[serhii73]

The new strategy tokenizes the text on characters that never occur inside a date expression and tries to parse token n-grams, longest first, so that the most complete date is found. Compared to the default translation-based "split" strategy it produces more predictable results on noisy real-world text.

The strategy is exposed through the new "strategy" parameter of search_dates and defaults to the existing behavior.

[codecov]

Codecov Report

❌ Patch coverage is 94.64286% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 97.07%. Comparing base (762dfff) to head (0470e33).

Files with missing lines	Patch %	Lines
dateparser/search/ngram_search.py	91.89%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #1351      +/-   ##
==========================================
- Coverage   97.12%   97.07%   -0.06%     
==========================================
  Files         235      236       +1     
  Lines        2924     2972      +48     
==========================================
+ Hits         2840     2885      +45     
- Misses         84       87       +3     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[serhii73]

Implementation summary

This PR adds an opt-in "ngram" strategy to search_dates, adapting a date-search algorithm from a production date extraction pipeline where it has consistently produced more predictable results than the split-based search on noisy real-world text. The default behavior is unchanged (strategy="split").

How the ngram strategy works

For example, in "Posted on 4 October 1957 by admin":

1. Tokenize the text on characters that never occur inside a date expression (whitespace and ,|()@), keeping each token's position in the original text.
2. Drop noise words (on, at, of, a) that may appear between date components.
3. Scan tokens left to right, trying n-grams of up to 7 tokens, longest first — 4 October 1957 by admin ✗ → … → 4 October 1957 ✓ — so the most complete date phrase wins over fragments like October 1957.
4. Reject blacklisted candidates that parse but are almost always false positives: bare 1–3 digit numbers, #123-style refs, separator runs, single letters.
5. Parse candidates with DateDataParser(languages=..., use_given_order=True); after a match, scanning resumes past it, so all non-overlapping dates are returned.
6. Return (substring, datetime) pairs, with substrings taken from the original text (punctuation-stripped), matching the existing search_dates contract.

Changes

• New module dateparser/search/ngram_search.py with _NgramDateSearch implementing the algorithm above.
• New strategy parameter on the public search_dates() and DateSearchWithDetection.search_dates(): "split" (default) or "ngram"; invalid values raise ValueError.
• Language detection and candidate-language resolution are shared between both strategies; the ngram path passes the full candidate-language list to the parser in given order and falls back to settings.DEFAULT_LANGUAGES like the split path.
• RETURN_TIME_SPAN post-processing extracted into a shared module-level _add_time_span_results() helper, so both strategies have feature parity.
• 12 new tests in TestNgramSearch: en/fr/ru extraction, noise-word/punctuation handling, blacklist, RELATIVE_BASE, RETURN_TIME_SPAN, module-level passthrough, add_detected_language, None when no dates found, and strategy validation.

Verification

• tests/test_search.py: 256 passed
• Full suite (pytest tests): 24072 passed, 20 skipped, 1 xfailed
• ruff check and ruff format: clean

Usage

from dateparser.search import search_dates

search_dates("Posted on 4 October 1957 by admin", strategy="ngram")
# [('4 October 1957', datetime.datetime(1957, 10, 4, 0, 0))]