| name | doc/semantic-chunking |
| description | Split arbitrary prose or a Markdown file into a sensible chunk hierarchy in a chunked-graph Document. |
| tools | ["Document"] |
| license | Apache-2.0 |
Semantic chunking
When the operator hands you a body of text (pasted prose, a spec, a Markdown
file) and asks to "import it" or "chunk it", turn it into a hierarchy of
chunks — do NOT dump it into one giant chunk.
Method
- Read the text's own structure first: existing headings, numbered sections,
topic shifts, and natural paragraphs are your chunk boundaries.
- Choose a hierarchy: a top-level chunk per major section, child chunks per
sub-section. Aim for chunks that are individually meaningful and editable —
a few sentences to a few paragraphs each, not one-liners and not whole pages.
- Give every chunk a short, descriptive title (it becomes the heading).
- Create them top-down so parents exist before children:
Document op=create_chunk document_id=<id> parent_id=<parent> title="…" body="…"
(omit parent_id for a child of the root, or pass the document's root chunk).
- Carry a
type when the operator's domain implies one (e.g. section,
requirement, task) and a status if they use a workflow.
Notes
- If the text is a loomcycle export (it has
<!-- loom: … --> comments), prefer
the deterministic import_md op instead (see the doc/md-import skill).
- Keep the operator's wording in the body; don't paraphrase unless asked.