What Are Embedding Models? API Pricing and Practical Uses

How are embeddings used in search and RAG?

Split documents into meaningful chunks, generate an embedding for each chunk and store the vectors with document identifiers and permissions. When a user asks a question, embed the query and retrieve relevant chunks.

For retrieval-augmented generation, pass selected text to a language model along with the question. The embedding model helps choose context; the language model writes the answer. Evaluate retrieval and generation separately so you can locate failures.

What do embedding APIs charge for?

Tokely documents input-token billing with no output charge for its embedding endpoint. Its listed rates per million input tokens are $0.01 for text-embedding-3-small, $0.065 for text-embedding-3-large and $0.05 for text-embedding-ada-002. Tokely embedding pricing.

At those rates, embedding 100 million tokens costs $1, $6.50 or $5 respectively. These calculations exclude your vector database, document processing, query traffic and later regeneration of changed content.

How do you generate embeddings through Tokely?

With the OpenAI Python SDK installed and TOKELY_API_KEY configured on your server:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["TOKELY_API_KEY"],
    base_url="https://api.tokely.me/v1",
)
result = client.embeddings.create(
    model="text-embedding-3-small",
    input=["How to reset an API key", "Understanding token pricing"],
)
for item in result.data:
    print(item.index, len(item.embedding))
print(result.usage)

This example follows the documented embeddings endpoint; it is not a live test result. Respect the endpoint’s per-input and request limits when batching documents.

Does a larger vector always improve search?

No. Quality depends on the model, language, document domain, chunk boundaries and retrieval setup. Higher dimensionality also increases storage and computation. Compare candidates using labeled queries and relevant-document judgments.

Use the same model and compatible settings for indexed documents and queries. Vectors from different embedding models generally do not share a meaningful coordinate system, even if their dimensions match. A model migration usually requires a new index or an explicit compatibility strategy.

How much vector storage might you need?

One million vectors with 1,536 float32 values each require about 6.144 GB for the raw values alone, using decimal gigabytes. Index structures, metadata, replicas and backups add overhead. Token pricing is only one part of a search system’s operating cost.

How do you protect customer data?

Apply tenant and permission filters during retrieval. Do not let a shared index return another customer’s document simply because its meaning is similar. Store source versions so updated or deleted content can be reflected in the index.

Use Tokely’s embeddings documentation to choose an enabled model, then evaluate retrieval quality before connecting the results to a generative answer workflow.