Text to Vectors

To use text with math, you first have to turn it into numbers. Converting text into a vector — an embedding — is the essential first step for semantic search, RAG, and similarity comparisons. This is how a sentence becomes a fixed list of numbers that captures its meaning, ready to be compared with any other text.

💡 In one line: Text-to-vector conversion runs text through an embedding model to produce a single fixed-length vector that represents its meaning.

The Goal

Take a piece of text — a word, sentence, or paragraph — and produce one fixed-length vector of numbers that represents its meaning. That single vector is what everything downstream (search, clustering, RAG) works with.

How Text Becomes a Vector

The conversion runs through a short pipeline.

Whiteboard
Whiteboard diagram

The model produces a representation for each token, then pools them (e.g. averaging, or using a special summary token) into a single vector for the whole text.

What the Vector Looks Like

The result is a list of floating-point numbers of fixed length — for example a 384-dimensional vector:

[0.021, -0.187, 0.043, 0.256, -0.099, ... ]

Each dimension is a learned feature. The individual numbers aren't human-readable, but the whole vector encodes the text's meaning.

Fixed Length, Regardless of Input

A single word and a whole paragraph both map to a vector of the same dimension. This uniformity is what lets you compare any two texts directly.

Token Embeddings vs. Text Embeddings

  • Token embeddings — one vector per token (the input layer inside an LLM).
  • Text embeddings — those token vectors pooled into one vector for the whole text.

For search and similarity, we use the text embedding — the single pooled vector.

Use the Same Model for Everything

Vectors are only comparable if they come from the same embedding model. Different models produce different, incompatible spaces — so embed everything you plan to compare with one model.

Code Example


 
Practical Notes
 
     
  • Batch-encode many texts at once for speed.
    •  
  • Normalise vectors if you'll compare them with cosine similarity.
    •  
  • Store vectors in a vector database for fast search at scale.
    •  
 
Summary
 
     
  • Text-to-vector conversion produces a single fixed-length embedding for a piece of text.
    •  
  • The pipeline is tokenize → embedding model → pool → vector.
    •  
  • The vector is a list of floats where each dimension is a learned feature.
    •  
  • Any text — word or paragraph — maps to the same dimension, enabling direct comparison.
    •  
  • Always use the same model, and store vectors in a vector database for search. EOF echo created