共用方式為


教學:第三部分 - 使用 Microsoft Foundry SDK 評估自訂聊天應用程式

在這個教學中,你會評估你在 教學系列第二部分中打造的聊天應用程式。 你在多個指標上評估應用程式的品質,然後不斷改進。 在此部分中,您會:

  • 建立評估數據集
  • 使用 Azure AI 評估工具評估聊天應用程式
  • 反覆運算並改善您的應用程式

本教學建立在 第二部分:使用 Microsoft Foundry SDK 建立自訂聊天應用程式的基礎上。

先決條件

備註

本教學課程使用 中樞型專案。 這裡展示的步驟和程式碼不適用於 Foundry 專案。 如需詳細資訊,請參閱 項目類型

  • 完成 教學系列的第二部分 ,建立聊天應用程式。
  • 使用你在第一部分創建的同一個 Microsoft Foundry 專案。
  • Azure AI 權限:擁有者或貢獻者角色,可修改模型端點速率限制並執行評估工作。
  • 務必完成 第二部分新增遙測記錄 的步驟。

建立評估資料集

請使用以下評估資料集,其中包含範例問題與預期答案。 利用此資料集搭配評估器及 get_chat_response() 目標函數,評估聊天應用程式在相關性、貼地性及連貫性指標上的表現。

  1. 你的 assets 資料夾裡建立一個名為 chat_eval_data.jsonl 的檔案。

  2. 將此資料集貼到檔案中:

    {"query": "Which tent is the most waterproof?", "truth": "The Alpine Explorer Tent has the highest rainfly waterproof rating at 3000m"}
    {"query": "Which camping table holds the most weight?", "truth": "The Adventure Dining Table has a higher weight capacity than all of the other camping tables mentioned"}
    {"query": "How much do the TrailWalker Hiking Shoes cost? ", "truth": "The Trailewalker Hiking Shoes are priced at $110"}
    {"query": "What is the proper care for trailwalker hiking shoes? ", "truth": "After each use, remove any dirt or debris by brushing or wiping the shoes with a damp cloth."}
    {"query": "What brand is TrailMaster tent? ", "truth": "OutdoorLiving"}
    {"query": "How do I carry the TrailMaster tent around? ", "truth": " Carry bag included for convenient storage and transportation"}
    {"query": "What is the floor area for Floor Area? ", "truth": "80 square feet"}
    {"query": "What is the material for TrailBlaze Hiking Pants?", "truth": "Made of high-quality nylon fabric"}
    {"query": "What color does TrailBlaze Hiking Pants come in?", "truth": "Khaki"}
    {"query": "Can the warrenty for TrailBlaze pants be transfered? ", "truth": "The warranty is non-transferable and applies only to the original purchaser of the TrailBlaze Hiking Pants. It is valid only when the product is purchased from an authorized retailer."}
    {"query": "How long are the TrailBlaze pants under warranty for? ", "truth": " The TrailBlaze Hiking Pants are backed by a 1-year limited warranty from the date of purchase."}
    {"query": "What is the material for PowerBurner Camping Stove? ", "truth": "Stainless Steel"}
    {"query": "Is France in Europe?", "truth": "Sorry, I can only queries related to outdoor/camping gear and equipment"}
    

    參考資料: 評估資料集的 JSONL 格式

使用 Azure AI 評估工具進行評估

建立一個評估腳本,產生目標函式包裝器、載入資料集、執行評估,並將結果記錄到你的 Foundry 專案。

  1. 在主資料夾裡建立一個名為 evaluate.py 的檔案。

  2. 新增下列程式代碼以匯入必要的連結庫、建立專案用戶端,以及設定一些設定:

    import os
    import pandas as pd
    from azure.ai.projects import AIProjectClient
    from azure.ai.projects.models import ConnectionType
    from azure.ai.evaluation import evaluate, GroundednessEvaluator
    from azure.identity import DefaultAzureCredential
    
    from chat_with_products import chat_with_products
    
    # load environment variables from the .env file at the root of this repo
    from dotenv import load_dotenv
    
    load_dotenv()
    
