Microsoft Foundry

Microsoft Foundry では、2 つの異なるコンテキスト パターンがサポートされています。 どちらも Foundry で管理されるリソースを使用しますが、エージェントにアタッチする方法が異なり、さまざまな問題が解決されます。

Pattern Agent Framework のメカニズム Behavior
ファイル検索 RAG プロバイダーホスト型ファイル検索ツール Foundry プロジェクトでアプリケーションが明示的にアップロードおよび管理するファイルとベクター ストアを検索します。
マネージド セマンティック メモリ FoundryMemoryProvider コンテキスト プロバイダー 会話からファクトと概要を抽出し、スコープ別に格納し、後で実行して関連するメモリを取得します。

モデル推論とサービスで管理される Foundry エージェントについては、Foundry モデル プロバイダーと Foundryエージェント サービスMicrosoft Microsoft参照してください。

ファイル検索 RAG を使用する

Foundry がキュレーションされたナレッジ ベースのドキュメント インジェストとベクター ストア ライフサイクルを所有する必要がある場合は、このパターンを使用します。 ファイル検索は、コンテキスト プロバイダーではなくホステッド ツールです。ツールの動作については、汎用 ファイル検索 ガイダンスを参照してください。 アプリケーション信頼のソースがAzure AI 検索インデックスである場合は、Azure AI 検索を使用します。

Foundry ベクター ストアとエージェントを作成する

ナレッジ ベース ファイルをアップロードし、ベクター ストアを作成し、 FileSearchToolをアタッチして、バージョン管理された FoundryAgentを作成します。

var endpoint = Environment.GetEnvironmentVariable("FOUNDRY_PROJECT_ENDPOINT") ?? throw new InvalidOperationException("FOUNDRY_PROJECT_ENDPOINT is not set.");
var deploymentName = Environment.GetEnvironmentVariable("FOUNDRY_MODEL") ?? "gpt-5.4-mini";

// Create an AI Project client and get an OpenAI client that works with the foundry service.
// WARNING: DefaultAzureCredential is convenient for development but requires careful consideration in production.
// In production, consider using a specific credential (e.g., ManagedIdentityCredential) to avoid
// latency issues, unintended credential probing, and potential security risks from fallback mechanisms.
AIProjectClient aiProjectClient = new(
    new Uri(endpoint),
    new DefaultAzureCredential());
OpenAIClient openAIClient = aiProjectClient.GetProjectOpenAIClient();

// Upload the file that contains the data to be used for RAG to the Foundry service.
OpenAIFileClient fileClient = openAIClient.GetOpenAIFileClient();
ClientResult<OpenAIFile> uploadResult = await fileClient.UploadFileAsync(
    filePath: "contoso-outdoors-knowledge-base.md",
    purpose: FileUploadPurpose.Assistants);

// Create a vector store in the Foundry service using the uploaded file.
VectorStoreClient vectorStoreClient = openAIClient.GetVectorStoreClient();
ClientResult<VectorStore> vectorStoreCreate = await vectorStoreClient.CreateVectorStoreAsync(options: new VectorStoreCreationOptions()
{
    Name = "contoso-outdoors-knowledge-base",
    FileIds = { uploadResult.Value.Id }
});

// Use the native OpenAI SDK FileSearchTool directly with the vector store ID.
#pragma warning disable OPENAI001
FileSearchTool fileSearchTool = new([vectorStoreCreate.Value.Id]);
#pragma warning restore OPENAI001

ProjectsAgentVersion agentVersion = await aiProjectClient.AgentAdministrationClient.CreateAgentVersionAsync(
    "AskContoso",
    new ProjectsAgentVersionCreationOptions(
        new DeclarativeAgentDefinition(model: deploymentName)
        {
            Instructions = "You are a helpful support specialist for Contoso Outdoors. Answer questions using the provided context and cite the source document when available.",
            Tools = { fileSearchTool }
        }));
FoundryAgent agent = aiProjectClient.AsAIAgent(agentVersion);

AgentSession session = await agent.CreateSessionAsync();

