Fashion RAG Search
Multimodal product search across 20K+ H&M items: text or image queries in, matching products and LLM-generated styling advice out.
- ownership
- Independent project
- when
- Sep 2025
- org
- Independent
The film, as text: Fashion RAG Search. Multimodal product search across 20K+ H&M items: text or image queries in, matching products and LLM-generated styling advice out. Shared CLIP space: CLIP embeds text and images into the same vector space. FAISS search: FAISS finds the nearest products in under 0.5s per query. Visual matches: Matches come back with their images and similarity scores. Styling advice: An LLM turns the retrieved items into outfit and occasion suggestions. Results: 90%+ retrieval accuracy and 85% recommendation accuracy. 20K+ H&M products searchable. under 0.5s per query. 90%+ retrieval accuracy. 85% recommendation accuracy. Independent project, Independent, Sep 2025.
How it works
- Multimodal retrieval: CLIP embeddings place text and images in one space, and FAISS similarity search returns matches across both modalities in under 0.5 seconds per query.
- Dataset: 20K+ H&M products prepared with Hugging Face datasets, Pandas and NumPy, with PIL for image handling.
- Styling advice: Groq and LangChain turn retrieved products into styling and occasion-based recommendations, shown with their images.
- Interactive app: a Streamlit interface shows matching products with images, descriptions and similarity scores.
Outcome
20K+
H&M products searchable
under 0.5s
per query
90%+
retrieval accuracy
85%
recommendation accuracy
Stack
- Retrieval
- CLIP
- FAISS
- PyTorch
- Multimodal RAG
- Generation
- Groq API
- LangChain
- Data
- Hugging Face datasets
- Pandas
- NumPy
- PIL
- Application
- Python
- Streamlit