Architecture

Workflow

1. Upload a PDF and submit conversion parameters
2. The API creates a job_id, stores metadata, and queues the job
3. The worker performs parsing, OCR, image-region detection, and page reconstruction
4. The system exports output.pptx, and the client downloads it after status tracking

Architecture Overview

flowchart LR
    User[Browser User] --> Web[Next.js Web]
    Web -->|/api/v1/*| API[FastAPI API]
    API --> Redis[(Redis / RQ Queue)]
    Redis --> Worker[RQ Worker]
    API --> Jobs[(Job Storage)]
    Worker --> Jobs
    Worker --> OCR[OCR / Parser Providers]
    Web -->|track / download| API

MCP in the Overall System

In addition to the browser flow, the current system can also be exposed to MCP clients through ppt-mcp.

flowchart LR
    Browser[Browser User] --> Web[Next.js Web]
    Web --> API[FastAPI API]
    API --> Redis[(Redis / RQ Queue)]
    Redis --> Worker[RQ Worker]
    MCPClient[MCP Client] --> PPTMCP[ppt-mcp]
    PPTMCP --> API

The responsibility split is:

• the main PDF2PPT service handles parsing, OCR, job orchestration, and PPT generation
• ppt-mcp does not reimplement conversion logic; it wraps the existing API as MCP tools
• browser users normally go through the Web UI, while AI clients normally go through ppt-mcp

Key Concepts

Concept	Meaning	Typical options
Parse Engine	Main document-processing pipeline	local_ocr remote_ocr baidu_doc mineru_cloud
OCR Provider	OCR / page-understanding backend	aiocr tesseract paddle_local baidu
AIOCR Chain	Remote vision-recognition mode	direct layout_block doc_parser
Scanned Page Mode	Export strategy for scanned pages	fullpage segmented

Job Directory and Artifacts

Typical job directory:

api/data/jobs/<job_id>/
├── input.pdf
├── ir.parsed.json
├── ir.ocr.json
├── ir.json
├── output.pptx
└── artifacts/
    ├── ocr/
    ├── image_regions/
    ├── page_renders/
    ├── image_crops/
    └── final_preview/

By default:

• jobs and directories are retained for 24 hours
• they are then removed by the cleanup daemon
• if debug artifacts are disabled, the system keeps only the necessary result files where possible