NFT Visual and Semantic Search
Finding similar items when the right answer does not share the same words.
Built an image search system over marketplace collections. It collected images and metadata, represented images and text as vectors, stored the index, and returned similar items by distance rather than relying only on names or keywords.
Marketplace collections are difficult to search by title alone. Two images can look related while using completely different names, and one collection can contain many variations of the same idea.
Collecting the material
The project used marketplace and API sources to collect images, collection information, and lookup metadata. The scraping work was part of the search system because the index is only useful when the underlying collection is organized and traceable.
Searching by meaning and appearance
Image and text embeddings were placed in a searchable index. A query could then return nearby items based on visual or semantic similarity, making the result useful even when the query and the stored item used different words.
selected tools
- Python
- Embeddings
- Vector search
- Selenium
- Marketplace APIs