Console.WriteLine(">> Asking about returns\n");
Console.WriteLine(await agent.RunAsync("Hi! I need help understanding the return policy.", session));

Console.WriteLine("\n>> Asking about shipping\n");
Console.WriteLine(await agent.RunAsync("How long does standard shipping usually take?", session));

Console.WriteLine("\n>> Asking about product care\n");
Console.WriteLine(await agent.RunAsync("What is the best way to maintain the TrailRunner tent fabric?", session));

// Cleanup
await fileClient.DeleteFileAsync(uploadResult.Value.Id);
await vectorStoreClient.DeleteVectorStoreAsync(vectorStoreCreate.Value.Id);
await aiProjectClient.AgentAdministrationClient.DeleteAgentAsync(agent.Name);

運用ナレッジ ベースの永続的なベクター ストアを、すべてのプロセス実行に対して作成するのではなく再利用します。

パッケージをインストールする

pip install agent-framework-foundry --pre

Foundry プロジェクト OpenAI クライアントを使用してファイルとベクター ストアを作成し、結果のファイル検索ツールをエージェントに渡します。

async def create_vector_store(client: FoundryChatClient) -> tuple[str, str]:
    """Create a vector store with sample documents."""
    file = await client.client.files.create(
        file=("todays_weather.txt", b"The weather today is sunny with a high of 75F."), purpose="assistants"
    )
    vector_store = await client.client.vector_stores.create(
        name="knowledge_base",
        expires_after={"anchor": "last_active_at", "days": 1},
    )
    result = await client.client.vector_stores.files.create_and_poll(vector_store_id=vector_store.id, file_id=file.id)
    if result.last_error is not None:
        raise Exception(f"Vector store file processing failed with status: {result.last_error.message}")

    return file.id, vector_store.id


async def delete_vector_store(client: FoundryChatClient, file_id: str, vector_store_id: str) -> None:
    """Delete the vector store after using it."""
    with contextlib.suppress(Exception):
        await client.client.vector_stores.delete(vector_store_id=vector_store_id)
    with contextlib.suppress(Exception):
        await client.client.files.delete(file_id=file_id)


async def main() -> None:
    print("=== Foundry Chat Client with File Search Example ===\n")

    # Initialize the Foundry chat client
    # Make sure you're logged in via 'az login' before running this sample
    client = FoundryChatClient(credential=AzureCliCredential())

    file_id, vector_store_id = await create_vector_store(client)

    # Create file search tool using instance method
    file_search_tool = client.get_file_search_tool(vector_store_ids=[vector_store_id])

    agent = Agent(
        client=client,
        instructions="You are a helpful assistant that can search through files to find information.",
        tools=[file_search_tool],
    )

    query = "What is the weather today? Do a file search to find the answer."
    print(f"User: {query}")
    result = await agent.run(query)
    print(f"Agent: {result}\n")

    await delete_vector_store(client, file_id, vector_store_id)

Note

Foundry ファイル検索の統合は、現在、Agent Framework Go については文書化されていません。 最新の hosted-tool サポートについては、 Agent Framework Go リポジトリ を参照してください。

マネージド セマンティック メモリを追加する

エージェントがセッション間で永続的なユーザーまたはアプリケーション コンテキストを呼び戻す必要がある場合は、 FoundryMemoryProvider を使用します。 Foundry メモリは、抽出されたファクトと概要を完全な会話トランスクリプトとは別に格納します。

パッケージをインストールする

dotnet add package Microsoft.Agents.AI.Foundry --prerelease

安定したスコープを持つ FoundryMemoryProvider を作成し、メモリ ストアが存在することを確認し、非同期更新を待ってから、新しく抽出されたメモリに依存します。

// Create an AIProjectClient for Foundry with Azure Identity authentication.
// WARNING: DefaultAzureCredential is convenient for development but requires careful consideration in production.
// In production, consider using a specific credential (e.g., ManagedIdentityCredential) to avoid
// latency issues, unintended credential probing, and potential security risks from fallback mechanisms.
DefaultAzureCredential credential = new();
AIProjectClient projectClient = new(new Uri(foundryEndpoint), credential);

