datefinder - extract dates from text

.. image:: https://github.com/akoumjian/datefinder/actions/workflows/python-package.yml/badge.svg :target: https://github.com/akoumjian/datefinder :alt: Build Status

.. image:: https://img.shields.io/pypi/dm/datefinder.svg :target: https://pypi.python.org/pypi/datefinder/ :alt: pypi downloads per day

.. image:: https://img.shields.io/pypi/v/datefinder.svg :target: https://pypi.python.org/pypi/datefinder :alt: pypi version

A python module for locating dates inside text. Use this package to extract date-like strings from documents and turn them into useful datetime/temporal objects.

As of 1.0.0, find_dates(...) defaults to the v2 compatibility engine. The original engine remains available as find_dates_legacy(...).

Installation

Requires Python 3.9+.

With pip

.. code-block:: sh

pip install datefinder

If a compatible prebuilt wheel is unavailable for your platform, pip will build from source and requires a Rust toolchain.

Note: I do not publish the version on conda forge and cannot verify its integrity.

What You Can Do With datefinder

datefinder is a Python date parser for extracting dates from unstructured text. It is useful when your data is not already normalized, for example:

• emails, tickets, and support conversations
• contracts, policies, and legal text
• logs, reports, and markdown/wiki pages
• scraped HTML and mixed-format documents

You can use it to:

• parse explicit calendar dates like January 4th, 2017 or 2024-11-03 18:00
• parse relative expressions like tomorrow, yesterday, and in 3 days
• parse multiple date formats in one pass (month-name, slash, ISO, hyphen)
• anchor relative parsing to a reference/base date
• return either compatibility datetimes or typed structured match objects

In short: if you need to find and parse dates from text in Python, especially inside large documents with mixed formatting, datefinder is designed for that.

Common workflows:

• migration from legacy date extraction code: use find_dates_legacy(...) for parity, then move to find_dates(...)
• modern typed extraction: use extract(...) to get match kinds, spans, confidence, and structured values
• command line processing: use datefinder --engine extract --json in shell pipelines

Example (Python):

.. code-block:: python

import datefinder
from datetime import datetime, timezone

text = "Meeting tomorrow; launch on 2024-11-03 18:00 UTC."
ref = datetime(2026, 3, 19, 12, 0, tzinfo=timezone.utc)

# Compatibility datetimes
print(list(datefinder.find_dates(text, base_date=ref)))

# Typed extraction
for match in datefinder.extract(text, reference_dt=ref):
    print(match.kind, match.text, match.value)

Example (CLI):

.. code-block:: sh

datefinder --reference "2026-03-19T12:00:00+00:00" --json \
  "Meeting tomorrow; launch on 2024-11-03 18:00 UTC."

How to Use

.. code-block:: python

In [1]: string_with_dates = """
   ...: ...
   ...: entries are due by January 4th, 2017 at 8:00pm
   ...: ...
   ...: created 01/15/2005 by ACME Inc. and associates.
   ...: ...
   ...: """

In [2]: import datefinder

In [3]: matches = datefinder.find_dates(string_with_dates)

In [4]: for match in matches:
   ...:     print(match)
   ...:
2017-01-04 20:00:00
2005-01-15 00:00:00

CLI

The package now includes a CLI entrypoint:

.. code-block:: sh

datefinder --json "tomorrow and 2024-12-10"

You can also run it as a module:

.. code-block:: sh

python -m datefinder --engine extract --json --reference "2026-03-18T00:00:00+00:00" "in 3 days"

Engine options:

• default: find_dates(...) (v2 compatibility default)
• legacy: find_dates_legacy(...)
• compat: find_dates_compat(...)
• extract: typed extract(...) output

Common options:

• --reference <ISO8601>: anchor for relative dates/times (equivalent to base_date/reference_dt)
• --first {month,day,year}: disambiguation for numeric dates
• --strict: stricter matching
• --json / --pretty: machine-readable output
• --source / --index: include source span details (default/legacy only)
• --locale <code>: locale hint for extract (repeatable)
• --no-month-only: disable month-only inference ("May" -> YYYY-05-01)
• --compact-numeric: enable compact numeric parsing (e.g. 20240315)
• --no-multiline: disable cross-line matching

Examples:

.. code-block:: sh

# default engine (v2 compatibility), anchored relative parsing
datefinder --reference "2026-03-19T12:00:00+00:00" --json "tomorrow and 2024-12-10"

# explicit legacy behavior, include source text and indices
datefinder --engine legacy --source --index --json "created 01/15/2005 by ACME"

# typed extract output with locale hints
datefinder --engine extract --locale en --locale fr --pretty --json "in 3 days and demain"

# read long input from stdin
cat document.txt | datefinder --engine extract --json

Relative and duration values:

• default / legacy / compat engines emit datetimes.
• extract emits typed values:
  • relative includes both resolved_datetime and delta_seconds.
  • duration includes total_seconds and normalized components.

