Find and retrieve artifacts

Different consumers ask different questions of the registry. An evaluation job wants the latest model. A backfill wants the newest version of every August partition. A dashboard wants the list of regions that have data, and an auditor wants everything a given run produced. Each of these is one call on flyte.remote.Artifact, or one flyte get artifact command.

The recipes below go from simple to specific. The Python calls need a client, so call flyte.init_from_config() first. All of them accept project= and domain= and default to your config.

from datetime import date, datetime

import flyte
from flyte.remote import Artifact

flyte.init_from_config()

By name and version

model = Artifact.get("trained-model")                  # latest version
model = Artifact.get("trained-model", version="v3")    # a pinned version

flyte get artifact trained-model        # every version, newest first
flyte get artifact trained-model v3     # one version

One partition

For a partitioned artifact, pass one value per partition key to get the latest version of that partition:

events = Artifact.get("raw_events", date=date(2026, 8, 2), region="us")

flyte get artifact raw_events --partition date=2026-08-02 --partition region=us

A date matches a daily partition, a datetime an hourly one, and a string that reads as an ISO date or timestamp is treated as a time value too. If a key name clashes with an argument of get (such as name or version), pass it in a partitions={...} mapping instead.

A range of partitions

listall takes a (start, end) tuple on the time key, inclusive at both ends, and a list on a string key to match any of its values. Add latest_per_partition=True to get one version per partition instead of every version:

august = list(
    Artifact.listall(
        "raw_events",
        date=(date(2026, 8, 1), date(2026, 8, 31)),
        region=["us", "eu"],
        latest_per_partition=True,
    )
)
for a in august:
    print(a.partitions, a.version)

flyte get artifact raw_events --partition date=2026-08-01..2026-08-31 --partition region=us,eu --latest-per-partition

This is the query a backfill plans with: compare the partitions that came back against the days you expect, and fill the gaps.

The values a partition key has

partition_values returns the distinct values of one key, sorted. Other keys narrow it:

Artifact.partition_values("raw_events", "region")                      # ["eu", "us"]
Artifact.partition_values("raw_events", "date", region="us")           # [date(2026, 8, 1), ...]

Time keys come back as date values, or datetime for hourly partitions. Artifact.get_schema(name) tells you which keys an artifact has.

By where it came from, or what it is

listall filters server-side on provenance and metadata. Combine as many as you need:

Artifact.listall(source_run="my-run")                                  # everything a run produced
Artifact.listall(source_run="my-run", source_action="a0")              # one action of that run
Artifact.listall(source_external_ref="s3://partner-bucket/drop/2026-08-18.csv")
Artifact.listall(kind="model", attrs={"framework": ["torch", "jax"]})  # any listed value matches
Artifact.listall("trained-model", created_after=datetime(2026, 8, 1), limit=10)

flyte get artifact --source-run my-run
flyte get artifact --kind model --attr framework=torch

listall returns an iterator, newest first.

Browsing names

for group in Artifact.list_names(search="events"):
    print(group)

flyte get artifact                    # every name, with its latest version and version count
flyte get artifact --search events

Passing what you found to a run

An Artifact binds to a task input of the matching type. The task receives a plain File, Dir, or DataFrame:

run = flyte.run(summarize, events=Artifact.get("raw_events", date=date(2026, 8, 2), region="eu"))

A task that takes a list accepts a list of artifacts. listall is an iterator, so materialize it first:

august = list(Artifact.listall("raw_events", date=(date(2026, 8, 1), date(2026, 8, 31)), latest_per_partition=True))
run = flyte.run(monthly_rollup, days=august)

To load the data in a notebook or script instead, see reading an artifact’s value directly.

Things to know

• get returns exactly one version. It rejects a range or a list (use listall) and a version combined with partition values, since a version already names one artifact.
• latest_per_partition needs an artifact name.
• A read by partition raises an error when that partition has no version yet.