Embeddings With Ollama

A different kind of model

Every lesson so far used llama3.2, a generation model: text in, new text out. nomic-embed-text is a different kind of model entirely, an embedding model, it can't chat or answer questions at all. Its only job is to turn text into a fixed-length list of numbers (a vector) that captures the text's meaning, the same concept the langchain course covers in its final RAG lessons and pgvector builds an entire course around, just running on your own machine here instead of calling out to Google.

Setup: pull the embedding model first (small, under 300 MB):

ollama pull nomic-embed-text

The code, piece by piece

response = ollama.embed(model="nomic-embed-text", input=SENTENCES)

embed() takes a list of strings and returns one vector per string, in the same order you passed them in. Passing a list instead of calling embed() once per sentence lets Ollama batch the work internally, which matters once you're embedding hundreds of chunks for a real RAG pipeline (Lesson 21 builds exactly that).

def cosine_similarity(a: list[float], b: list[float]) -> float:
    dot_product = sum(x * y for x, y in zip(a, b))
    norm_a = math.sqrt(sum(x * x for x in a))
    norm_b = math.sqrt(sum(x * x for x in b))
    return dot_product / (norm_a * norm_b)

Cosine similarity measures the angle between two vectors, ignoring their length: 1.0 means they point in exactly the same direction (as similar in meaning as this model can represent), 0.0 means they're unrelated. This is hand-rolled here with plain Python so the math is visible; the pgvector course uses Postgres's <=> operator to do the same computation directly in the database instead.

cosine_similarity(dog_vec, puppy_vec)
cosine_similarity(dog_vec, tax_vec)

Two sentences about pets should score meaningfully higher than a pet sentence compared against a sentence about taxes, even though none of the words overlap between "dog" and "puppy" beyond the general topic. That's the entire point of an embedding: it captures meaning, not just shared words.