V2 Typed API

This repository includes a v2 extraction API with typed match objects and first-class support for relative expressions and durations.

.. code-block:: python

import datefinder
from datetime import datetime, timezone

matches = datefinder.extract(
    "in 3 days we deploy on 2024-11-03 18:00",
    reference_dt=datetime.now(timezone.utc),
)
for m in matches:
    print(m.kind, m.text, m.value)

There is also a compatibility helper for migrating existing code:

.. code-block:: python

for dt in datefinder.find_dates_compat("tomorrow and 2024-12-10"):
    print(dt)

If you need the original parser behavior exactly:

.. code-block:: python

for dt in datefinder.find_dates_legacy("April 9, 2013 at 6:11 a.m."):
    print(dt)

Rust kernel source is under rust/datefinder-kernel and is required for v2/default runtime behavior.

Rust Portability

• Compiled Rust extensions are platform-specific, they do not run on every system by default.
• Release wheel targets:
  • Linux glibc: x86_64 and aarch64 (manylinux2014)
  • Linux musl: x86_64 and aarch64 (musllinux_1_2)
  • macOS: x86_64 and arm64
  • Windows: x86_64
• If no compatible wheel is available, pip builds from source and requires a Rust toolchain.

Conformance and Ambiguity Reports

Build a reproducible corpus from legacy tests and generate differential reports between legacy behavior and find_dates_compat:

.. code-block:: sh

python scripts/build_conformance_corpus.py
python scripts/diff_legacy_v2.py

This writes:

• conformance/legacy_parity_cases.jsonl
• conformance/reports/legacy_v2_diff_report.md
• conformance/reports/ambiguity_showcase.md
• conformance/reports/behavior_change_changelog.md

The ambiguity showcase also supports interpretation judgments in conformance/interpretation_judgments.jsonl to assess whether legacy behavior is semantically preferable for ambiguous real-world cases.

See also:

• CONTRIBUTING.md for developer setup and validation commands.
• RELEASE.md for release checklist.

Benchmark Snapshot

The command below generates a local benchmark snapshot comparing:

• v2: datefinder.extract(...)
• legacy: datefinder.find_dates_legacy(...)
• dateparser: dateparser.search.search_dates
• duckling_http: Duckling POST /parse

Run:

.. code-block:: sh

# optional: run duckling locally
docker run --rm -p 8000:8000 rasa/duckling:latest

python bench/bench_readme_compare.py \
  --iterations-small 12 \
  --iterations-large 2

Latest local snapshot (2026-03-19 UTC):

+------------------+--------+---------------+-------------------+----------------------+-------------------------+---------------+------------------+----------------------+ | dataset | size | v2 median (s) | legacy median (s) | dateparser median (s)| duckling_http median (s)| v2 vs legacy | v2 vs dateparser | v2 vs duckling_http | +==================+========+===============+===================+======================+=========================+===============+==================+======================+ | core_corpus | 498 | 0.000236 | 0.003042 | 0.180596 | 0.050266 | 12.91x | 766.74x | 213.41x | +------------------+--------+---------------+-------------------+----------------------+-------------------------+---------------+------------------+----------------------+ | seattle_html_76k | 74838 | 0.037436 | 0.281466 | 0.771712 | 25.353595 | 7.52x | 20.61x | 677.24x | +------------------+--------+---------------+-------------------+----------------------+-------------------------+---------------+------------------+----------------------+ | test_data_560k | 552301 | 0.239391 | 2.840845 | n/a | n/a | 11.87x | n/a | n/a | +------------------+--------+---------------+-------------------+----------------------+-------------------------+---------------+------------------+----------------------+

Notes:

• n/a means unavailable/failed for that dataset in this run.
• dateparser/duckling_http are skipped by default for documents larger than 200k bytes unless forced.