| name | generate-pptx |
| description | Generate a PowerPoint deck from ingested source documents using the agnostic Document tree (parametrised by slide_count, style, audience, tone) |
| model | deepseek/deepseek-v4-pro |
| tier | pro |
| inputs | [{"name":"source_document_ids","type":"list[str]","required":true},{"name":"brief","type":"str","required":true},{"name":"slide_count","type":"int","required":false,"description":"Target number of slides (soft cap — renderer truncates, default 10)"},{"name":"style","type":"str","required":false,"description":"e.g. \"clinical\", \"corporate\", \"casual\", \"academic\" (default professional)"},{"name":"audience","type":"str","required":false,"description":"Target audience (default general)"},{"name":"tone","type":"str","required":false,"description":"e.g. \"warm\", \"concise\", \"authoritative\" (default neutral)"}] |
| outputs | [{"name":"file_path","type":"str"},{"name":"file_url","type":"str"},{"name":"title","type":"str"},{"name":"slide_count","type":"int"},{"name":"model","type":"str"},{"name":"source_document_ids","type":"list[str]"}] |
generate-pptx
Generates a PowerPoint (.pptx) deck from one or more ingested source
documents stored in the OfficePlane database.
What it does
- Loads the requested source documents from Postgres via Prisma (chapters,
sections, summaries).
- Prompts the LLM to produce an agnostic Document JSON tree aligned to
the CommonMark / Pandoc AST vocabulary — NOT the legacy
modules → lessons → blocks schema.
- Parses the JSON into a
Document dataclass tree via parse_document.
- Forces
doc.meta.render_hints["max_slides"] = slide_count so the renderer
hard-truncates if the LLM overshot.
- Renders the tree to
.pptx bytes via render_pptx.
- Writes the result to
/data/workspaces/<job_id>/output.pptx.
Slide layout strategy
The renderer maps the document tree to slides as follows:
- Title slide — always first if
doc.meta.title is non-empty.
- Section divider slide — for every level-1 Section that has child Sections
beneath it (produced by
_LAYOUT_SECTION_HEADER).
- Content slide — for every leaf-bearing section (a Section whose children
contain at least one non-Section block). Batches heading / paragraph / list /
callout / quote / divider blocks into a single slide body.
- Big-primitive slides — Table, Figure, or Code blocks each get their own
dedicated slide to avoid crowding.
Plan the section nesting depth and branching factor so the total slide count
(title + dividers + content + big-primitives) stays close to slide_count.
Parameters
| Parameter | Default | Notes |
|---|
slide_count | 10 | Renderer hard-truncates to this cap |
style | professional | Visual/verbal register: clinical, corporate, casual, academic |
audience | general | Who will view the deck |
tone | neutral | Voice: warm, concise, authoritative, etc. |
Expected JSON shape
{
"type": "document",
"schema_version": "1.0",
"meta": {
"title": "My Deck",
"language": "en",
"render_hints": {"max_slides": 10}
},
"children": [
{
"type": "section",
"id": "s1",
"level": 1,
"heading": "Introduction",
"children": [
{"type": "paragraph", "text": "Opening bullet."},
{
"type": "list",
"ordered": false,
"items": [
{"type": "paragraph", "text": "Key point one"},
{"type": "paragraph", "text": "Key point two"}
]
}
]
}
],
"attributions": [
{
"node_id": "s1",
"document_id": "<source-doc-uuid>",
"section_id": "<source-section-id>"
}
]
}
Key rules:
- Sections nest recursively. Use
level 1..6 to indicate depth.
- Block types:
heading, paragraph, list, table, figure, code,
callout, quote, divider.
- Each content slide should have at most 6 bullet items.
- Headings must be 60 characters or fewer; bullets 18 words or fewer.
- Never emit
"modules" or "lessons" — use nested sections instead.
- Every non-trivial node should have an
id (short string).
- Provide at least one attribution per major section pointing back to a
source
document_id.
