse=True)
results = searcher.search(myquery, sortedby=facet)

Sort by ascending size and then descending price:

mf = sorting.MultiFacet()
mf.add_field("size")
mf.add_field("price", reverse=True)
results = searcher.search(myquery, sortedby=mf)

# or...
sizes = sorting.FieldFacet("size")
prices = sorting.FieldFacet("price", reverse=True)
results = searcher.search(myquery, sortedby=[sizes, prices])

Sort by the “category” field, then by the document’s score:

cats = sorting.FieldFacet("category")
scores = sorting.ScoreFacet()
results = searcher.search(myquery, sortedby=[cats, scores])

Accessing column values

Per-document column values are available in Hit objects just like stored field values:

schema = fields.Schema(title=fields.TEXT(stored=True),
                       price=fields.NUMERIC(sortable=True))

...

results = searcher.search(myquery)
for hit in results:
    print(hit["title"], hit["price"])

ADVANCED: if you want to access abitrary per-document values quickly you can get a column reader object:

with ix.searcher() as s:
    reader = s.reader()

    colreader = s.reader().column_reader("price")
    for docnum in reader.all_doc_ids():
        print(colreader[docnum])

Grouping

It is often very useful to present “faceted” search results to the user. Faceting is dynamic grouping of search results into categories. The categories let users view a slice of the total results based on the categories they’re interested in.

For example, if you are programming a shopping website, you might want to display categories with the search results such as the manufacturers and price ranges.

Manufacturer Price
Apple (5) $0 - $100 (2)
Sanyo (1) $101 - $500 (10)
Sony (2) $501 - $1000 (1)
Toshiba (5)  

You can let your users click the different facet values to only show results in the given categories.

Another useful UI pattern is to show, say, the top 5 results for different types of found documents, and let the user click to see more results from a category they’re interested in, similarly to how the Spotlight quick results work on Mac OS X.

The groupedby keyword argument

You can use the following objects as groupedby values:

A FacetType object
Uses this object to group the documents. See below for the available facet types.
A field name string
Converts the field name into a FieldFacet (see below) and uses it to sort the documents. The name of the field is used as the facet name.
A list or tuple of field name strings
Sets up multiple field grouping criteria.
A dictionary mapping facet names to FacetType objects
Sets up multiple grouping criteria.
A Facets object
This object is a lot like using a dictionary, but has some convenience methods to make setting up multiple groupings a little easier.

Examples

Group by the value of the “category” field:

results = searcher.search(myquery, groupedby="category")

Group by the value of the “category” field and also by the value of the “tags” field and a date range:

cats = sorting.FieldFacet("category")
tags = sorting.FieldFacet("tags", allow_overlap=True)
results = searcher.search(myquery, groupedby={"category": cats, "tags": tags})

# ...or, using a Facets object has a little less duplication
facets = sorting.Facets()
facets.add_field("category")
facets.add_field("tags", allow_overlap=True)
results = searcher.search(myquery, groupedby=facets)

To group results by the intersected values of multiple fields, use a MultiFacet object (see below). For example, if you have two fields named tag and size, you could group the results by all combinations of the tag and size field, such as ('tag1', 'small'), ('tag2', 'small'), ('tag1', 'medium'), and so on:

# Generate a grouping from the combination of the "tag" and "size" fields
mf = MultiFacet("tag", "size")
results = searcher.search(myquery, groupedby={"tag/size": mf})

Getting the faceted groups

The Results.groups("facetname") method returns a dictionary mapping category names to lists of document IDs:

myfacets = sorting.Facets().add_field("size").add_field("tag")
results = mysearcher.search(myquery, groupedby=myfacets)
results.groups("size")
# {"small": [8, 5, 1, 2, 4], "medium": [3, 0, 6], "large": [7, 9]}

If there is only one facet, you can just use Results.groups() with no argument to access its groups:

results = mysearcher.search(myquery, groupedby=myfunctionfacet)
results.groups()

