what the widgets should feel like
The SDK widgets, rendered & alive.
Each widget from the tutorial notebooks, built as a working prototype in the
brand system — filters filter, selections select, the url4 recipe updates as you edit.
This is the target behavior and look for eng, not the SDK itself.
These are mock prototypes — no backend, no real inference (consistent with the SDK's
“simulated but honest” stance). Every widget mounts from
data-widget="…" via SFWidgets.mount(),
so any one is liftable into another page. The .value line under each is the
real object it produces.01 · Authentication
Plug in providers before you compose. Keys stay in memory; grading is simulated.
In [1]
sf.setup() # interactive panel — connect providers
Out
In [2]
sf.connections() # live status; click a row to authenticate
Out
02 · Models
The catalog service — search, sort, filter down to provider/model ids.
In [3]
b = sf.models.browse() # search + sort + price filter; click to add b
Out
In [4]
sf.models.get("anthropic/claude-opus-4.8") # a Model card
Out
03 · Fusions
Compose, cost, run, then read the result. The recipe is the identity.
In [5]
c = fusion.compose(options=sf.models.list(max_price=20)) c # add/remove members, reduce, judge, prompts; url4 updates live
Out
In [6]
r = fusion.run() # configure + watch: cost, progress, score r
Out
In [7]
run.stats() # which models carried the fusion
Out
In [8]
i = run.inspect() # master/detail, wrong-only, per-participant reasoning i
Out
04 · Custom scripts
Write your own reduce/loop logic — it runs for real. Attach with reduce_with / loop_with.
In [9]
reducer = sf.script(reduce) # from a function, a .py path, or a code string reducer
Out
05 · Leaderboard
One board: compare against baselines and other fusions, fork the recipe.
In [10]
board = sf.leaderboard("GPQA Diamond") board # benchmark switch, mine-only, fork; .to_frame() for a table
Out