Case study
PixelQuery
A text-to-image retrieval system for fashion: natural language in, ranked product images out, powered by CLIP embeddings and FAISS.
Problem
Keyword search fails when shoppers describe style, fabric, or vibe instead of SKU language. The system needed semantic retrieval that still respected catalog filters and could compare embedding models.
Approach
- Encode queries and product images with OpenAI CLIP (ViT-B-32 / ViT-L-14).
- Index vectors in FAISS for fast approximate nearest-neighbor search.
- Expand queries with related terms; filter by category, color, price, material.
- Re-rank by similarity plus popularity / featured signals; compare models side by side.
Result
Interactive retrieval that maps free-form language onto a fashion catalog with tunable confidence thresholds and multi-model comparison for experiment loops.
Stack
Python · CLIP · FAISS · Vector search · Metadata filters · Query expansion