By default, the values in the dictionary returned by groups() are lists of document numbers in the same relative order as in the results. You can use the Searcher object’s stored_fields() method to take a document number and return the document’s stored fields as a dictionary:

for category_name in categories:
    print "Top 5 documents in the %s category" % category_name
    doclist = categories[category_name]
    for docnum, score in doclist[:5]:
        print "  ", searcher.stored_fields(docnum)
    if len(doclist) > 5:
        print "  (%s more)" % (len(doclist) - 5)

If you want different information about the groups, for example just the count of documents in each group, or you don’t need the groups to be ordered, you can specify a whoosh.sorting.FacetMap type or instance with the maptype keyword argument when creating the FacetType:

# This is the same as the default
myfacet = FieldFacet("size", maptype=sorting.OrderedList)
results = mysearcher.search(myquery, groupedby=myfacet)
results.groups()
# {"small": [8, 5, 1, 2, 4], "medium": [3, 0, 6], "large": [7, 9]}

# Don't sort the groups to match the order of documents in the results
# (faster)
myfacet = FieldFacet("size", maptype=sorting.UnorderedList)
results = mysearcher.search(myquery, groupedby=myfacet)
results.groups()
# {"small": [1, 2, 4, 5, 8], "medium": [0, 3, 6], "large": [7, 9]}

# Only count the documents in each group
myfacet = FieldFacet("size", maptype=sorting.Count)
results = mysearcher.search(myquery, groupedby=myfacet)
results.groups()
# {"small": 5, "medium": 3, "large": 2}

# Only remember the "best" document in each group
myfacet = FieldFacet("size", maptype=sorting.Best)
results = mysearcher.search(myquery, groupedby=myfacet)
results.groups()
# {"small": 8, "medium": 3, "large": 7}

Alternatively you can specify a maptype argument in the Searcher.search() method call which applies to all facets:

results = mysearcher.search(myquery, groupedby=["size", "tag"],
                            maptype=sorting.Count)

(You can override this overall maptype argument on individual facets by specifying the maptype argument for them as well.)

Facet types

FieldFacet

This is the most common facet type. It sorts or groups based on the value in a certain field in each document. This generally works best (or at all) if each document has only one term in the field (e.g. an ID field):

# Sort search results by the value of the "path" field
facet = sorting.FieldFacet("path")
results = searcher.search(myquery, sortedby=facet)

# Group search results by the value of the "parent" field
facet = sorting.FieldFacet("parent")
results = searcher.search(myquery, groupedby=facet)
parent_groups = results.groups("parent")

By default, FieldFacet only supports non-overlapping grouping, where a document cannot belong to multiple facets at the same time (each document will be sorted into one category arbitrarily.) To get overlapping groups with multi-valued fields, use the allow_overlap=True keyword argument:

facet = sorting.FieldFacet(fieldname, allow_overlap=True)

This supports overlapping group membership where documents have more than one term in a field (e.g. KEYWORD fields). If you don’t need overlapping, don’t use allow_overlap because it’s much slower and uses more memory (see the secion on allow_overlap below).

QueryFacet

You can set up categories defined by arbitrary queries. For example, you can group names using prefix queries:

# Use queries to define each category
# (Here I'll assume "price" is a NUMERIC field, so I'll use
# NumericRange)
qdict = {}
qdict["A-D"] = query.TermRange("name", "a", "d")
qdict["E-H"] = query.TermRange("name", "e", "h")
qdict["I-L"] = query.TermRange("name", "i", "l")
# ...

qfacet = sorting.QueryFacet(qdict)
r = searcher.search(myquery, groupedby={"firstltr": qfacet})

By default, QueryFacet only supports non-overlapping grouping, where a document cannot belong to multiple facets at the same time (each document will be sorted into one category arbitrarily). To get overlapping groups with multi-valued fields, use the allow_overlap=True keyword argument:

facet = sorting.QueryFacet(querydict, allow_overlap=True)

RangeFacet

The RangeFacet is for NUMERIC field types. It divides a range of possible values into groups. For example, to group documents based on price into buckets $100 “wide”:

pricefacet = sorting.RangeFacet("price", 0, 1000, 100)

The first argument is the name of the field. The next two arguments are the full range to be divided. Value outside this range (in this example, values below 0 and above 1000) will be sorted into the “missing” (None) group. The fourth argument is the “gap size”, the size of the divisions in the range.

