1. Prepare inputs
A portable docset can contain:
my-docset/
vocabulary.yaml
prose/
overview.md
openapi.yaml # optional; spec.json is also supported
Define names and aliases in the vocabulary. Bound the input to one coherent subject. Include authoritative specs when available; prose supplies behavioral context.
2. Extract
uv sync --extra build
uv run scull-schemafier run \
--concept myconcept \
--docset /path/to/my-docset \
--out schemas/ \
--model provider/model-name
Configure the selected provider's credentials first. Local OpenAI-compatible endpoints can be
selected with --model-api-base. See configuration.
Explicit --vocab, --spec, --oas, and --prose flags override discovered inputs. Keep output
outside the docset.
3. Review
Exit status 2 means output was written but needs review. Check evidence, field assignments, conflicts, and omissions. Do not treat extraction confidence as approval.
uv run scull validate
uv run scull-schemafier audit-assignments --schemas schemas --vocab /path/to/my-docset/vocabulary.yaml
Follow review and release, then build an index if you need semantic search. Exact lookup works without an index.