Compartilhar via

Usar o Azure OpenAI no Fabric com a API REST (versão prévia)

Importante

Esse recurso está em versão prévia.

Este documento mostra exemplos de como usar o Azure OpenAI no Fabric com a API REST.

Inicialização

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)
}

Bate-papo

GPT-4o e GPT-4o-mini são modelos de linguagem otimizados para interfaces de conversação.

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)

Saída

==========================================================================================
| 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.
==========================================================================================

Incorporações

Uma inserção é um formato especial de representação de dados que os algoritmos e modelos de machine learning podem utilizar facilmente. Ela contém o significado semântico repleto de informações de um texto, representado por um vetor de números de ponto flutuante. A distância entre duas inserções no espaço vetorial está relacionada à semelhança semântica entre duas entradas originais. Por exemplo, se dois textos forem semelhantes, suas representações de vetor também deverão ser semelhantes.

Para acessar os pontos de extremidade de inserções do Azure OpenAI no Fabric, você pode enviar uma solicitação de API usando o seguinte formato:

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

deployment_name pode ser text-embedding-ada-002.

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)

Saída:

==========================================================================================
| 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, ... ]
==========================================================================================

Conteúdo relacionado