The “gap” can be a list instead of a single value. In that case, the values in the list will be used to set the size of the initial divisions, with the last value in the list being the size for all subsequent divisions. For example:

pricefacet = sorting.RangeFacet("price", 0, 1000, [5, 10, 35, 50])

...will set up divisions of 0-5, 5-15, 15-50, 50-100, and then use 50 as the size for all subsequent divisions (i.e. 100-150, 150-200, and so on).

The hardend keyword argument controls whether the last division is clamped to the end of the range or allowed to go past the end of the range. For example, this:

facet = sorting.RangeFacet("num", 0, 10, 4, hardend=False)

...gives divisions 0-4, 4-8, and 8-12, while this:

facet = sorting.RangeFacet("num", 0, 10, 4, hardend=True)

...gives divisions 0-4, 4-8, and 8-10. (The default is hardend=False.)

Note

The ranges/buckets are always inclusive at the start and exclusive at the end.

DateRangeFacet

This is like RangeFacet but for DATETIME fields. The start and end values must be datetime.datetime objects, and the gap(s) is/are datetime.timedelta objects.

For example:

from datetime import datetime, timedelta

start = datetime(2000, 1, 1)
end = datetime.now()
gap = timedelta(days=365)
bdayfacet = sorting.DateRangeFacet("birthday", start, end, gap)

As with RangeFacet, you can use a list of gaps and the hardend keyword argument.

ScoreFacet

This facet is sometimes useful for sorting.

For example, to sort by the “category” field, then for documents with the same category, sort by the document’s score:

cats = sorting.FieldFacet("category")
scores = sorting.ScoreFacet()
results = searcher.search(myquery, sortedby=[cats, scores])

The ScoreFacet always sorts higher scores before lower scores.

Note

While using sortedby=ScoreFacet() should give the same results as using the default scored ordering (sortedby=None), using the facet will be slower because Whoosh automatically turns off many optimizations when sorting.

FunctionFacet

This facet lets you pass a custom function to compute the sorting/grouping key for documents. (Using this facet type may be easier than subclassing FacetType and Categorizer to set up some custom behavior.)

The function will be called with the index searcher and index document ID as arguments. For example, if you have an index with term vectors:

schema = fields.Schema(id=fields.STORED,
                       text=fields.TEXT(stored=True, vector=True))
ix = RamStorage().create_index(schema)

...you could use a function to sort documents higher the closer they are to having equal occurances of two terms:

def fn(searcher, docnum):
    v = dict(searcher.vector_as("frequency", docnum, "text"))
    # Sort documents that have equal number of "alfa" and "bravo" first
    return 0 - (1.0 / (abs(v.get("alfa", 0) - v.get("bravo", 0)) + 1.0))

facet = sorting.FunctionFacet(fn)
results = searcher.search(myquery, sortedby=facet)

StoredFieldFacet

This facet lets you use stored field values as the sorting/grouping key for documents. This is usually slower than using an indexed field, but when using allow_overlap it can actually be faster for large indexes just because it avoids the overhead of reading posting lists.

StoredFieldFacet supports allow_overlap by splitting the stored value into separate keys. By default it calls the value’s split() method (since most stored values are strings), but you can supply a custom split function. See the section on allow_overlap below.

MultiFacet

This facet type returns a composite of the keys returned by two or more sub-facets, allowing you to sort/group by the intersected values of multiple facets.

MultiFacet has methods for adding facets:

myfacet = sorting.RangeFacet(0, 1000, 10)

mf = sorting.MultiFacet()
mf.add_field("category")
mf.add_field("price", reverse=True)
mf.add_facet(myfacet)
mf.add_score()

You can also pass a list of field names and/or FacetType objects to the initializer:

prices = sorting.FieldFacet("price", reverse=True)
scores = sorting.ScoreFacet()
mf = sorting.MultiFacet("category", prices, myfacet, scores)

Missing values

  • When sorting, documents without any terms in a given field, or whatever else constitutes “missing” for different facet types, will always sort to the end.
  • When grouping, “missing” documents will appear in a group with the key None.

