Book Rating Prediction API is a FastAPI-based service that predicts a numerical rating (e.g., 1 to 5) based on a given user review of a book. It uses a fine-tuned ModernBERT model exported to ONNX for efficient inference. The API provides endpoints for health checks and real-time rating predictions from raw text inputs.

 ```Review-Rating-Prediction-API/
├── app/
│   ├── __init__.py
│   ├── BERT_regressor.py       # Loads the ONNX model and handles predictions
│   ├── config.py			    # Contains configuration
│   ├── exception_handler.py   #exception handling logic for FastAPI
│   ├── main.py # FastAPI entry point with endpoints
│   ├── model.py # Defines the machine learning model wrapper
│   ├── schema.py # Pydantic data schemas for request and response validation
│   └── example/
│       └── client.py # Example client for testing the API
├── notebooks/
│   ├── convert_onnx.ipynb # Converts the trained PyTorch model to ONNX format
│   ├── expand_dataset.ipynb # Data augmentation
│   ├── ML.ipynb # Classic ML model experiments
│   └── ModernBERT.ipynb # Fine-tuning and evaluation of the ModernBERT 
├── model/
│   ├── bert_regressor.onnx # Exported ONNX model used for inference
│   └── tokenizer/ # Tokenizer files corresponding to the ModernBERT 
├── requirements.txt # Python dependencies for the project
├── .gitignore
└── README.md``` 

• python -m venv env

• Windows .\.venv\Scripts\Activate.ps1
• macOS/Linux source .venv/bin/activate

pip install -r requirements.txt

Set the GLOBAL_CONFIG variable as follows config.py:

 
GLOBAL_CONFIG = { 
"MODEL_PATH": "../model/bert_regressor.onnx", 
"BERT_MODEL": "answerdotai/ModernBERT-base", 
"MAX_LEN": 2048, 
"DEVICE": "cpu" 
}  
 

uvicorn main:app --reload

FastAPI automatically provides a Swagger-based user interface. You can test the API using this interface

1. Start the API server: uvicorn main:app --reload
2. Go to the following address from your browser: http://localhost:8000/docs
3. From the Swagger interface: -You can see if the model is ready by testing the /health endpoint. -Select the /predict endpoint, enter a text in the reviewText field and press ‘Execute’ to get the prediction.

You can send requests to API endpoints using the curl command from the command line.

• /predict Endpoint

  • curl -X POST http://localhost:8000/predict \ -H "Content-Type: application/json" \ -d "{\"reviewText\": \"This book is amazing!\"}" 

• /health Endpoint

  • curl http://localhost:8000/health

You can send a request to the API via PowerShell.

• (Invoke-WebRequest -Uri "http://localhost:8000/predict" -Method POST -Headers @{ "Content-Type" = "application/json" } -Body '{"reviewText":"This book is amazing!').Content | ConvertFrom-Json | Select-Object -ExpandProperty rating

You can call the API programmatically using Python and the requests library:

1. Start the API server: uvicorn main:app --reload
2. Run python client.py

• URL: /health
• Method: GET
• Description: It is used to check whether the model is loaded.
• Example Request: curl http://localhost:8000/health
• Success Response: Status Code: 200 OK { "rating": 0.0 }
• Error:

{ 
"error": true, 
"message": "ONNX model session not initialized", 
"traceback": "..." 
}  

• URL: /predict
• Method:: POST
• Description: Estimates a score based on the given review text.
• • Example Request:

 curl -X POST http://localhost:8000/predict -H "Content-Type: application/json" -d "{\"reviewText\": \"This product is amazing! The quality is top notch.\"}"

• Success Response: Status Code: 200 OK

{
"rating": 5.0
} 

• Error: 400 Bad Request: If the review text is missing or incorrect.

{
"error": true,
"message": "Invalid request data. Please check the 'reviewText' field."
}

• Validation Errors

FastAPI automatically throws this error when the API client sends invalid data (for example, a missing or incorrect type of field).

Example Cases:

• If the reviewText field is missing in the POST /predict request
• If the wrong data type was sent (e.g. reviewText: int)

• Internal Server Error

Failure to load the model while the application is running, NoneType errors, exception in the predict function of the model, unexpected Python errors such as file access.

Example Cases:

• If there is no model file
• model_wrapper.session does not exist (NoneType error)
• if there is an error inside the predict() function (e.g. the tokeniser is not loaded properly)

• FileNotFoundError

Thrown if the file being accessed on the file system does not exist.

Example Cases:

• File not found when trying to load model file (Model(model_path=...)).
• File path access error received from the user

• TimeoutError

This error occurs if the asynchronous operation is not completed within a certain time.

Example Cases:

• An estimate or data processing that takes too long.
• Connection timeout when receiving data from external services.

• RuntimeError

The program encounters an unexpected situation at runtime.

Example Cases:

• The model could not be initialised (model_wrapper.session is None).
• The ONNX session could not be created or there is missing configuration.