PowerPoint Spelling & Grammar Corrector

An AI-powered FastAPI service that corrects spelling and grammar in PowerPoint files using Claude Sonnet 4.5 via the Langdock API. Consultants upload a .pptx to Langdock, Langdock’s sandbox sends it to the backend, and the backend returns a corrected file as a base64 payload.

The service preserves formatting at the run level and supports an optional highlighted output mode for reviewed changes. It is designed for Railway deployment in the EU region with in-memory job processing and no file writes during correction.

Architecture

Langdock custom action
  |
  | POST /jobs?highlight=false
  | Body: { file_base64, file_name }
  v
FastAPI backend on Railway
  |
  | Background job
  | - decode base64 into BytesIO
  | - extract text from slides
  | - call Langdock LLM API
  | - preserve formatting and optionally highlight changes
  | - store result in jobs dict
  v
GET /jobs/{job_id}
  |
  | { status: processing | done | error }
  v
Langdock action returns corrected PPTX file

API

GET /health

{ "status": "ok" }

POST /jobs?highlight=false

Headers:

Authorization: Bearer <API_KEY>

Body:

{
  "file_base64": "...",
  "file_name": "presentation.pptx"
}

Response:

{ "job_id": "<uuid>" }

GET /jobs/{job_id}

Headers:

Authorization: Bearer <API_KEY>

Responses:

{ "status": "processing" }

{
  "status": "done",
  "file_base64": "...",
  "file_name": "presentation_corrected.pptx",
  "corrections_count": 12,
  "changes": [
    { "slide": 1, "original": "Teh strategc obiectives for Q3", "corrected": "The strategic objectives for Q3" }
  ]
}

{ "status": "error", "message": "..." }

Local Development

python -m venv venv
venv\Scripts\activate
pip install -r requirements-dev.txt
python -m uvicorn app.main:app --reload --host 0.0.0.0 --port 8000

Environment Variables

Variable	Required	Purpose
API_KEY	Yes	Bearer token required on /jobs endpoints
LANGDOCK_API_KEY	Yes	Langdock API key for Claude calls
LANGDOCK_API_URL	Yes	Langdock completions endpoint
PORT	Yes in Railway	Port for FastAPI / Uvicorn

Optional:

Variable	Purpose
LANGDOCK_MODEL	Override the Claude model if needed

Langdock Integration Setup

1. Create a custom integration in Langdock.
2. Add a file input named document.
3. Add a boolean input named highlight, defaulting to false.
4. Add an API key auth field that stores the backend bearer token.
5. Paste the reference action from langdock/action_correct_pptx.js.
6. Configure the action to call the Railway service URL for POST /jobs and GET /jobs/{job_id}.
7. Set the auth field slug to apiKey, or update the script to use your chosen data.auth.<slug> value.

Important: Langdock documents a 2-minute execution timeout for custom actions. The reference script polls in the sandbox, but it must still fit inside that timeout budget. If a job is still processing when the sandbox limit is reached, the action should return a clear in-progress error and let the user retry.

Railway Deployment

1. Create a Railway service from this repository.
2. Use the root Dockerfile.
3. Set the Railway region to EU West Metal, Amsterdam in the Railway dashboard.
4. Add the runtime variables API_KEY, LANGDOCK_API_KEY, LANGDOCK_API_URL, and PORT.
5. Configure the health check path as /health.

Tests

venv\Scripts\python.exe -m pytest tests/ -v --tb=short

Notes

• Processing is fully in memory.
• No files are written to disk during correction.
• The FastAPI job store is ephemeral; a restart clears pending jobs.
• The backend keeps the original McKinsey-style correction prompt and run-level formatting preservation logic.