Cookbook¶
This page collects short, runnable recipes for common search tasks. Each one maps to a self-contained script in the examples/ directory of the repository, so you can run it end to end:
git clone https://github.com/priya-sundaram-dev/whoosh
cd whoosh
pip install -e .
python examples/quickstart.py
All recipes use only the standard library plus Whoosh itself — there are no extra runtime dependencies.
Quick start¶
examples/quickstart.py — the shortest path from an empty directory to a
working search: define a Schema, add a couple of
documents, parse a user query with QueryParser, and
print results with keyword highlighting.
A guided tour¶
examples/tutorial.py — a longer, commented walk-through that builds a small
product catalogue in an in-memory index. It covers:
an in-memory index with
RamStorageupserting records with
writer.update_documentsingle-field and multi-field parsing (
QueryParser/MultifieldParser)combining a parsed query with an exact-term filter
sorting results by a numeric field
grouping results with a
FieldFacethighlighting matched terms in stored text
The prose version of the same material lives in TUTORIAL.md at the repo root.
“Did you mean …?” spelling correction¶
examples/did_you_mean.py — Whoosh has spelling correction built in, with no
external dependencies. The recipe shows both single-word suggestions via
searcher.suggest and whole-query correction via
searcher.correct_query, and prints both a plain-text and an HTML
“did you mean” prompt. See also “Did you mean… ?” Correcting errors in user queries.
Autocomplete / search-as-you-type¶
examples/autocomplete.py — three pure-Python approaches to
search-as-you-type:
term completion with
reader.expand_prefixprefix matching with
whoosh.query.Prefix“matches anywhere” fuzzy completion with an n-gram analyzer (see Indexing and searching N-grams)
Pick the one that fits your latency and index-size budget.
Whoosh vs. SQLite FTS5¶
examples/benchmark_vs_sqlite.py — an honest, reproducible micro-benchmark
comparing build time, on-disk size, and average query latency against SQLite’s
FTS5 extension over the same corpus. Use it to decide when Whoosh’s pure-Python,
zero-dependency, deeply programmable model is the right trade-off for your
project, and when an embedded C engine is a better fit.
Highlighting and snippets¶
examples/highlighting.py — turn raw matches into the “keyword in context”
snippets you see on a real search-results page. Call Hit.highlights(fieldname)
and Whoosh finds the best-scoring passages, trims them to a readable length,
and wraps each matched term in markup. The recipe covers:
the one-liner:
hit.highlights("body")off a stored fieldchoosing where snippets are cut —
ContextFragmenter(a window around each match) vsSentenceFragmenter(whole sentences)choosing how matches are marked —
HtmlFormatterwith your own tag and CSS class, orUppercaseFormatterfor plain textfast “pinpoint” highlighting: index the field with
chars=Trueand usePinpointFragmenterso long documents are highlighted without being re-tokenizedhighlighting a field you did not store, by passing the original text to
hit.highlights("body", text=...)
See also How to create highlighted search result excerpts for the full highlighting reference.
Custom analyzers (build your own text pipeline)¶
examples/custom_analyzers.py — the feature that sets Whoosh apart: instead
of a fixed set of “language modes”, you compose your own text-processing
pipeline from a tokenizer and a chain of filters using the | operator:
from whoosh.analysis import RegexTokenizer, LowercaseFilter, StopFilter
analyzer = RegexTokenizer() | LowercaseFilter() | StopFilter()
The first item must be a tokenizer; everything after it is a filter. Attach the
analyzer to a field (TEXT(analyzer=analyzer)) and Whoosh runs the same
pipeline at index time and query time, so the two always agree. The recipe
covers:
watching a pipeline take shape one stage at a time — tokenize, lowercase, drop stop words, then stem with
StemFilteraccent folding with
CharsetFilterand the bundledaccent_mapsocafematchescafénormalising tokens with
SubstitutionFiltersowi-fi,wi_fiandwificollapse to one termcharacter
NgramFilterfor substring / “matches anywhere” searchwiring a custom analyzer onto a field and confirming, with a real index, that
runfindsrunning/runner/ranandZURICHfindsZürich
See also About analyzers for the full analysis reference.