// Get the ChatClient from the AIProjectClient's OpenAI property using the deployment name.
// The stateInitializer can be used to customize the Foundry Memory scope per session and it will be called each time a session
// is encountered by the FoundryMemoryProvider that does not already have state stored on the session.
// If each session should have its own scope, you can create a new id per session via the stateInitializer, e.g.:
// new FoundryMemoryProvider(projectClient, memoryStoreName, stateInitializer: _ => new(new FoundryMemoryProviderScope(Guid.NewGuid().ToString())), ...)
// In our case we are storing memories scoped by user so that memories are retained across sessions.
FoundryMemoryProvider memoryProvider = new(
    projectClient,
    memoryStoreName,
    stateInitializer: _ => new(new FoundryMemoryProviderScope("sample-user-123")));

ChatClientAgent agent = projectClient.AsAIAgent(
    new ChatClientAgentOptions()
    {
        Name = "TravelAssistantWithFoundryMemory",
        ChatOptions = new()
        {
            ModelId = deploymentName,
            Instructions = "You are a friendly travel assistant. Use known memories about the user when responding, and do not invent details."
        },
        AIContextProviders = [memoryProvider]
    });

AgentSession session = await agent.CreateSessionAsync();

Console.WriteLine("\n>> Setting up Foundry Memory Store\n");

// Ensure the memory store exists (creates it with the specified models if needed).
await memoryProvider.EnsureMemoryStoreCreatedAsync(deploymentName, embeddingModelName, "Sample memory store for travel assistant");

// Clear any existing memories for this scope to demonstrate fresh behavior.
await memoryProvider.EnsureStoredMemoriesDeletedAsync(session);

Console.WriteLine(await agent.RunAsync("Hi there! My name is Taylor and I'm planning a hiking trip to Patagonia in November.", session));
Console.WriteLine(await agent.RunAsync("I'm travelling with my sister and we love finding scenic viewpoints.", session));

// Memory extraction in Microsoft Foundry is asynchronous and takes time to process.
// WhenUpdatesCompletedAsync polls all pending updates and waits for them to complete.
Console.WriteLine("\nWaiting for Foundry Memory to process updates...");
await memoryProvider.WhenUpdatesCompletedAsync();

Console.WriteLine("Updates completed.\n");

Console.WriteLine(await agent.RunAsync("What do you already know about my upcoming trip?", session));

Console.WriteLine("\n>> Serialize and deserialize the session to demonstrate persisted state\n");
JsonElement serializedSession = await agent.SerializeSessionAsync(session);
AgentSession restoredSession = await agent.DeserializeSessionAsync(serializedSession);
Console.WriteLine(await agent.RunAsync("Can you recap the personal details you remember?", restoredSession));

Console.WriteLine("\n>> Start a new session that shares the same Foundry Memory scope\n");

Console.WriteLine("\nWaiting for Foundry Memory to process updates...");
await memoryProvider.WhenUpdatesCompletedAsync();

AgentSession newSession = await agent.CreateSessionAsync();
Console.WriteLine(await agent.RunAsync("Summarize what you already know about me.", newSession));

パッケージをインストールする

pip install agent-framework-foundry --pre

AIProjectClientを使用してメモリ ストアを作成し、FoundryMemoryProviderをエージェントにアタッチします。

async def main() -> None:
    endpoint = os.environ["FOUNDRY_PROJECT_ENDPOINT"]
    async with (
        AzureCliCredential() as credential,
        AIProjectClient(endpoint=endpoint, credential=credential) as project_client,
    ):
        # Generate a unique memory store name to avoid conflicts
        memory_store_name = f"agent_framework_memory_{datetime.now(timezone.utc).strftime('%Y%m%d')}"
        # Specify memory store options
        options = MemoryStoreDefaultOptions(
            chat_summary_enabled=False,
            user_profile_enabled=True,
            user_profile_details="Avoid irrelevant or sensitive data, such as age, financials, precise location, and credentials",
        )
        memory_store_definition = MemoryStoreDefaultDefinition(
            chat_model=os.environ["FOUNDRY_MODEL"],
            embedding_model=os.environ["AZURE_OPENAI_EMBEDDING_MODEL"],
            options=options,
        )
        print(f"Creating memory store '{memory_store_name}'...")
        try:
            # Create a memory store
            memory_store = await project_client.beta.memory_stores.create(
                name=memory_store_name,
                description="Memory store for Agent Framework with FoundryMemoryProvider",
                definition=memory_store_definition,
            )
        except Exception as e:
            print(f"Failed to create memory store: {e}")
            return

