What Are Vectors and Why They Power Modern AI

01 What Is a Vector?

In mathematics a vector has direction and magnitude, and in computing it can be seen as a set of meaningful numbers that represent the features of an object. For example, a bird could be encoded as [red: 0.913, singing: 0.823, size: 0.534], turning descriptive attributes into a numeric vector.

02 What Can Vectors Do?

Vectors allow computers to measure similarity between items. By comparing the distance between two vectors, the system can infer how alike the underlying objects are. For instance, an apple [red: 0.92, sweet: 0.83, round: 0.78] and a strawberry [red: 0.85, sweet: 0.75, round: 0.62] have close vectors, indicating similar characteristics.

Apple: [red: 0.92, sweet: 0.83, round: 0.78] Strawberry:

[red: 0.85, sweet: 0.75, round: 0.62]

03 Vector Similarity vs. Keyword Search

Traditional keyword search matches exact terms, so a document containing the words "AI" and "doctor" is retrieved, while another that mentions "machine learning" but not the keywords is ignored. Semantic search converts queries and documents into vectors, allowing the system to find documents whose meaning is close to the query even without exact keyword overlap.

04 How to Measure Vector Similarity

Similarity is usually computed by measuring the distance between vectors; smaller distance means higher similarity. A common method is cosine similarity. Example vectors:

Text 1: [0.82, 0.61, 0.97] Text 2: [0.90, 0.73, 0.98] Using the cosine formula, the similarity might be 0.99, indicating the two texts are almost identical in meaning.

05 Who Computes Vectors?

Embedding models are responsible for turning complex data—text, images, audio—into numeric vectors. For example, the sentence "I like watching movies." might become [0.7, 0.2, 0.5, 0.9], capturing the key concepts "like" and "movies".

Similarly, an image of a dog could be encoded as [0.8, 0.6, 0.4, 0.7], representing visual features such as shape and color.

06 Who Stores and Compares Vectors?

Vector databases (e.g., Milvus, ChromDB, Pinecone) are specialized systems that store high‑dimensional vectors and enable fast similarity search. Unlike traditional relational databases that store structured fields, vector databases handle "numeric feature" data such as image or document embeddings.

07 Where Are Vectors Applied?

One prominent use case is Retrieval‑Augmented Generation (RAG). Documents are split into chunks, each chunk is embedded into a vector, and a user query is also embedded. By comparing the query vector with chunk vectors (using cosine similarity), the most relevant chunks are retrieved and fed to a large language model to generate an answer.

Example chunk vectors:

Paragraph 1: [0.1, 0.7, 0.2] Paragraph 2: [0.9, 0.1, 0.4] Paragraph 3: [0.3, 0.8, 0.2] The query "What is the principle of C‑RAG?" might be encoded as [0.2, 0.8, 0.1]. The system finds the closest paragraph vectors, retrieves the relevant text, and the LLM answers the question.

08 Limitations of Vectors

High computational cost : Vectors often have hundreds or thousands of dimensions, making similarity calculations resource‑intensive.

Potential representation error : Poorly trained embedding models can produce vectors that do not accurately capture the underlying data.

Storage demands : Storing millions or billions of high‑dimensional vectors requires optimized storage and indexing strategies.

Despite these challenges, vectors remain a fundamental tool that enables modern AI systems to understand and compare complex data across text, images, and beyond.