다음을 통해 공유


REST API와 함께 Fabric에서 Azure OpenAI 사용(미리 보기)

중요합니다

이 기능은 미리 보기로 제공됩니다.

이 문서에서는 REST API를 사용하여 패브릭에서 Azure OpenAI를 사용하는 방법의 예를 보여 줍니다.

초기화

from synapse.ml.mlflow import get_mlflow_env_config
from trident_token_library_wrapper import PyTridentTokenLibrary

mlflow_env_configs = get_mlflow_env_config()
mwc_token = PyTridentTokenLibrary.get_mwc_token(mlflow_env_configs.workspace_id, mlflow_env_configs.artifact_id, 2)

auth_headers = {
    "Authorization" : "MwcToken {}".format(mwc_token)
}

채팅

GPT-4o 및 GPT-4o-mini는 대화형 인터페이스에 최적화된 언어 모델입니다.

import requests

def print_chat_result(messages, response_code, response):
    print("==========================================================================================")
    print("| OpenAI Input    |")
    for msg in messages:
        if msg["role"] == "system":
            print("[System] ", msg["content"])
        elif msg["role"] == "user":
            print("Q: ", msg["content"])
        else:
            print("A: ", msg["content"])
    print("------------------------------------------------------------------------------------------")
    print("| Response Status |", response_code)
    print("------------------------------------------------------------------------------------------")
    print("| OpenAI Output   |")
    if response.status_code == 200:
        print(response.json()["choices"][0]["message"]["content"])
    else:
        print(response.content)
    print("==========================================================================================")


deployment_name = "gpt-4o" # deployment_id could be one of {gpt-4o or gpt-4o-mini}
openai_url = mlflow_env_configs.workload_endpoint + f"cognitive/openai/openai/deployments/{deployment_name}/chat/completions?api-version=2025-04-01-preview"
payload = {
    "messages": [
        {"role": "system", "content": "You are an AI assistant that helps people find information."},
        {"role": "user", "content": "Does Azure OpenAI support customer managed keys?"}
    ]
}

response = requests.post(openai_url, headers=auth_headers, json=payload)
print_chat_result(payload["messages"], response.status_code, response)

출력

==========================================================================================
| OpenAI Input    |
[System]  You are an AI assistant that helps people find information.
Q:  Does Azure OpenAI support customer managed keys?
------------------------------------------------------------------------------------------
| Response Status | 200
------------------------------------------------------------------------------------------
| OpenAI Output   |
As of my last training cut-off in October 2023, Azure OpenAI Service did not natively support customer-managed keys (CMK) for encryption of data at rest. Data within Azure OpenAI is typically encrypted using Microsoft-managed keys. 

However, you should verify this information on the official Azure documentation or by contacting Microsoft support, as cloud service features and capabilities are frequently updated.
==========================================================================================

포함(Embeddings)

포함은 기계 학습 모델 및 알고리즘에서 쉽게 활용할 수 있는 특수한 형식의 데이터 표현입니다. 부동 소수점 숫자의 벡터로 표현되는 텍스트의 정보가 풍부한 의미 체계 의미를 포함합니다. 벡터 공간에서 두 포함 간의 거리는 두 원본 입력 간의 의미론적 유사성과 관련이 있습니다. 예를 들어 두 텍스트가 비슷한 경우 벡터 표현도 유사해야 합니다.

Fabric에서 Azure OpenAI 포함 엔드포인트에 액세스하려면 다음 형식을 사용하여 API 요청을 보낼 수 있습니다.

POST <url_prefix>/openai/deployments/<deployment_name>/embeddings?api-version=2024-02-01

deployment_nametext-embedding-ada-002일 수 있습니다.

import requests

def print_embedding_result(prompts, response_code, response):
    print("==========================================================================================")
    print("| OpenAI Input    |\n\t" + "\n\t".join(prompts))
    print("------------------------------------------------------------------------------------------")
    print("| Response Status |", response_code)
    print("------------------------------------------------------------------------------------------")
    print("| OpenAI Output   |")
    if response_code == 200:
        for data in response.json()['data']:
            print("\t[" + ", ".join([f"{n:.8f}" for n in data["embedding"][:10]]) + ", ... ]")
    else:
        print(response.content)
    print("==========================================================================================")

deployment_name = "text-embedding-ada-002"
openai_url = mlflow_env_configs.workload_endpoint + f"cognitive/openai/openai/deployments/{deployment_name}/embeddings?api-version=2025-04-01-preview"
payload = {
    "input": [
        "empty prompt, need to fill in the content before the request",
        "Once upon a time"
    ]
}

response = requests.post(openai_url, headers=auth_headers, json=payload)
print_embedding_result(payload["input"], response.status_code, response)

출력:

==========================================================================================
| OpenAI Input    |
	empty prompt, need to fill in the content before the request
	Once upon a time
------------------------------------------------------------------------------------------
| Response Status | 200
------------------------------------------------------------------------------------------
| OpenAI Output   |
	[-0.00258819, -0.00449802, -0.01700411, 0.00405622, -0.03064079, 0.01899395, -0.01295485, -0.01426286, -0.03512142, -0.01831212, ... ]
	[0.02129045, -0.02013996, -0.00462094, -0.01146069, -0.01123944, 0.00199124, 0.00228992, -0.01370478, 0.00855917, -0.01470356, ... ]
==========================================================================================