---
title: "Embeddings with OpenAI SDKs"
description: "Create text embeddings with the stock OpenAI SDKs, from Python, TypeScript or curl."
---

> Documentation Index
> Fetch the complete documentation index at: https://docs.ezghcloud.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Embeddings with OpenAI SDKs

EZGH Inference's embeddings endpoint is OpenAI's. Point an OpenAI SDK at your project's base URL
and give it an EZGH API key; nothing else changes.

## Before you start

- **Base URL:** your project's, with its organization and project IDs (in the console, the
  Quickstart on a model's page shows it filled in):

  ```text
  https://inference-api.ezghcloud.com/v1/organizations/<org-id>/projects/<project-id>/openai/v1
  ```

- **API key:** an [API key](/iam/api-keys) whose owner has `inference.models.invoke` on the
  project (`InferenceFullAccess`), in `EZGH_API_KEY`. OAuth access tokens and OIDC API keys work
  too. The API takes bearer credentials only, never cookies.

## Create embeddings

### Python

```sh
pip install openai
```

```python
import os
from openai import OpenAI

client = OpenAI(
    base_url="https://inference-api.ezghcloud.com/v1/organizations/<org-id>/projects/<project-id>/openai/v1",
    api_key=os.environ["EZGH_API_KEY"],
)
r = client.embeddings.create(model="qwen3-embedding-0.6b", input=["Instruct: Given a web search query, retrieve relevant passages\nQuery: what is a vector database"])
print(len(r.data[0].embedding))
```
### TypeScript

```sh
npm install openai
```

```ts
import OpenAI from "openai"

const client = new OpenAI({
  baseURL: "https://inference-api.ezghcloud.com/v1/organizations/<org-id>/projects/<project-id>/openai/v1",
  apiKey: process.env.EZGH_API_KEY,
})
const r = await client.embeddings.create({
  model: "qwen3-embedding-0.6b",
  input: ["Instruct: Given a web search query, retrieve relevant passages\nQuery: what is a vector database"],
})
console.log(r.data[0].embedding.length)
```
### curl

```sh
curl https://inference-api.ezghcloud.com/v1/organizations/<org-id>/projects/<project-id>/openai/v1/embeddings \
  -H "Authorization: Bearer $EZGH_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "qwen3-embedding-0.6b", "input": ["Instruct: Given a web search query, retrieve relevant passages\nQuery: what is a vector database"]}'
```

The response is OpenAI's, with the model version that ran:

```json
{
  "object": "list",
  "data": [{ "object": "embedding", "index": 0, "embedding": [0.0123, -0.0456] }],
  "model": "qwen3-embedding-0.6b:1.0.0",
  "usage": { "prompt_tokens": 7, "total_tokens": 7 }
}
```

See [create embeddings](/inference-api/CreateEmbeddings/) in the API reference.

## Queries and documents

`qwen3-embedding-0.6b` is instruction-aware. For retrieval, write each **query** with a one-line
task description, and send **documents** as they are:

```text
Instruct: Given a web search query, retrieve relevant passages
Query: what is a vector database
```

The instruction is part of your input; there's no separate parameter. Use the same instruction
for every query in an index, in English even when the text isn't.

## Request fields

| Field | Values |
| --- | --- |
| `model` | `qwen3-embedding-0.6b`, or a pinned version, `qwen3-embedding-0.6b:1.0.0`. See [Models](/inference/models). |
| `input` | A string, or an array of 1 to 128 strings. Empty strings are refused. Each input is at most 8,192 tokens. Arrays of token IDs are refused with `token_input_unsupported`. |
| `dimensions` | Optional: 32 to 1,024 (the default). Shorter vectors are truncated and normalized (Matryoshka). |
| `encoding_format` | `float` (the default) or `base64`: little-endian 32-bit floats, base64-encoded. |
| `user` | Accepted and ignored. |

Other fields are ignored. A request body is at most 1 MiB.

The responses carry `x-ratelimit-*` headers; see [Limits and errors](/inference/limits), with
every error code. The OpenAI SDKs retry `429` and `5xx` answers, honoring `Retry-After`, except
the monthly token limit (`insufficient_quota`).

## LangChain

LangChain's `OpenAIEmbeddings` needs two settings:

- `check_embedding_ctx_length=False`: by default it splits text with OpenAI's tokenizer and sends
  token IDs, which the API refuses (`token_input_unsupported`).
- `chunk_size=128`: it sends up to 1,000 inputs per request by default, and a request takes at
  most 128, so embedding more than 128 documents at once would fail with `invalid_request`.

```python
import os
from langchain_openai import OpenAIEmbeddings

embeddings = OpenAIEmbeddings(
    model="qwen3-embedding-0.6b",
    base_url="https://inference-api.ezghcloud.com/v1/organizations/<org-id>/projects/<project-id>/openai/v1",
    api_key=os.environ["EZGH_API_KEY"],
    check_embedding_ctx_length=False,
    chunk_size=128,
)
```

## Data

Inputs and embeddings are never stored, logged or recorded. A refused request appears in
[Trails](/inference#trails), with its error code; a successful one doesn't.

Source: https://docs.ezghcloud.com/inference/embeddings/index.mdx