เวกเตอร์ฝังตัว: แนวคิดหลัก

Embeddings are digital fingerprints for your data — they turn words, images, or whole documents into vectors that capture meaning and context.

Computers understand numbers, not raw language or images. Embeddings bridge that gap, giving the machine a numerical description it can actually process.

An embedding lives in a vector space — a coordinate system with many dimensions. Each piece of data is a point, and similar items sit closer together.

A vector is just an ordered list of numbers like [0.1, 0.5, -0.2, 0.8]. Each number captures a learned feature — not usually human-readable on its own.

An embedding model does the conversion: it processes your input and outputs a vector encoding its meaning. Similar meanings produce numerically similar vectors.

Embeddings often have hundreds or thousands of dimensions — think 768 or 1536. More dimensions capture finer nuance in the data's meaning.

Here's the power: data with similar meaning produces vectors that sit close in the space, while unrelated data lands far apart. That proximity drives search.

We measure closeness with distance metrics. Cosine similarity compares the angle between vectors; Euclidean distance measures the straight line between them.

Real embeddings come from complex models, but we can sketch vectors in code — notice how similar concepts get numerically similar values.

Recap: embeddings are numeric vectors of complex data that let computers process meaning. They live in a vector space where proximity means similarity, measured by metrics like cosine.