Semantic book recommendation system powered by Sentence-BERT, Streamlit, Airflow, and Google Cloud.

Live app | GitHub profile

Bookwise AI is an end-to-end recommendation platform that lets users describe the kind of book they want in natural language and returns semantically relevant titles. Instead of relying only on keyword matching, the system converts both user intent and enriched book descriptions into dense vector embeddings, then ranks candidates by semantic similarity.

The project demonstrates a complete applied AI workflow: dataset cleaning, metadata enrichment, embedding generation, vector-based retrieval, Streamlit deployment, feedback logging, Airflow-based data asset orchestration, and Terraform-managed Google Cloud infrastructure.

Most simple recommender demos depend on explicit genres, tags, or manual filters. Bookwise AI is designed around a more realistic discovery problem: a user may say "I want a slow, emotional novel about memory and family" without knowing the exact category or title. The system handles that open-ended intent by mapping the query into the same semantic space as the book corpus.

For recruiters and technical reviewers, this project highlights practical strengths across machine learning, data engineering, application development, and cloud-aware deployment.

flowchart LR
    A[GoodBooks Dataset] --> B[Cleaning Notebooks]
    B --> C[Google Books Metadata Enrichment]
    C --> D[Enriched Book Dataset]
    D --> E[Sentence-BERT Embedding Generation]
    E --> F[Precomputed Embedding Matrix]

    G[User Natural Language Query] --> H[Streamlit App]
    H --> I[Sentence-BERT Query Embedding]
    I --> J[Cosine Similarity Search]
    F --> J
    D --> J
    J --> K[Ranked Book Recommendations]
    K --> L[User Feedback]
    L --> M[Google Sheets Logging]

    B --> N[Airflow Upload DAGs]
    C --> N
    E --> N
    N --> O[Google Cloud Storage]
    P[Terraform] --> O
    P --> Q[BigQuery Datasets]

bookwise-ai/
|-- airflow/
|   |-- dags/                 # Airflow DAGs for uploading data artifacts
|   |-- data/                 # Pipeline data used by Airflow tasks
|   `-- docker/               # Local Airflow Docker Compose setup
|-- assets/
|   `-- bookwise-ai-cover.png # README cover image
|-- data/                     # Project data and embeddings
|-- infrastructure/           # Terraform configuration for GCP resources
|-- notebooks/                # Cleaning, enrichment, embedding, and modeling notebooks
|-- streamlit/
|   |-- app.py                # Streamlit recommendation app
|   |-- enriched_data.csv     # App-ready enriched book catalog
|   |-- embeddings.npy        # Precomputed semantic vectors
|   `-- requirements.txt      # App dependencies
|-- LICENSE
`-- README.md

1. Clean the source book dataset and retain the fields that matter for discovery.
2. Enrich each book with descriptive metadata through the Google Books API.
3. Encode book descriptions with Sentence-BERT and persist the embedding matrix.
4. Load the enriched catalog and embeddings in the Streamlit application.
5. Encode the user's query with the same embedding model.
6. Rank books with cosine similarity and return the top recommendations.
7. Log user queries and feedback for future product and model analysis.

The project includes Airflow DAGs for managing key data artifacts:

Terraform provisions the GCP storage and warehouse foundation:

• Google Cloud Storage bucket for versioned data assets.
• BigQuery datasets for raw and clean analytical layers.
• Region and project configuration managed through variables.

git clone https://github.com/sntk-76/bookwise-ai.git
cd bookwise-ai

python -m venv .venv
.venv\Scripts\activate
pip install -r streamlit/requirements.txt

streamlit run streamlit/app.py

On macOS/Linux, activate the environment with:

source .venv/bin/activate

Feedback logging requires Streamlit secrets containing the Google service-account configuration expected by streamlit/app.py.

cd airflow/docker
docker-compose up

Then open the Airflow UI at http://localhost:8080 and trigger the upload DAGs as needed.

cd infrastructure
terraform init
terraform plan
terraform apply

The Terraform configuration expects Google Cloud credentials and project values to be available through the variables in infrastructure/variables.tf.

• Uses semantic retrieval instead of keyword-only matching.
• Separates data preparation, embedding generation, application serving, and cloud upload workflows.
• Keeps inference efficient by using precomputed book embeddings.
• Adds a product feedback loop through Google Sheets logging.
• Demonstrates both ML application delivery and data engineering discipline in one coherent project.

• Replace full-matrix cosine search with FAISS, Annoy, or ScaNN for larger catalogs.
• Add genre, language, rating, and publication-year filters.
• Introduce user sessions and personalized recommendation history.
• Build a feedback analytics dashboard from logged query and satisfaction data.
• Containerize the Streamlit app for deployment to a managed cloud runtime.

This project is licensed under the MIT License.

• GoodBooks-10K Dataset
• Google Books API
• SentenceTransformers
• Streamlit
• Apache Airflow
• Terraform