Welcome to Find a Fragrance, an intelligent fragrance recommendation engine that understands what you're looking for — even when you don't know the name. This project uses semantic search, powered by AI-generated vector embeddings, to help users find fragrances based on style, mood, and descriptive intent.
Instead of relying on keyword matching, Find a Fragrance uses semantic embeddings to represent both user queries and fragrance profiles in high-dimensional vector space.
User query:
"I'm looking for something fresh and woody, but not too strong."
Even if those exact words don’t appear in any fragrance description, the system understands the intent and recommends colognes that match the vibe — like crisp green notes or light cedar blends.
User input is passed into a Sentence Transformer model (e.g., all-MiniLM-L6-v2), which converts text into a vector embedding.
Each fragrance in our database has a precomputed embedding, generated from its description, notes, and accords.
Cosine similarity is used to compare the user’s embedding against every stored fragrance vector.
The top matches are returned based on highest similarity — with the option to include filters like price, gender, or season.
Semantic search provides a human-like way to find fragrances. It's not about guessing the right keywords — it's about expressing what you want and letting AI interpret it.
- All results displayed: When looking at specific brands, or scents based off a season, all available fragrances are displayed. Adding a 'see more' button is desired for a more user-friendly UI in the front-end application.
- In-memory similarity: Not optimized for very large datasets.
- No hybrid filtering yet: Can’t combine vector search with metadata filtering (e.g., price or season).
- Embeddings are static: Based only on initial descriptions.
- Explainability is low: Hard to show why a result is relevant.
- ✅ Integrate pgvector or FAISS for scalable vector indexing
- ✅ Support hybrid search (e.g., vector + filter by season/gender)
- ✅ Use multi-modal embeddings (notes + tone + metadata)
- ✅ Rerank results with user behavior (clicks, favorites)
- ✅ Add confidence or similarity scores to increase transparency
- Backend: Python + Flask
- Embedding Model: Sentence Transformers
- Vector Math: Numpy + Cosine Similarity
- Frontend: HTML/CSS/JS/Bootstrap
git clone https://github.com/your-username/Find a Fragrance.git
cd Find a Fragrance
python3 -m venv venv
source venv/bin/activateRun locally:
flask run
Note: .env data is not included for the database configuration (secrey key), API to populate database, or links in the footer.
Have fragrance data, design ideas, or feature suggestions? Pull requests and issues are welcome!