To search for similar products with embeddings, you usually do this:

1) Convert each product into text

Create a text representation that captures the product well, for example:

• title
• brand
• category
• key attributes
• short description

Example:

Nike Air Zoom Pegasus 40 running shoes, men's, black, size 10, breathable mesh, cushioned sole

2) Generate an embedding for each product

Use an embedding model to turn each product text into a vector, e.g. 1536 dimensions.

You do this for:

• every product in your catalog
• the user’s search query

3) Store product embeddings in a vector database

Common options:

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

Store:

• product_id
• embedding vector
• metadata like price, category, availability

4) Embed the search query

If a user searches:

lightweight black running shoes for men

Generate an embedding for that query using the same model.

5) Find nearest neighbors

Compare the query embedding to all product embeddings using a similarity metric:

• cosine similarity
• dot product
• Euclidean distance

Return the top-k most similar products.

---

Basic flow

Offline indexing

1. Build text for each product
2. Create embeddings
3. Save embeddings in vector DB

Online search

1. Convert user query to embedding
2. Search nearest vectors
3. Optionally filter by metadata
4. Rank and return results

---

Example in Python-like pseudocode

# indexing
for product in products:
    text = f"{product['title']} {product['brand']} {product['category']} {product['description']}"
    product['embedding'] = embed(text)
    vector_db.upsert(product_id=product['id'], vector=product['embedding'], metadata=product)

# search
query = "lightweight black running shoes for men"
query_vec = embed(query)
results = vector_db.search(vector=query_vec, top_k=10)

---

Good practices

Use rich product text

Include the important attributes that affect similarity:

• category
• brand
• color
• material
• use case
• size/specs
• tags

Normalize product text

Make formatting consistent so embeddings are more reliable.

Combine semantic search with filters

Example:

• semantic similarity for “similar products”
• filters for category = shoes, price < 100, in_stock = true

Re-rank results if needed

You can use a second-stage ranking model or business rules to improve relevance.

---

Example query strategies

“Search by example”

User clicks a product and wants similar items:

• embed the clicked product
• search nearest neighbors

“Search by text”

User types natural language:

• embed the query
• retrieve matching products

“Hybrid search”

Combine:

• keyword search
• embedding similarity

This often works better than embeddings alone, especially for exact attributes like model numbers or sizes.

---

A simple product similarity formula

If vectors are normalized, cosine similarity works well:

similarity = dot(query_vector, product_vector)

Higher score = more similar.

---

If you want, I can also show:

• a full code example with Python + pgvector
• how to do this with OpenAI embeddings
• how to build a “similar products” endpoint in an API