timefuzz

Fuzzy, natural-language time parsing that goes past dateparser.

timefuzz resolves human phrases like "sometime next week", "the Tuesday after my birthday", or "end of Q3" into concrete datetimes or ranges — with a confidence score, so your app knows when to ask the user to confirm.

Rust core, thin Python API. Ships as compiled wheels: pip install timefuzz, no Rust toolchain required.

import timefuzz as tf
from datetime import date, datetime

now = datetime(2026, 7, 12, 15, 30)

tf.parse("next friday", now=now)
# Instant(when=datetime(2026, 7, 17, 9, 0), confidence=0.95,
#         interpretation='next Friday, default 09:00')

tf.parse("sometime next week", now=now)
# Range(start=datetime(2026, 7, 13, 0, 0), end=datetime(2026, 7, 19, 23, 59, 59),
#       confidence=0.8, interpretation='the calendar week after this one')

tf.parse("the tuesday after my birthday", now=now,
         anchors={"my birthday": date(2026, 8, 3)})
# Instant(when=datetime(2026, 8, 4, 9, 0), confidence=0.9,
#         interpretation='first Tuesday strictly after 2026-08-03')

tf.parse("end of q3", now=now)
# Range(start=datetime(2026, 9, 1, 0, 0), end=datetime(2026, 9, 30, 23, 59, 59),
#       confidence=0.85, interpretation='last month of Q3')

Why another date parser?

Existing parsers (dateparser, parsedatetime) handle "in 3 days" and "next Friday" but fall over on:

• Ranges & vagueness — "sometime next week" is a span, not an instant.
• Anchored relatives — "the Tuesday after my birthday" needs a user-supplied anchor date.
• Business calendars — "end of Q3", "next business day", "the 2nd Monday of March", "10 business days after the invoice date".
• Honest ambiguity — "next weekend" said on a Wednesday has two defensible readings; timefuzz returns both instead of silently guessing.
• Confidence — callers want to know how fuzzy a result is, so they can confirm low-confidence parses with the user instead of silently guessing.

Every parse returns the resolved value plus a confidence score plus the interpretation the parser chose, so schedulers, reminder apps, and chat bots can decide when to double-check.

Install

pip install timefuzz

Wheels are published for CPython 3.10–3.13 on Linux (manylinux), macOS (x86-64 + arm64), and Windows (x86-64). An sdist falls back to building from source (requires a Rust toolchain) if no wheel matches.

What it returns

Every parse yields one of three shapes:

If nothing matches at all, ParseError is raised with the reason. An unknown anchor ("after my graduation" with no "my graduation" anchor) returns Ambiguous with an empty candidate list and an explanatory reason. Candidates are ordered most-likely-first.

match tf.parse(user_text):
    case tf.Instant(when=when, confidence=c) if c >= 0.8:
        schedule(when)
    case tf.Instant(when=when):
        confirm_with_user(when)          # low confidence -> ask first
    case tf.Range(start=s, end=e):
        offer_slot_picker(s, e)
    case tf.Ambiguous(candidates=cands, reason=why):
        disambiguate(cands, why)

API

def parse(
    text: str,
    now: datetime | None = None,        # reference moment (default: datetime.now())
    anchors: dict[str, date] | None = None,
    config: Config | None = None,
) -> Instant | Range | Ambiguous: ...

@dataclass(frozen=True)
class Config:
    default_time: time = time(9, 0)     # time attached to date-only results
    week_start: Weekday = Weekday.MON   # or Weekday.SUN
    next_skips_today: bool = True       # "next friday" said on a Friday
    tz: tzinfo | None = None            # naive by default; tz-aware opt-in
    holidays: Callable[[date], bool] | None = None   # business-day hook

What the grammar understands (v0.4)

• Relative offsets: in 3 days, 2 weeks ago, 3 days from now, in a week, tomorrow, yesterday, today, now
• Bare durations (the in is optional): 30 minutes, 2 hrs, 3 days, half an hour, half a day, a couple hours, a few days, back in a couple hours
• Period spans: next week, this month, last quarter, next year, bare month names (august)
• Weekday navigation: next friday, this tuesday, last monday, bare friday, friday next week, monday last week
• Weekends: this weekend, next weekend, last weekend, sometime next weekend, the weekend after the wedding
• Ordinal-in-period: 2nd monday of march, 2nd monday of march 2027, last friday of october, first monday of next month
• Anchored: the tuesday after my birthday, 2nd tuesday after my birthday, the day before the wedding, 2 weeks after my birthday, the week after my birthday, 3 business days after the invoice date, bare on my birthday (anchors supplied by you)
• Clock times: next friday at 3pm, tomorrow at 7:30, friday at midnight, bare 3pm / 15:45 / noon (next occurrence)
• Vague spans: sometime next week, sometime in august, early march, mid q3, mid month, late next month, sometime early next month
• Business calendar: next business day, in 3 business days, 5 business days ago, end of q3, start of q4, end of month, end of next month, first business day of next month, last business day of the month, eow/eom/eoq/eoy

See docs/grammar_reference.md for the complete rule catalogue and docs/cookbook.md for recipes.

Conventions (the fine print)

These are deliberate, documented choices — see the config knobs above:

• "next Friday" when today is Friday is culture-dependent. Default: skip today (next_skips_today=True), i.e. you get the Friday seven days out.
• Date-only results get default_time (09:00 by default); an explicit trailing time (… at 3pm) overrides it.
• Arithmetic offsets keep the clock: in 3 days = now + 72h; tomorrow = tomorrow at default_time.
• The in is optional. 30 minutes resolves exactly like in 30 minutes. Approximate amounts (a couple hours) resolve too, but at 0.7 so you confirm rather than schedule silently, and half a month stays a ParseError because there is no exact half of a calendar month.
• Ranges are inclusive, 00:00:00 through 23:59:59 of the last day.
• Weekends are Sat–Sun, whatever week_start says.
• Naive by default. Set Config(tz=...) to get tz-aware results; math is wall-clock, so "tomorrow 09:00" is 09:00 across a DST jump.
• Month names and quarters assume the current cycle and roll forward if already past (the interpretation string says so).
• Genuinely contested phrases return Ambiguous rather than a guess: next weekend midweek, sunday said on a Sunday, july said during July.
• Confidence is deterministic, not learned — the same phrase always gets the same score, offsets past ~10 years are trusted slightly less, and stacked hedges (sometime early …) cap lower. See docs/confidence.md.

Development

# build + test
pip install maturin
maturin develop --release
pytest

# rust checks
cargo clippy --all-targets
cargo fmt --check
cargo bench          # informational: full-corpus, single-phrase, tokenizer-only

The test suite is corpus-driven: tests/corpus*.jsonl map phrases (plus a fixed reference now) to expected outputs. Adding a phrase = adding a line.

Roadmap

• v0.2 — ✅ richer anchored phrases + business-calendar rules
• v0.3 — ✅ confidence-model refinement, more Ambiguous candidates
• v0.4 — ✅ corpus expansion, benches, cookbook growth, ✅ bare durations (30 minutes, half an hour, a couple hours)
• v1.0 — ✅ i18n-ready grammar structure (Locale seam, English-only), ✅ clock-time support, ✅ written stability policy; remaining: a soak period on 0.3.x, then the freeze

See CHANGELOG.md for details.

License

MIT