mattfanto / feddit-sentiment-analysis

A simple exercise used as a interview challenge

Feddit Sentiment Analysis

A web API predicting if comments on a given subfeddit or category are positive or negative. The project is implemented with FastAPI and Docker.

How-to-run

The webapp sentiment analysis model relies on external API for the classification. By default, a simple mocked model is enabled, if you don't care about the model just run:

docker-compose up -d --build

If you want to use the external API you have 2 options:

• AWS Comprehend (recommended)
• OpenAI ChatGPT (not recommended)

The OpenAI approach is an alternative if you don't work with AWS. Please keep in mind the following limitation for the OpenAI approach:

• I didn't implement input sanitization
• I didn't optimize the prompt
• is much more expensive
• is 2 times slower

Depending on your preferences and possibility (e.g. you may not have AWS account available), you can choose one model or the other via ENV configuration.

• OpenAI approach:

SENTIMENT_MODEL=openai OPENAI_API_KEY=<your-openai-key> docker-compose up --build

• AWS Comprehend approach:

PREDICTIONS_MODEL=aws_comprehend \
AWS_ACCESS_KEY_ID=<your-aws-access-key-id> \
AWS_SECRET_ACCESS_KEY=<your-aws-secret-access-key> \
docker-compose up --build

API

Check http://0.0.0.0:8000/docs for OpenAPI documentation.

Monitoring

To monitor the application I provided 2 tools:

• JSON logger, printing to stdout and eventually collected by a log collector
• Prometheus' metrics, exposed at http://localhost:8000/metrics and accessible at http://localhost:9090

Prometheus is more a proof of concept, I didn't implement any alerting or dashboard. In general if available there are better tools than Prometheus covering better distributed tracing aspect like NewRelic, Sentry, DataDog, Elastic APM, etc.

Tests

For the sake of simplicity and time saving I created only 2 tests:

• Sample smoke test running against a live service
• Simple unit test for the AWS model