Vector Embedding
A dense numeric representation of text that captures meaning and intent, so machines can compare content by similarity rather than keywords.
A vector embedding is a dense list of numbers that represents a piece of text as a point in a meaning space. Instead of just matching keywords, embedding models learn to place text so that similar concepts land near each other. This allows machines to compare content by its actual intent.
This is how modern search engines like Google actually read your content. When Google processes a search query or crawls a webpage, it translates that information into embeddings. It then uses these numbers to judge query intent, content relevance, and the relationships between concepts, rather than just looking for literal word overlap.
These embeddings are the foundation of vector search and the retrieval part of retrieval-augmented generation. Google has even opened up its own model, called EmbeddingGemma, which offers a rare window into how its embeddings behave. When working with these models, choices like dimension size and data compression allow you to trade a bit of accuracy for faster speed and lower storage costs.
A vector embedding is a dense list of numbers that represents a piece of text as a point in "meaning space." Embedding models learn to place text so that similar meanings land near each other, letting machines compare content by intent rather than by matching keywords.
This is how modern search actually reads your content. When Google processes a query or crawls a page, it produces embeddings to judge query intent, content relevance, and the relationships between concepts and entities — not just literal word overlap. Embeddings are the input to vector search and the "retrieve" half of retrieval-augmented generation.
Google's open EmbeddingGemma model gives a rare window into how its own embeddings behave, and choices like dimension size and quantisation (see vector embedding optimization) trade accuracy against speed and storage.