How do I use vector search for recommendations?

Vector search is a great way to build recommendations when you can represent items and users as embeddings. The core idea is:

• turn each item into a vector
• turn the user’s current interests into a vector
• find the item vectors closest to the user vector

Basic workflow

1. Create embeddings for items

For each product, article, movie, song, etc., generate an embedding from:

• title
• description
• tags
• category
• metadata
• user-generated text like reviews

Example:

• “wireless noise-canceling headphones”
• “over-ear bluetooth headset with ANC”

These should end up near each other in vector space.

2. Build a user/query vector

You can create a vector for a user from:

• items they clicked
• items they purchased
• search terms
• profile data
• session behavior

Common approach:

• average the embeddings of items they liked
• weight recent interactions more heavily
• combine text from search/history into one embedding

3. Store embeddings in a vector database

Use a vector index such as:

• FAISS
• Milvus
• Pinecone
• Weaviate
• pgvector in Postgres
• Elasticsearch/OpenSearch vector search

4. Retrieve nearest neighbors

For a given user vector, run nearest-neighbor search to find top-k similar items.

This gives you:

• “users who like X also like Y” style recommendations
• personalized “for you” results
• similar item recommendations

5. Filter and rerank

After vector retrieval, apply business rules:

• exclude already purchased items
• filter out out-of-stock items
• enforce freshness/diversity
• rerank using a separate model or heuristics

Simple example

If a user liked:

• “The Martian”
• “Project Hail Mary”

You embed those books, average the vectors, and search for nearby books. Results might include:

• “Artemis”
• “Red Mars”
• “Seveneves”

Common recommendation patterns

Item-to-item recommendations

“Because you viewed this, you might like that.”

• Embed each item
• Find nearest neighbors for the item

User-to-item recommendations

“Recommended for you.”

• Build a user embedding from history
• Search across the item catalog

Session-based recommendations

Use the last few clicks in the current session to build a temporary vector and recommend relevant items.

Why vector search works well

It captures semantic similarity beyond exact keywords:

• “sneakers” and “trainers”
• “laptop” and “notebook computer”
• “mindfulness meditation” and “stress relief breathing”

Important caveats

Vector search alone is usually not enough for production recommendations:

• it may over-recommend similar items
• it doesn’t automatically handle popularity, freshness, or constraints
• it can miss collaborative signals unless embeddings are built from behavior
• it may not be ideal for very sparse catalogs without good features

Best practice

Use vector search as the candidate generation step, then combine it with:

• collaborative filtering
• popularity signals
• business rules
• ranking models

If you want to implement it

A typical stack looks like:

1. Embed items with an embedding model
2. Index vectors in a vector DB
3. Build user/session vectors from interactions
4. Search top-k nearest items
5. Filter and rerank
6. Log clicks/purchases to improve embeddings

If you want, I can show you:

• a concrete Python example using FAISS or pgvector
• how to build user embeddings
• how to combine vector search with collaborative filtering