Recency

A Search can be biased or filtered by record age via Search.recency(...), using one of two typed strategies, plus an optional reference "now".

Hard window filter — Range

Range(from_, to) restricts results to the inclusive [from_, to] window of unix-second timestamps. from_ is named with a trailing underscore because from is a Python keyword. Candidates are pruned before fusion; a very narrow window over a large corpus can return fewer than top_k hits.

from valise import Search, Range

q = Search().text("body", "fox").recency(Range(from_=1_700_000_000,
                                                to=1_700_086_400))

Exponential decay — HalfLife

HalfLife(days) applies an exponential recency decay with the given half-life (in days). Combine with .now(unix_secs) to set the reference time the decay is measured from (otherwise the current wall clock is used).

from valise import Search, HalfLife

q = (
    Search()
    .text("body", "fox")
    .recency(HalfLife(days=7.0))
    .now(1_700_000_000)
)

A third strategy treats recency as one more RRF channel (newest-first, contributing weight/(k + rank + 1); requires RRF fusion):

from valise import Rrf, RrfChannel, Search

hits = store.search("notes", Search()
    .text("body", "release notes")
    .recency(RrfChannel(half_life_days=14, weight=0.5))
    .fuse(Rrf(k=60))
    .top_k(10))