Give shoppers a better way to find products: upload a photo, type what they mean, or keep browsing while recommendations adapt to what they view.
Search
Synonym-aware and image-led retrieval
Recommendations
Viewed-product signals become recommendation context
Operations
Bulk backfills plus live product updates
New shoppers see products they are likely to love from the first signals they give.
Flexible Vector Database learns from viewed products, liked items, saved styles, and session behavior to recommend similar products before the shopper has to search again.
Runner
Browsing running gear
Recently viewed
3 product signals
Recommended similar product
Because they browsed running gear
Why this match works
Runner shoes combine cushioning and stability. Products in this family match the shopper's active lifestyle and performance needs.
Chic
Browsing chic styles
Recently viewed
3 product signals
Recommended similar product
Because they browsed chic styles
Why this match works
Chic styles balance elegance and trend. Products in this family feel curated and polished, matching the shopper's refined taste.
Traveller
Browsing travel gear
Recently viewed
3 product signals
Recommended similar product
Because they browsed travel gear
Why this match works
Travel gear prioritises durability and versatility. Products in this family suit the shopper who is always on the move.
Two shoppers, two ways to find the right product. One types what they need. The other uploads a photo of their shiba inu. Flexible Vector Database understands both and returns the perfect match.
Full query: comfortable shoes for rainy days
Match found
Drop a product photo
Match found
Keep embedding generation with Compute, then let Flexible Vector Database handle the catalog work: indexing, image search, semantic search, and recommendations.
Compute side
Capacity, runtime, and generation remain in Compute, so your storefront team does not manage model serving.
Catalog side
The retrieval layer uses generated embeddings to keep product discovery, recommendations, and catalog updates in sync.
You should not need a second bill just to make your catalog searchable. Generate embeddings with the model running in Compute, then let the vector database use them for search and recommendations.
The embedding model runs where capacity is already managed.
Text, image, and product vectors are created on that same deployed model.
Retrieval and recommendation workflows use the generated vectors directly.
Keep expensive urgency out of slow-moving catalog work, and reserve live indexing for the updates that directly affect shopper experience.
Bulk
Use Bulk for seasonal imports, backfills, and nightly catalog syncs that do not need to appear instantly.
Live
Use Live for stock, price, availability, and merchandising changes that should affect search right away.
One retrieval layer across search and recommendations.
Included
Run embeddings in Compute and consume the output directly in retrieval.
Included
Search for shopper language, not only exact catalog labels.
Included
Support visual discovery when the product starts from a look, not a keyword.
Included
Use viewed-product signals to recommend adjacent and complementary items.
Included
Backfill catalogs through a free, non-real-time path built for large imports.
Included
Push urgent product changes into retrieval immediately.
One workflow feeds image search, semantic search, and recommendation surfaces across your storefront.
Start with Compute, index once, and use the same retrieval layer for search, context, and recommendations.