🤖 AI / ML

Embeddings

Vector Representations · Dense Vectors · Vector Embeddings

Definition

Embeddings are dense numeric vectors of fixed dimensionality that encode objects — products, users, queries — so that geometric distance between vectors reflects semantic similarity.

What an embedding is

A computer does not know that sneakers and running shoes are related concepts. For it to find
semantically similar objects, they have to be represented as numbers — and in a way that puts
similar meanings close together in the number space.

An embedding is a vector, a list of numbers, encoding an object:

Nike Air Max sneakers    → [0.23, -0.11, 0.87, ..., 0.04]  # 128 numbers
Adidas running shoes     → [0.21, -0.09, 0.91, ..., 0.07]  # close
iPhone smartphone        → [-0.85, 0.62, -0.31, ..., 0.55]  # far

The distance between vectors — cosine or Euclidean — reflects semantic similarity. Finding similar
products becomes finding the vectors nearest to the current item’s vector.

How embeddings are trained in e-commerce

From behaviour (Item2Vec). Session sequences are the training data: a shopper viewed [A, B, C,
D]. The model learns to predict the context, so items viewed together receive close vectors. This
behavioural similarity captures patterns invisible in the descriptions.

From text. An LLM or a specialised model encodes the product description into a vector. This
suits semantic search: a query such as lightweight shoes for summer runs finds relevant products
even when that exact wording never appears in a description.

A two-tower model. Separate embeddings for users and for items. Closeness between a user vector
and an item vector is personalized relevance. This is the backbone of modern recommenders.

Applications in e-commerce

Scenario	What is encoded	Task
Similar items	Products	Nearest-neighbour lookup by vector
Semantic search	Query plus products	Matching a query to the catalogue
Personal recommendations	Users plus products	Two-tower matching
Interest profile	Behaviour history	The mean vector of views as a taste profile

Tip: the mean vector of a shopper’s viewed items is a simple and effective taste profile. The
nearest items they have not seen yet make strong recommendations at cold start.

Common mistakes

Training on clicks alone. Clicks are a noisy signal. Fine-tune on purchases or add-to-cart
events, weighting by signal strength.
One embedding set for search and recommendations. These are different tasks: search optimises
relevance to a query, recommendations optimise engagement and conversion.
No retraining. The catalogue changes and behaviour moves. Embeddings trained six months ago do
not reflect seasonal trends or new arrivals.

Frequently asked questions

How are product embeddings produced?

There are several routes. From text — an LLM or a specialised model encodes the product description into a vector. From behaviour — Item2Vec trains on sequences of views and purchases, so items with similar usage patterns receive similar vectors. Hybrid approaches combine textual and behavioural signals.

How many dimensions should an embedding have?

It depends on the task. 32–64 dimensions suit light models for real-time inference. 128–256 balances precision and speed and is typical for e-commerce recommendations. 512–1536 belongs to text and semantic search. More dimensions means more precision but slower search and more memory.

How do you search embeddings quickly?

Exact nearest neighbour search scales poorly. For millions of items, approximate nearest neighbour algorithms such as FAISS, HNSW and ScaNN are used. They give up about 1% of accuracy and return results 10 to 100 times faster. Dedicated vector databases automate that layer.