        print(f"Created memory store: {memory_store.name} ({memory_store.id})")
        print(f"Description: {memory_store.description}\n")
        print("==========================================")

        # Create the chat client
        client = FoundryChatClient(project_client=project_client)
        # Create the Foundry Memory context provider
        memory_provider = FoundryMemoryProvider(
            project_client=project_client,
            memory_store_name=memory_store.name,
            scope="user_123",  # Scope memories to a specific user, if not set, the session_id
            # will be used as scope, which means memories are only shared within the same session
            update_delay=0,  # Do not wait to update memories after each interaction (for demo purposes)
            # In production, consider setting a delay to batch updates and reduce costs
        )

        # Create an agent with the memory context provider
        async with Agent(
            name="MemoryAgent",
            client=client,
            instructions="""You are a helpful assistant that remembers past conversations.
                The memories from previous interactions are automatically provided to you.""",
            context_providers=[memory_provider, InMemoryHistoryProvider(load_messages=False)],
            default_options={"store": False},
        ) as agent:
            try:
                # note that we will use the service side storage, nor load messsages from the history provider,
                # but we include it to demonstrate that it can be used alongside the Foundry provider for other use cases.
                session = agent.create_session()

                # First interaction - establish some preferences
                print("=== First conversation ===")
                query1 = "I prefer dark roast coffee and I'm allergic to nuts"
                print(f"User: {query1}")
                result1 = await agent.run(query1, session=session)
                print(f"Agent: {result1}\n")

                # Wait for memories to be processed
                print("Waiting for memories to be stored...")
                await asyncio.sleep(8)

                # Second interaction - test memory recall
                print("=== Second conversation ===")
                query2 = "Can you recommend a coffee and snack for me?"
                print(f"User: {query2}")
                result2 = await agent.run(query2, session=session)
                print(f"Agent: {result2}\n")

                # Third interaction - continue the conversation
                print("=== Third conversation ===")
                query3 = "What do you remember about my preferences?"
                print(f"User: {query3}")
                result3 = await agent.run(query3, session=session)
                print(f"Agent: {result3}\n")

                print(f"Stored memories from: {memory_store.name} ({memory_store.id})")
                res = await project_client.beta.memory_stores.search_memories(name=memory_store.name, scope="user_123")
                for memory in res.memories:
                    print(f"Memory: {memory.memory_item.content}")

            except Exception as e:
                print(f"An error occurred: {e}")

            finally:
                await project_client.beta.memory_stores.delete(memory_store_name)

このサンプルでは、サービス側とローカルのトランスクリプトの読み込みが無効になるため、後の応答ではチャット履歴再生ではなくセマンティック メモリが示されます。

Note

Microsoft Foundry メモリ統合は現在、Agent Framework Go では使用できません。 最新の状態については、 Agent Framework Go リポジトリ を参照してください。

実稼働に関する考慮事項

  • 運用ナレッジ ベースの永続的ベクター ストアを再利用します。
  • アプリケーション所有の安定したメモリ スコープ識別子を使用し、スコープを選択する前にアクセスを承認します。
  • 後続の操作が新しく書き込まれたメモリに依存している場合は、非同期抽出を待機します。
  • 完全な会話レコードが必要な場合は、履歴プロバイダーに正確なトランスクリプトを保持します。
  • コンプライアンス要件に合わせてリテンション期間、リージョン、モデルのデプロイを構成します。

次のステップ

より深く進む: