How to Generate Text Embeddings With OpenAI (and Ollama)

What the embeddings API returns

You POST a string and the model returns a vector - a list of floats (768 dims for nomic-embed-text, 1536 for text-embedding-3-small). That vector is the numeric representation of the text's meaning, and it's the input every vector store (pgvector, Pinecone, Chroma, OpenSearch) expects at ingestion time.

Calling the embeddings endpoint

The Ollama /api/embeddings route uses the same request shape as OpenAI's /v1/embeddings, so the pattern is identical:

import requests

EMBED_URL = "http://localhost:11434/api/embeddings"  # or OpenAI's endpoint
MODEL = "nomic-embed-text"  # or "text-embedding-3-small" on OpenAI

def embed_docs(docs):
    pairs = []
    for doc in docs:
        r = requests.post(
            EMBED_URL,
            json={"model": MODEL, "prompt": doc},
            timeout=60,
        )
        r.raise_for_status()
        vector = r.json()["embedding"]   # list of 768 floats
        pairs.append((doc, vector))
    return pairs

The return value - a list of (doc, vector) tuples - is exactly what a vector store's bulk-insert method takes for ingestion.

Switching to OpenAI

Point at OpenAI's endpoint and use input instead of prompt (the only structural difference in the request body):

import openai

client = openai.OpenAI()  # reads OPENAI_API_KEY from env

def embed_docs_openai(docs):
    resp = client.embeddings.create(model="text-embedding-3-small", input=docs)
    return [(doc, item.embedding) for doc, item in zip(docs, resp.data)]

The OpenAI SDK batches the list in one request - much faster than one call per doc for large corpora.

What to watch out for in production

• Batch, don't loop - sending a list of strings in one request is 10x faster than calling the endpoint once per document. The OpenAI SDK handles this for you; with Ollama you need to loop (it doesn't support batch input yet).
• Pin the model version - switching from nomic-embed-text to text-embedding-3-small changes the vector dimensions and invalidates every stored vector. Re-embed the whole corpus when you change models.
• Store the model name alongside the vector - in your DB row or metadata, record which model produced each vector. When you upgrade the model later you know which rows need re-embedding.
• Verify the dimension - print len(vector) after the first call to confirm you're getting the expected size (768 for nomic-embed-text, 1536 for text-embedding-3-small) before inserting thousands of rows.