Vector
A set of numerical values that represents the meaning of a term or product. Two vectors that are close together indicate related concepts, which is how a search engine connects "comfortable" with "ergonomic" even though the words look nothing alike.
Definition
Definition
A vector is an ordered list of numbers that represents something, in search, typically a word, a query, a product, or an image. In the context of semantic search and recommendation systems, vectors are the output of embedding models: each item is converted into a fixed-length array of floating-point numbers (typically 256 to 1,536 values) that encodes its meaning.
Two vectors are "close" if the items they represent are semantically similar, measured using distance metrics like cosine similarity. A vector for "laptop" and a vector for "notebook computer" will be close; a vector for "laptop" and a vector for "coffee table" will be far apart.
What it's used for
What it's used for
Vectors make it computationally feasible to compare meaning at scale. Checking whether two words are related by looking up a rule table is slow and incomplete.
Comparing the cosine distance between two 512-dimensional vectors is fast and generalizes to any language, any domain, and any combination of terms, including ones the system has never explicitly encountered before. This is what allows a search engine to match "ergonomic office chair" with a query for "back pain home office" without a synonym rule ever being written.
Doofinder
Doofinder computes vectors for products during indexing using language models trained on eCommerce-relevant data. At query time, the user's query is converted to a vector and the engine retrieves the products whose vectors are closest, the semantic nearest neighbors.
This vector retrieval runs in parallel with or as a complement to keyword retrieval, with the final results blending signals from both approaches.
Example
Case study
Two queries arrive at the same store: "lightweight men's dress shoe" and "formal footwear low weight men". Both express the same intent using completely different words.
Vector search computes embeddings for both queries and finds that they land near the same region of the vector space, close to products tagged as men's formal shoes with attributes indicating lightweight construction. Both queries return the same relevant set of products.