What is a vector embedding?

An embedding is a list of numbers that captures the meaning of text. Similar texts get similar vectors, so you can measure closeness mathematically, which powers search, clustering, and retrieval.

Which Ollama model should I use for embeddings?

Use a dedicated embedding model such as nomic-embed-text . It is small and fast, and it is built for embeddings rather than chat, so it produces better vectors than a general chat model.

What is the difference between /api/embed and /api/embeddings?

/api/embed is current: it takes an input field (a string or array) and returns embeddings , supporting batches. /api/embeddings is legacy, takes a single prompt , and returns embedding .

Why generate embeddings locally instead of in the cloud?

Local embedding keeps text on your machine, so sensitive documents or proprietary code never reach a third-party server. It also has no per-call cost and works fully offline once the model is pulled.

Build / AI / Ollama

Local Vector Embeddings

Generate vector embeddings locally with the /api/embed endpoint and models like nomic-embed-text.

TL;DR

Turn text into vectors with the /api/embed endpoint.
Use a small model like nomic-embed-text for embeddings.
Batch a string or array of texts to /api/embed.

Jump to section

Embedding Basics
Generate Vectors
Python
Why Local
Tips
Warnings
In Practice
FAQ

Embedding Basics

Pull A Model

Download a dedicated embedding model.

ollama pull nomic-embed-text

/api/embed

Send text and get back a vector array.

curl http://localhost:11434/api/embed -d '{
  "model": "nomic-embed-text",
  "input": "The sky is blue"
}'

Vectors

The response holds a numeric array per input.

# { "embeddings": [[0.01, -0.2, ...]] }

Generate Vectors

Single Input

Pass one string to embed it.

"input": "one sentence to embed"

Batch Inputs

Pass an array to embed many at once.

"input": ["first text", "second text"]

Response Shape

Each input maps to one vector in order.

# embeddings[0] matches input[0]

Python

ollama.embed

Call the embed helper with text.

r = ollama.embed(
    model="nomic-embed-text", input="hi")

Read Vectors

The result exposes the embeddings list.

vec = r["embeddings"][0]

Batch

Pass a list to embed several texts.

ollama.embed(model="nomic-embed-text",
    input=["a", "b"])

Why Local

Privacy

Text never leaves your machine to be embedded.

# No cloud call, no data egress

No Per-Call Cost

Local embedding has no usage billing.

# Unlimited embeds after download

Offline

Works with no internet once pulled.

# Index private data fully offline

Tips

Pull a dedicated embedding model like nomic-embed-text; it is small, fast, and purpose-built for turning text into vectors.
Embed many texts at once by passing an array to input, which is far faster than one request per chunk in a loop.

Warnings

Use the same embedding model for documents and queries; mixing models produces vectors you cannot compare meaningfully.
Legacy /api/embeddings uses a prompt field and returns embedding; current /api/embed uses input and returns embeddings.

In Practice

Generate embeddings for two sentences in one batch call, then compare them with cosine similarity.

One batch call embeds both sentences in a single request.
Each vector comes back in the same order as the inputs.
Cosine similarity measures how close the two vectors are.
Similar sentences score near 1.0; unrelated ones near 0.

import ollama, numpy as np

r = ollama.embed(
    model="nomic-embed-text",
    input=["I love cats", "I adore cats"],
)
a, b = r["embeddings"]
a, b = np.array(a), np.array(b)

cos = a @ b / (np.linalg.norm(a) *
               np.linalg.norm(b))
print(round(float(cos), 3))