Your First Embedding

What we're building

A tiny program that turns one sentence into a vector: a fixed-length list of numbers, produced by an AI model trained specifically to make "meaning" measurable. This is the "embed" stage of Naive RAG's four stages, done in isolation before it's part of anything bigger.

What an embedding actually is

Gemini's chat model, from Lesson 1, takes text in and produces text out. An embedding model is different: it takes text in and produces a list of floating-point numbers out, always the same length no matter how long the input text was. That list is called a vector, and it's positioned in a very high-dimensional space such that two pieces of text with similar meaning end up with vectors that are close together in that space, and two pieces of text with unrelated meaning end up far apart.

Nothing about a vector is human-readable on its own, printing one just shows a list of numbers like -0.0153, 0.0106, 0.0268, .... Its value comes entirely from comparing it to other vectors, which Lesson 3 does next.

The code, piece by piece

response = client.models.embed_content(
    model=EMBEDDING_MODEL,
    contents=[text],
    config=types.EmbedContentConfig(output_dimensionality=EMBEDDING_DIMENSIONS),
)

embed_content is Gemini's embedding endpoint, the sibling of generate_content from Lesson 1. contents takes a list, so you can embed several pieces of text in one network call, this lesson passes just one. EmbedContentConfig(output_dimensionality=768) asks Gemini to return a 768-number vector specifically; left unset, this model's vectors are 3072 numbers long. A smaller, fixed length keeps every vector in this course comparable to each other and to this repo's pgvector course, which uses the same 768 setting.

values = response.embeddings[0].values

response.embeddings is a list (matching the list you passed in as contents), so [0] is "the embedding for the first, and only, piece of text we sent." .values is the actual list of floats.