Generating Documents (Word + PowerPoint)

OfficePlane generates .docx and .pptx files from natural language prompts via two job endpoints. Both endpoints return immediately with a job ID and stream progress over SSE.

Generating a Word Document

POST /api/jobs/invoke/generate-docx
Content-Type: application/json

{
  "topic": "Overview of quantum-safe cryptography standards",
  "page_count": 8,
  "audience": "engineering team",
  "tone": "technical"
}

Response (202 Accepted):

{
  "job_id": "job_01j2kqx8v",
  "status": "queued",
  "stream_url": "/api/jobs/job_01j2kqx8v/stream",
  "skill": "generate-docx"
}

Parameters

Parameter	Type	Required	Description
topic	string	yes	Subject matter of the document
page_count	integer	no (default: 5)	Target page count
audience	string	no	Intended reader — influences vocabulary and depth
tone	string	no	Writing register: technical, executive, educational

---

Generating a PowerPoint Presentation

POST /api/jobs/invoke/generate-pptx
Content-Type: application/json

{
  "topic": "Q3 infrastructure cost reduction initiative",
  "slide_count": 12,
  "style": "professional",
  "audience": "executive leadership",
  "tone": "executive"
}

Response (202 Accepted):

{
  "job_id": "job_01j3mnp2q",
  "status": "queued",
  "stream_url": "/api/jobs/job_01j3mnp2q/stream",
  "skill": "generate-pptx"
}

Parameters

Parameter	Type	Required	Description
topic	string	yes	Subject matter of the presentation
slide_count	integer	no (default: 10)	Target number of slides
style	string	no	Visual register: professional, minimal, bold
audience	string	no	Intended audience — influences slide density
tone	string	no	Writing register: technical, executive, educational

---

The Document JSON Schema the Model Emits

The LLM is prompted to emit a document tree conforming to schema_version 1.0 (see The Document Tree) before the renderer converts it to OOXML. The intermediate JSON is stored alongside the output file and forms the basis for subsequent edits.

A minimal generation output looks like:

{
  "schema_version": "1.0",
  "id": "doc_01j2kqx8v",
  "title": "Quantum-Safe Cryptography Standards",
  "sections": [
    {
      "id": "sec_01",
      "title": "Introduction",
      "depth": 0,
      "children": [],
      "blocks": [
        { "id": "blk_01", "type": "heading", "level": 1, "content": "Introduction" },
        { "id": "blk_02", "type": "paragraph", "content": "Post-quantum cryptography addresses..." }
      ]
    }
  ]
}

The renderer iterates sections depth-first, maps each Block to the corresponding OOXML element, and writes the file. Because the intermediate JSON is persisted, every generation is immediately editable without re-ingestion.

---

Example Results

The following results are from smoke tests run against the default flash-tier model:

Prompt	Format	Slides / Pages	Generation Time	File Size
"Overview of quantum-safe cryptography, 8 pages, technical"	DOCX	8 pp	18 s	42 KB
"Q3 infra cost reduction, 12 slides, executive"	PPTX	12 slides	24 s	98 KB
"Onboarding guide for new engineers, 6 pages"	DOCX	6 pp	14 s	31 KB
"Product roadmap pitch deck, 15 slides"	PPTX	15 slides	31 s	124 KB

Times measured end-to-end from job submission to file available, on a single worker with DeepSeek v4-flash.

---

Workspace Layout

Each generation job creates an isolated workspace directory:

/data/workspaces/<job_id>/
├── document.json        # Intermediate document tree (schema_version 1.0)
├── output.docx          # or output.pptx
└── revisions/
    └── rev_00.json      # Initial revision snapshot (parent_revision_id: null)

Future versions will write incremental rev_N.json snapshots for each agent edit applied to the workspace before final render.

The workspace is retained for 24 hours after job completion, then pruned by the cleanup worker. Download the output file before it expires:

GET /api/jobs/{job_id}/output

Returns the binary file with the appropriate Content-Type header (application/vnd.openxmlformats-officedocument.wordprocessingml.document or .presentationml.presentation).

---

Streaming Progress

Connect to the SSE stream to observe generation in real time:

GET /api/jobs/{job_id}/stream

event: start
data: {"job_id": "job_01j2kqx8v", "skill": "generate-docx"}

event: delta
data: {"text": "Drafting section: Introduction"}

event: delta
data: {"text": "Drafting section: Background and Standards"}

event: tool_call
data: {"tool": "write_file", "path": "document.json"}

event: stop
data: {"job_id": "job_01j2kqx8v", "document_id": "doc_01j2kqx8v", "duration_ms": 18340}