Moderation api

welcome to moderation_api_py a simple api that uses different AI models to moderate texts and images in a fast and efficient way using FastAPI and Celery.

Celery is used to perform the moderation tasks in the background and store the results in a Redis database.

features:

• Image moderation¹:
  • nsfw score
  • normal score
  • summary
• Text moderation² :
  • toxicity
  • severe toxicity
  • obscene
  • identity attack (refer to detoxify docs)
  • insult
  • threat
  • sexual explicit content
  • summary
• Video moderation :working (WIP)
• Moderation tasks execution queue with celery
• Moderation task monitoring dashboard with flower web portal on http://localhost:5555
• Authentication by api keys middleware is implemented.

how to run :

With docker:

Create a docker-compose.yml file:

services:
  api:
    image: ghcr.io/brahimabd98/moderation_api_py:main
    ports:
      - "8000:8000"
    environment:
      DATABASE_URL: "redis://redis:6379/0"
  redis:
    image: redis:alpine
    ports:
      - "6379:6379"
    volumes:
      - redis-data:/data

docker image :

docker pull ghcr.io/brahimabd98/moderation_api_py:main

locally with GitHub

Start a new terminal session and clone the project with this command :

git clone github.com/brahimabd98/moderation_api_py/

cd into the project :

cd moderation_api_py

Start a new venv (optional):

python -m venv .venv

Install the project's dependencies:

pip install -r requirements requirements.txt

set up the .env file (checkout .env.example)

cp .env.example .env

start the project

fastapi run main.py --port 8000

Redis:

• make sure you have a running Redis db instance, you can check their docs for more info

docker image:

You can use the provided docker image :

docker pull ghcr.io/brahimabd98/moderation_api_py:main

screenshots:

swagger ui http://localhost:8000/docs

flower dashboard http://localhost:5555

task list:

task:

Planned Improvements:

• testing
• task re-execution policy
• rate limiting: limit the amount of request.
• api gateway
• task priority
• grafana and Prometheus integration

Non-goals (subject to change) :

• multiple image processing
• html sanitization
• censorship: apply blur filter or text censorship of certain words
• measuring model performance

credits :

• text moderation : detoxify
• image moderation : Falconsai/nsfw_image_detection

disclaimers :

• ethical considerations: as the author of the different AI models, these kinds of classifications shouldn't be taken at face value and human intervention should present for the final decision.
• performance: yet to be determined.
• Error handling: currently there's no proper error handling, so the server is prone to crashing
• logging: right there's no good structured logging we plan to add in the future

credit: Brahim, 2024

footnotes:

Footnotes

1. detoxify. ↩

2. Falconsai/nsfw_image_detection. ↩