Using overlapping groups

The common supported workflow for grouping and sorting is where the given field has one value for document, for example a path field containing the file path of the original document. By default, facets are set up to support this single-value approach.

Of course, there are situations where you want documents to be sorted into multiple groups based on a field with multiple terms per document. The most common example would be a tags field. The allow_overlap keyword argument to the FieldFacet, QueryFacet, and StoredFieldFacet allows this multi-value approach.

However, there is an important caveat: using allow_overlap=True is slower than the default, potentially much slower for very large result sets. This is because Whoosh must read every posting of every term in the field to create a temporary “forward index” mapping documents to terms.

If a field is indexed with term vectors, FieldFacet will use them to speed up allow_overlap faceting for small result sets, but for large result sets, where Whoosh has to open the vector list for every matched document, this can still be very slow.

For very large indexes and result sets, if a field is stored, you can get faster overlapped faceting using StoredFieldFacet instead of FieldFacet. While reading stored values is usually slower than using the index, in this case avoiding the overhead of opening large numbers of posting readers can make it worthwhile.

StoredFieldFacet supports allow_overlap by loading the stored value for the given field and splitting it into multiple values. The default is to call the value’s split() method.

For example, if you’ve stored the tags field as a string like "tag1 tag2 tag3":

schema = fields.Schema(name=fields.TEXT(stored=True),
                       tags=fields.KEYWORD(stored=True))
ix = index.create_in("indexdir")
with ix.writer() as w:
    w.add_document(name="A Midsummer Night's Dream", tags="comedy fairies")
    w.add_document(name="Hamlet", tags="tragedy denmark")
    # etc.

...Then you can use a StoredFieldFacet like this:

ix = index.open_dir("indexdir")
with ix.searcher() as s:
    sff = sorting.StoredFieldFacet("tags", allow_overlap=True)
    results = s.search(myquery, groupedby={"tags": sff})

For stored Python objects other than strings, you can supply a split function (using the split_fn keyword argument to StoredFieldFacet). The function should accept a single argument (the stored value) and return a list or tuple of grouping keys.

Using a custom sort order

It is sometimes useful to have a custom sort order per-search. For example, different languages use different sort orders. If you have a function to return the sorting order you want for a given field value, such as an implementation of the Unicode Collation Algorithm (UCA), you can customize the sort order for the user’s language.

The whoosh.sorting.TranslateFacet lets you apply a function to the value of another facet. This lets you “translate” a field value into an arbitrary sort key, such as with UCA:

from pyuca import Collator

# The Collator object has a sort_key() method which takes a unicode
# string and returns a sort key
c = Collator("allkeys.txt")

# Make a facet object for the field you want to sort on
nf = sorting.FieldFacet("name")

# Wrap the facet in a TranslateFacet with the translation function
# (the Collator object's sort_key method)
tf = sorting.TranslateFacet(facet, c.sort_key)

# Use the facet to sort the search results
results = searcher.search(myquery, sortedby=tf)

(You can pass multiple “wrapped” facets to the TranslateFacet, and it will call the function with the values of the facets as multiple arguments.)

The TranslateFacet can also be very useful with numeric fields to sort on the output of some formula:

# Sort based on the average of two numeric fields
def average(a, b):
    return (a + b) / 2.0

# Create two facets for the fields and pass them with the function to
# TranslateFacet
af = sorting.FieldFacet("age")
wf = sorting.FieldFacet("weight")
facet = sorting.TranslateFacet(average, af, wf)

results = searcher.search(myquery. sortedby=facet)

Remember that you can still sort by multiple facets. For example, you could sort by a numeric value transformed by a quantizing function first, and then if that is equal sort by the value of another field:

# Sort by a quantized size first, then by name
tf = sorting.TranslateFacet(quantize, sorting.FieldFacet("size"))
results = searcher.search(myquery, sortedby=[tf, "name"])

Expert: writing your own facet

TBD.