    # create a project client using environment variables loaded from the .env file
    project = AIProjectClient.from_connection_string(
        conn_str=os.environ["AIPROJECT_CONNECTION_STRING"], credential=DefaultAzureCredential()
    )
    
    connection = project.connections.get_default(connection_type=ConnectionType.AZURE_OPEN_AI, include_credentials=True)
    
    evaluator_model = {
        "azure_endpoint": connection.endpoint_url,
        "azure_deployment": os.environ["EVALUATION_MODEL"],
        "api_version": "2024-06-01",
        "api_key": connection.key,
    }
    
    groundedness = GroundednessEvaluator(evaluator_model)
    

    References: AIProjectClientDefaultAzureCredentialazure-ai-evaluation.

  3. 新增程式代碼以建立包裝函式,以實作查詢和回應評估的評估介面:

    def evaluate_chat_with_products(query):
        response = chat_with_products(messages=[{"role": "user", "content": query}])
        return {"response": response["message"].content, "context": response["context"]["grounding_data"]}
    

    參考資料: azure-ai-evaluation評估目標功能

  4. 最後,新增程式碼來執行評估、在本地查看結果,並在 Foundry 入口網站取得評估結果的連結:

    # Evaluate must be called inside of __main__, not on import
    if __name__ == "__main__":
        from config import ASSET_PATH
    
        # workaround for multiprocessing issue on linux
        from pprint import pprint
        from pathlib import Path
        import multiprocessing
        import contextlib
    
        with contextlib.suppress(RuntimeError):
            multiprocessing.set_start_method("spawn", force=True)
    
        # run evaluation with a dataset and target function, log to the project
        result = evaluate(
            data=Path(ASSET_PATH) / "chat_eval_data.jsonl",
            target=evaluate_chat_with_products,
            evaluation_name="evaluate_chat_with_products",
            evaluators={
                "groundedness": groundedness,
            },
            evaluator_config={
                "default": {
                    "query": {"${data.query}"},
                    "response": {"${target.response}"},
                    "context": {"${target.context}"},
                }
            },
            azure_ai_project=project.scope,
            output_path="./myevalresults.json",
        )
    
        tabular_result = pd.DataFrame(result.get("rows"))
    
        pprint("-----Summarized Metrics-----")
        pprint(result["metrics"])
        pprint("-----Tabular Result-----")
        pprint(tabular_result)
        pprint(f"View evaluation results in AI Studio: {result['studio_url']}")
    

    References: azure-ai-evaluationAIProjectClient.

設定評估模型

評估腳本會多次呼叫模型。 考慮增加評估模型每分鐘的代幣數量。

在本教學課程系列的第 1 部分中,您已建立一個 .env 檔案,指定評估模型 gpt-4o-mini的名稱。 如果您有可用的配額,請嘗試增加此模型的每分鐘令牌限制。 如果您沒有足夠的配額可增加此值,請不要擔心。 腳本的設計目的是要處理限制錯誤。

  1. 在 Foundry 入口網站的專案中,選擇Models + endpoints
  2. 選取 gpt-4o-mini
  3. 請選取 ,再編輯
  4. 如果你有配額, 請將每分鐘代幣數上限 提高到 30 個或以上。
  5. 選擇 儲存並關閉

執行評估指令碼

  1. 從你的控制台,使用 Azure CLI 登入你的 Azure 帳號:

    az login
    
  2. 安裝必要的套件:

    pip install azure-ai-evaluation[remote]
    

    參考資料:azure-ai-evaluation SDKEvaluation SDK 文件。

確認你的評估設定

在進行完整評估(需時 5–10 分鐘)前,先透過以下快速測試確認 SDK 與專案連線是否正常運作:

from azure.ai.projects import AIProjectClient
from azure.identity import DefaultAzureCredential

# Verify connection to project
client = AIProjectClient.from_config(credential=DefaultAzureCredential())
print("Evaluation SDK is ready! You can now run evaluate.py")

如果你看到 "Evaluation SDK is ready!",你的設定就完成了,可以繼續進行。

參考資料: AIProjectClientDefaultAzureCredential.

