Traditional product recommendations often depend on rules.
For example:
Customer bought shoes
↓
Show other shoes
AI-powered recommendations can go further.
Instead of matching only categories, a system can compare the semantic characteristics of products.
What Does Semantic Recommendation Mean?
Suppose a customer is looking for:
“Lightweight laptop for programming and travel”
The system can find products with similar characteristics even if the product descriptions use different wording.
Product Embeddings
Each product can be converted into a vector.
For example:
Product Description
↓
Embedding Model
↓
Product Vector
Store that vector alongside the product information.
Search by Similarity
When a user searches for something, the query is also converted into a vector.
The system then compares the query vector with product vectors.
User Query
↓
Vector
↓
Similarity Search
↓
Relevant Products
What Should Be Embedded?
Depending on the product, you might combine:
- Name
- Description
- Category
- Features
- Tags
The goal is to create a useful semantic representation.
Combine Vectors With Filters
Vector similarity alone is not always enough.
Suppose a customer searches for a laptop under $1,000.
Your system can combine:
Semantic Similarity
+
Price Filter
+
Availability
This gives more practical recommendations.
Store Product Metadata
A product record might contain:
id
name
description
price
category
stock
embedding
This keeps product information and vectors together.
Recommendation Example
Imagine a customer views:
Travel Backpack for Laptop
Your system can search for products with similar semantic characteristics.
It might return:
- Lightweight laptop backpack
- Waterproof travel backpack
- Business travel bag
The recommendation is based on similarity rather than only an exact category match.
Watch Performance
As the number of products increases, vector searches can require more database resources.
Monitor:
- Query speed
- Vector index size
- CPU
- RAM
- Database storage
Add User Behavior Later
You can make recommendations more sophisticated by combining semantic similarity with user behavior.
For example:
Product Similarity
+
Previous Views
+
Purchases
+
Wishlist
This can create a more personalized recommendation system.
Final Thoughts
Vector search does not replace normal product filtering.
It adds another layer.
Supabase PostgreSQL with vector capabilities can provide a practical foundation for experimenting with semantic recommendations while keeping product data and vector data close together.