Move from SDK 0.6 to 0.7

SDK 0.7 keeps the released supervised, research and causal interfaces and adds a simpler hosted setup plus explicit target-column convenience.

Select the endpoint explicitly when local#

Precedence is constructor base_url → WELT_BASE_URL → https://welt.ergodic.dev. The default changed from localhost to the hosted origin. If you develop against a local server, select it deliberately:

python

from welt import Client

with Client(base_url="http://localhost:8000") as client:
    models = client.models()

Public model-catalogue reads remain keyless. Authenticated methods validate missing credentials before issuing requests and explain how to create a workspace key. Invalid/expired credentials and insufficient permissions remain typed server failures. Keys are never saved implicitly.

Use a DataFrame target#

Existing fit(X, y) remains valid. The additive form accepts exactly one target source, requires a directly supplied DataFrame, and removes its named target from features before upload:

python

# Continue with the real train table and target name from Start.
model.fit(train, target=target)

Select Classifier or Regressor explicitly. Conflicting y/target, missing or duplicate target names and invalid schema reject locally before upload. Query rows still contain features only. Existing fitted predictor identities remain pinned; upgrading the client does not refit or change a remote model.

Reconnect after waiting#

Job.result(timeout=...) budgets initial inspection, polling, GET retry waits and final result retrieval. It never cancels or resubmits work. Async requests are bounded using the remaining wait budget. Synchronous HTTP phases receive the remaining budget and the deadline is checked after each request; separate network phases and custom blocking transports do not provide a hard wall-clock interrupt. Use the preserved job_id with the same endpoint/current workspace credentials. See Wait and recover.

Keep the release boundary#

0.7 excludes the held resumable CSV and large batch candidates. CSV/Parquet convenience reads still load into memory. Full all-model qualification, public Ergodic distribution and calibrated/conformal confidence retain their own gates. The 0.6 documentation remains available with its original contract and pinned notebook sources.