Embeddings

SGL Grid serves embedding models alongside chat, reasoning and coding models. An embedding turns a piece of text into a fixed-length vector of floats you can store in a vector database and compare with cosine similarity — the backbone of semantic search and retrieval-augmented generation (RAG). Embeddings are a distinct modality from chat: they use the OpenAI-compatible /v1/embeddings endpoint (not /v1/chat/completions), return a vector rather than text, and are billed on input tokens only (there is no generation).

Endpoint

POST https://grid.x402compute.cc/v1/embeddings

Request body (OpenAI-compatible):

Field	Type	Notes
`model`	string	An embedding model id (see the catalog below).
`input`	string \| string[]	One string, or an array to embed a batch in one call.
`dimensions`	number	Optional. Truncate a Matryoshka model to fewer dims (e.g. nomic `768→256`).
`input_type`	`"query"` \| `"document"`	Optional. Asymmetric-retrieval hint for models that support it.
`tier`	`"standard"` \| `"confidential"`	Optional. Route to any node, or attested nodes only.

The response matches OpenAI’s shape:

{
  "object": "list",
  "data": [{ "object": "embedding", "index": 0, "embedding": [0.01, -0.03, ...] }],
  "model": "nomic-embed-text-v1.5",
  "usage": { "prompt_tokens": 6, "total_tokens": 6 }
}

Vectors are returned in the same order as input and are L2-normalized, so a plain dot product equals cosine similarity.

Model catalog

All models are tiny, CPU-friendly BERT encoders — great even on modest hardware.

Model id	Dimensions	Notes
`nomic-embed-text-v1.5`	768 (Matryoshka → 512/256/128/64)	Recommended. 8K input, strong general retrieval.
`bge-small-en-v1.5`	384	Smallest/fastest English retriever.
`bge-base-en-v1.5`	768	Balanced quality/size.
`bge-large-en-v1.5`	1024	Highest-quality BGE English.
`mxbai-embed-large-v1`	1024 (→ 512/256)	Top MTEB retrieval.
`all-minilm-l6-v2`	384	Classic lightweight sentence encoder.

List what’s live right now (embeddings appear once at least one node serves them):

curl "https://grid.x402compute.cc/v1/models?type=embedding"

Quickstart

curl https://grid.x402compute.cc/v1/embeddings \
  -H "Authorization: Bearer x402c_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nomic-embed-text-v1.5",
    "input": ["The Singularity grid serves confidential embeddings.", "Hello world"]
  }'

from openai import OpenAI

client = OpenAI(base_url="https://grid.x402compute.cc/v1", api_key="x402c_your_api_key")
r = client.embeddings.create(model="nomic-embed-text-v1.5", input=["hello world"])
print(len(r.data[0].embedding))   # 768

from singularity_grid import GridClient

grid = GridClient(api_key="scg_your_api_key")
vecs = grid.embed("nomic-embed-text-v1.5", ["hello world", "bonjour"])
print(len(vecs), len(vecs[0]))     # 2 768

import { GridClient } from "@singularity-layer/grid";

const grid = new GridClient({ apiKey: "scg_your_api_key" });
const res = await grid.embed({ model: "nomic-embed-text-v1.5", input: ["hello world"] });
console.log(res.data[0].embedding.length); // 768

Matryoshka dimensions

nomic-embed-text-v1.5 and mxbai-embed-large-v1 are Matryoshka models — you can ask for a shorter vector to save storage and speed up search, at a small recall cost. The server truncates and re-normalizes:

r = client.embeddings.create(model="nomic-embed-text-v1.5", input="cat", dimensions=256)
len(r.data[0].embedding)  # 256

Query vs document

Some models score asymmetric retrieval better when you tell them whether a text is a search query or a stored document. Pass input_type (Singularity SDK / raw API):

grid.embed("nomic-embed-text-v1.5", "a king rules a kingdom", input_type="query")
grid.embed("nomic-embed-text-v1.5", "a queen rules a kingdom", input_type="document")

Billing

Embeddings are billed on input tokens only at the model’s per-token rate — there is no output/generation cost. Pay with credits (API key) or per-call with an x402 wallet, exactly like chat. A timed-out or failed request is never charged.

A minimal RAG loop

Index: embed your documents and store (id, vector, text) in a vector DB (pgvector, Pinecone, …).
Retrieve: embed the user’s question, find the nearest vectors (cosine).
Generate: pass the top matches as context to a chat model on the grid.

The embedding model is the search index step — not the conversation. See the API guide for the chat call.

Run an embedding model yourself and earn from other people’s search/RAG apps — see Provide Compute.

Overview

User Guide

Slayer Agent

Core flows

Staking Engine

Singularity Cloud Network

Contract Addresses

ERC8004

Skills

Singularity Payments

Agentic Fundraisers

Enterprise

Endpoint

Model catalog

Quickstart

Matryoshka dimensions

Query vs document

Billing

A minimal RAG loop

​Endpoint

​Model catalog

​Quickstart

​Matryoshka dimensions

​Query vs document

​Billing

​A minimal RAG loop

Endpoint

Model catalog

Quickstart

Matryoshka dimensions

Query vs document

Billing

A minimal RAG loop