Embeddings, vectors, dimensions, and semantic similarity
Practice this question🔴 MUST KNOW · Embedding means converting content into numbers that help computers compare meaning.
Overview
- An embedding is a list of numbers representing useful features of content, often its meaning.
- That list is a vector; its number of entries is its dimension.
- Semantic similarity means closeness in meaning, such as “car accident” and “vehicle collision,” even when words differ.
Why this matters
Exact word matching misses paraphrases. Embeddings give a searchable numerical representation of text, images, or other content.
An everyday example
A librarian puts books about similar subjects on nearby shelves even when their titles use different words.
A technical example
An embedding model maps “claim progress” and “claim status” to vectors that tend to be close, helping search return the same relevant article.
How it works
- Select an embedding model
- Encode content
- Receive a fixed-size numerical vector
- Store it with source metadata
- Encode a query with the compatible model
- Compare vectors.
Architecture and flow
Text → embedding model → vector + metadata → searchable index.
The insurance assistant example
A customer asks “Has anyone looked at my accident papers?” and search finds the article about claim review stages.
When to use it
Semantic search, duplicate detection, related-document discovery, and memory retrieval.
When to avoid it
Use exact lookup for a claim number. Do not use vector similarity to decide whether a customer is authorized or a statement is true.
Interview answer
“Embeddings turn content into numerical vectors so we can retrieve semantically related information even when wording differs.”
Benefits, trade-offs, and production considerations
- Each coordinate is generally not a human-labeled feature such as “insurance-ness.” Higher dimension is not automatically better.
- Domain, language, normalization, storage, and retrieval evaluation matter more than choosing the longest vector.