I'm building product recommendations with embeddings, what should I use?

For product recommendations with embeddings, the best choice usually depends on what you’re recommending from and what data you have. In practice, most teams use a hybrid approach:

Good default

Use two towers / dual-encoder embeddings for retrieval:

• one embedding for the user/query context
• one embedding for the product/item
• then retrieve similar items with nearest-neighbor search

This is a strong starting point because it scales well and works for:

• “users like this”
• “similar products”
• “recommended for you”
• search-to-product matching

What to use for item embeddings

A product embedding should combine:

• title
• description
• category / taxonomy
• brand
• attributes (size, color, material, etc.)
• maybe image embedding if visuals matter

Common options:

1. Text embeddings for product metadata

   • easiest and often enough to start
   • works well if titles/descriptions are rich

2. Multimodal embeddings

   • combine text + image
   • better for fashion, furniture, home goods, etc.

3. Learned collaborative embeddings

   • from user-item interactions
   • good when you have lots of click/purchase data

Best recommendation setup

A strong production architecture is:

1. Candidate generation

   • use embeddings + ANN index (FAISS, ScaNN, Pinecone, Weaviate, Milvus, etc.)
   • fetch top 100–1000 candidates fast

2. Ranking

   • use a separate ranker with more features:
     • price
     • availability
     • popularity
     • user affinity
     • recency
     • business rules
   • often a gradient-boosted model or neural ranker

If you have limited data

Start with:

• pretrained text embeddings for products
• average user embeddings from clicked/bought items
• ANN retrieval
• simple re-ranking

This gives a fast baseline.

If you have lots of interaction data

Train:

• matrix factorization / collaborative filtering embeddings
• or a two-tower retrieval model
• optionally fine-tune with content features

This usually beats pure content embeddings for personalization.

My practical recommendation

If you’re unsure, start with:

• product text embeddings
• user embedding = average of recently clicked/purchased product embeddings
• ANN search for retrieval
• then add behavioral/collaborative training later

That’s simple, scalable, and usually good enough to launch.

If you want, I can recommend a concrete stack based on your case:

• e-commerce catalog size
• whether you have click/purchase data
• text only vs image + text
• real-time vs batch recommendations