Custom scoring & sorting (control the ranking)¶
examples/scoring_and_sorting.py — ranking is where a search library earns
its keep. Whoosh gives you several independent levers, and this recipe runs each
one against a real index so you can see the ranking change:
tuning the default
BM25Fmodel —Bcontrols document-length normalisation andK1controls term-frequency saturation; per-field values use a<field>_Bkeyword (for exampleBM25F(B=0.75, body_B=0.2))mixing models per field with
MultiWeighting(for exampleTF_IDFfor titles,BM25Feverywhere else)scoring with your own function via
FunctionWeighting, which receives(searcher, fieldname, text, matcher)and returns a float — ideal for experiments and business rulesskipping relevance altogether and sorting by a stored, sortable field with
search(q, sortedby="views", reverse=True)— faster than scoring and often exactly what “newest first” / “most viewed” UIs need
Pass any weighting model to the searcher:
from whoosh import scoring
with ix.searcher(weighting=scoring.BM25F(B=0.0, K1=2.0)) as s:
results = s.search(q)
See also scoring module and Sorting and faceting for the full reference.
Closing indexes cleanly (and avoiding Windows file-lock errors)¶
Whoosh keeps an index’s on-disk files open while a reader or searcher is alive, so it can answer queries without re-opening files each time. If you let those objects be cleaned up by the garbage collector instead of closing them, the files stay open until the object is actually collected.
On POSIX systems that is usually harmless. On Windows, an open file
handle prevents the file from being deleted or replaced, so deleting or
rebuilding an index while a reader is still open surfaces as
PermissionError: [WinError 32] The process cannot access the file because
it is being used by another process. The robust fix is not to sprinkle
gc.collect() calls around — it is to close what you open.
Every reader and searcher is a context manager, so a with block releases
the handles deterministically as soon as the block exits, even on error:
from whoosh.qparser import QueryParser
qp = QueryParser("body", ix.schema)
q = qp.parse("pure AND search")
with ix.searcher() as searcher: # searcher closes on exit
results = searcher.search(q, limit=10)
titles = [hit["title"] for hit in results]
with ix.reader() as reader: # readers are context managers too
total = reader.doc_count()
When you are completely finished with an index object, call ix.close() to
release any cached readers it is holding on your behalf:
ix.close()
After everything is closed, the index directory can be deleted or rebuilt
immediately — including on Windows — with no gc.collect() workaround.
If you use AsyncWriter, remember that its background
thread must finish (via commit()) before the segment’s files are released.
Track any writers you create and join them before tearing down the index.
A complete, runnable version of this pattern lives in
examples/resource_management.py.
A command-line folder search tool¶
examples/search_cli.py — a tiny, dependency-free command-line program that
indexes a folder of text, Markdown, reStructuredText, or source files and lets
you search it straight from your terminal. No server, no external service:
# Index the current directory (creates ./.whoosh_index/)
python examples/search_cli.py index .
# Search it, with highlighted snippets
python examples/search_cli.py search "full text search"
# Re-index only changed/new files and drop deleted ones (fast; uses mtimes)
python examples/search_cli.py index . --update
# Choose which extensions to index, or emit HTML <mark> highlights
python examples/search_cli.py index ~/notes --ext .md,.txt
python examples/search_cli.py search "ranking" --html
It demonstrates several everyday patterns in one place: a
Schema with a unique ID
path, writer.update_document for idempotent upserts,
writer.delete_by_term to prune deleted files, incremental indexing driven by
a stored NUMERIC mtime, field-boosted
MultifieldParser queries, and result highlighting with
ContextFragmenter. It is a single file you can copy
into your own project and adapt.
A full-text search API with FastAPI¶
examples/fastapi_app.py — a small, production-shaped REST API that adds
full-text search to a web service. It exposes PUT /documents/{id} (an
idempotent upsert), DELETE /documents/{id}, and GET /search with
pagination and highlighted snippets:
pip install "whoosh3" fastapi "uvicorn[standard]"
uvicorn fastapi_app:app --reload
curl -X PUT localhost:8000/documents/1 \
-H 'content-type: application/json' \
-d '{"title": "Getting started with Whoosh", "body": "pure-python search"}'
curl 'localhost:8000/search?q=python&page=1&page_size=10'
The search logic lives in a small, framework-free SearchIndex class so it
is easy to unit-test without an HTTP server (run python fastapi_app.py for a
self-contained demo). It shows the pattern you actually need in a service: a
persistent on-disk index opened once at startup and
closed at shutdown (so file handles are
released — important on Windows), writer.update_document upserts keyed on a
unique ID, BM25F ranking, searcher.search_page for
pagination, and highlighted snippets via
HtmlFormatter. See Adding full-text search to your Python app for the
broader “adding search to your app” guide, including a Django variant.
Migrating from Whoosh 2.x / whoosh-reloaded¶
Already using the original Whoosh or Whoosh-Reloaded? The
MIGRATING.md guide at the repo root explains what changed: the import
package is still whoosh, the on-disk index format is unchanged, and the
public API is the same. In most cases the only change you make is the package
you install.