Embeddings Feature Guide

Turn text into numeric vectors for semantic search, RAGE, and pgvector storage.

For API reference, see api/embeddings.md. This page covers practical usage patterns.

Recommended Models

ModelDimensionsUse Case embeddinggemma768Google's embedding model qwen3-embedding1024Multilingual all-minilm384Lightweight, fast mxbai-embed-large1024High quality nomic-embed-text768General purpose

CLI

ollama run embeddinggemma "Hello world"
Output: JSON array of floats
echo "Hello world" | ollama run embeddinggemma

Python

import ollama

Single embedding
single = ollama.embed(model='mxbai-embed-large', input='Hello world')
print(len(single['embeddings'][0]))  # 1024

Batch embedding
batch = ollama.embed(
    model='mxbai-embed-large',
    input=[
        'The quick brown fox jumps over the lazy dog.',
        'The five boxing wizards jump quickly.',
        'Jackdaws love my big sphinx of quartz.',
    ]
)
print(len(batch['embeddings']))  # 3

JavaScript

import ollama from 'ollama'

const single = await ollama.embed({
    model: 'mxbai-embed-large',
    input: 'Hello world',
})
console.log(single.embeddings[0].length)  // 1024
const batch = await ollama.embed({
    model: 'mxbai-embed-large',
    input: ['First', 'Second', 'Third'],
})
console.log(batch.embeddings.length)  // 3

Key Facts

Embeddings are L2-normalized (unit-length) — cosine similarity = dot product

Always use the same model for indexing and querying

Batch embedding is more efficient than individual calls

Use truncate: false to get errors instead of silent truncation

mindX RAGE Integration

mindX uses mxbai-embed-large and nomic-embed-text (already installed) for RAGE semantic retrieval:

# Embed agent memory for pgvector storage
memories = ["BDI cycle completed successfully", "Rate limiter triggered on Gemini"]
batch = ollama.embed(model='mxbai-embed-large', input=memories)

Each embedding is a 1024-dim vector ready for pgvector
for i, embedding in enumerate(batch['embeddings']):
    # INSERT INTO memories (content, embedding) VALUES ($1, $2::vector)
    pass

---