Sentiment Classification API (v2 — FastAPI + JWT)

A REST API rebuild of my original Flask customer review sentiment classifier, same trained Naive Bayes model and Indonesian text preprocessing pipeline, now served through a proper REST layer with JWT authentication, request/response validation, auto-generated docs, and Docker support.

Original Flask version (v1, built for my thesis):

This project doesn't retrain or change the model — it re-engineers how the model is served. The dataset, the trained Naive Bayes model, and the 85.6% accuracy result all carry over unchanged from v1.

What's new compared to v1

	v1 (Flask)	v2 (this repo)
Framework	Flask	FastAPI
Auth	None	JWT (register/login)
API style	Informal calls	REST endpoints
Validation	Manual	Pydantic schemas
Docs	None	Auto-generated at /docs
Deployment	Local only	Dockerized

Tech stack

• FastAPI + Uvicorn — web framework and server
• SQLAlchemy + PyMySQL — ORM and MySQL driver
• python-jose + bcrypt — JWT tokens and password hashing
• scikit-learn + joblib — loading the trained Naive Bayes model and vectorizer
• Sastrawi — Indonesian stopword removal
• stemmid — Indonesian stemming

Text preprocessing pipeline

New review text goes through the same cleaning steps used before training, so it lines up with what the vectorizer's vocabulary actually knows:

lowercase -> normalize elongation ("mantapp" -> "mantap") -> remove stopwords (Sastrawi) -> stem (stemmid)

The elongation-normalization step is new in v2 — it collapses repeated letters used for emphasis in informal reviews ("mantapp", "kerennnn") so slang spellings map onto words the model already recognizes, instead of being treated as unknown and ignored. Trade-off: a few real Indonesian words with a genuine double letter ("maaf", "saat") get affected too; in practice this has minimal impact since they aren't strong sentiment words.

Option A — Run locally with your existing Laragon MySQL

1. Start Laragon, make sure MySQL is running, and create a new database — don't reuse your v1 database name, to keep the two projects' data fully separate. E.g. sentiment_db_v2.
2. Copy .env.example to .env and set DATABASE_URL to match (adjust user/password if yours differ from Laragon's default root with no password).
3. Create and activate a virtual environment:
   • Windows: python -m venv venv then venv\Scripts\activate
4. Install dependencies: pip install -r requirements.txt (this includes a GitHub-based install for stemmid, so it needs an internet connection)
5. Run the server: uvicorn app.main:app --reload
6. Open http://127.0.0.1:8000/docs — interactive Swagger UI. Register a user, log in, copy the token into the "Authorize" button, and call /predict right from the browser.

Option B — Run with Docker (no local MySQL needed)

docker compose up --build

Spins up the API and a MySQL container together. Same URL: http://127.0.0.1:8000/docs

Wiring in your trained model

1. From your v1 project, export the trained model and vectorizer once with:

   import joblib
   joblib.dump(model, "naive_bayes_model.pkl")
   joblib.dump(vectorizer, "vectorizer.pkl")

2. Create an ml/ folder in this project (same level as app/).
3. Copy both .pkl files in, keeping those exact filenames.
4. Restart uvicorn — the model loads once at startup. Without these files, /predict still runs using placeholder logic, so the API stays testable at every stage.

API endpoints

Method	Path	Auth required	Description
POST	/auth/register	No	Create an account
POST	/auth/login	No	Get a JWT access token
POST	/predict	Yes	Classify a review, save it to history
GET	/predict/history	Yes	List your past predictions
GET	/docs	No	Interactive Swagger UI

Known limitations

• The Naive Bayes / bag-of-words approach can't generalize to word forms it never saw in training. Elongation normalization helps with slang emphasis, but genuine typos or new vocabulary still won't be recognized.
• Confidence scores near 0.50 indicate the model found little to no recognizable signal in the input — not that the sentiment is uncertain.

Suggested next steps

1. Write tests with pytest + FastAPI's TestClient (register → login → predict → assert).
2. Push to GitHub with a screenshot or GIF of the /docs page.
3. Deploy to Railway or Render (both have free tiers with managed MySQL/Postgres) — point DATABASE_URL at their managed database instead of localhost.
4. Add the live URL and this repo's link to your LinkedIn Projects entry.