Quickstart

Create and run a reviewed semantic pipeline from Python.

1. Sign in and select a workspace

Create and confirm an account at app.relifold.com. Then create an organization and open its workspace.

from scratchpad.api import ScratchpadClient

account = ScratchpadClient.sign_in_with_password(
    "https://api.relifold.com",
    email="you@example.com",
    password="your-password",
)
workspace = account.select_workspace()

If the account has more than one organization or workspace, pass an ID, slug, or name as organization= or workspace=. Relifold never silently selects the first result.

2. Add a dataset

Upload a small CSV, JSON, or JSONL file from Data, or use the client:

workspace.upload(
    space="products",
    filename="products.jsonl",
    content=b'{"name":"Apple Inc."}\n{"name":"Apple"}\n',
)

3. Create a pipeline version

In the browser, create a pipeline and ask the Planner for a draft such as: “Group product names only when they identify the same brand. Preserve distinct brands and non-brand categories.” Inspect the generated inputs, output, keys, prompt, and parameters.

4. Add a review checkpoint

Choose the grouping task, request an automated sample, and prepare a small human curation sample. Human reasoning is optional but useful when a simple good/bad verdict would be ambiguous.

5. Run and inspect

run = workspace.run(
    "product-brands",
    "v1",
    max_cost_usd=5.0,
)
session_id = run["session_id"]
print(workspace.session(session_id))

The session page shows task state, elapsed time, calls, failed calls, cost, inputs, and outputs. After reviewing results, create an editable successor version and accept only the changes you want.

Next

Read Spaces, then use the operator chooser to understand the generated pipeline.