開始評估

  • 執行評估腳本:

    python evaluate.py
    

評估過程需時5至10分鐘。 你可能會看到逾時警告和速率限制錯誤。 腳本會自動處理這些錯誤並繼續處理。

解譯評估輸出

在主控台輸出中,你會看到每個問題的答案,接著是一張總結指標的表格,顯示相關性、紮實性和連貫性分數。 GPT輔助指標的分數範圍從0(最差)到4(最佳)不等。 尋找低接地性分數來辨識參考文獻支持不足的回答,低相關性分數則用來辨識離題回答。

你可能會看到很多 WARNING:opentelemetry.attributes: 訊息和逾時錯誤。 您可以放心地忽略這些訊息。 它們不會影響評估結果。 評估腳本設計用來處理速率限制錯誤並持續處理。

評估結果輸出還包含一個連結,可在 Foundry 入口網站查看詳細結果,您可以並排比較評估流程並追蹤隨時間的改進。

====================================================
'-----Summarized Metrics-----'
{'groundedness.gpt_groundedness': 1.6666666666666667,
 'groundedness.groundedness': 1.6666666666666667}
'-----Tabular Result-----'
                                     outputs.response  ... line_number
0   Could you specify which tent you are referring...  ...           0
1   Could you please specify which camping table y...  ...           1
2   Sorry, I only can answer queries related to ou...  ...           2
3   Could you please clarify which aspects of care...  ...           3
4   Sorry, I only can answer queries related to ou...  ...           4
5   The TrailMaster X4 Tent comes with an included...  ...           5
6                                            (Failed)  ...           6
7   The TrailBlaze Hiking Pants are crafted from h...  ...           7
8   Sorry, I only can answer queries related to ou...  ...           8
9   Sorry, I only can answer queries related to ou...  ...           9
10  Sorry, I only can answer queries related to ou...  ...          10
11  The PowerBurner Camping Stove is designed with...  ...          11
12  Sorry, I only can answer queries related to ou...  ...          12

[13 rows x 8 columns]
('View evaluation results in Foundry portal: '
 'https://xxxxxxxxxxxxxxxxxxxxxxx')

反覆運算和改善

評估結果顯示,回應往往並未充分基於參考文件。 為了提升基礎性,請修改 assets/grounded_chat.prompty 檔案中的系統提示,鼓勵模型更有效地使用參考文件。

目前提示(有問題):

If the question is not related to outdoor/camping gear and clothing, just say 'Sorry, I only can answer queries related to outdoor/camping gear and clothing. So, how can I help?'
If the question is related to outdoor/camping gear and clothing but vague, ask clarifying questions.

改進提示

If the question is related to outdoor/camping gear and clothing, answer based on the reference documents provided.
If you cannot find information in the reference documents, say: 'I don't have information about that specific topic. Let me help with related products or try a different question.'
For vague questions, ask clarifying questions to better assist.

更新提示後:

  1. 儲存檔案。

  2. 再執行一次評估腳本:

    python evaluate.py
    
  3. 將新的評估結果與上一次的測試做比較。 你應該能看到可靠性分數的提升。

嘗試其他調整,例如:

  • 將系統提示改為專注於準確性而非完整性
  • 使用不同模型測試(例如,gpt-4-turbo 如果有的話)
  • 調整上下文檢索以返回更相關的文件

每次迭代都能幫助你了解哪些變更能改善特定指標。

清理資源

為避免不必要的 Azure 成本,如果不再需要,請刪除你在教學中建立的資源。 若要管理資源,您可以使用 Azure 入口網站