Use tools from an MCP Server

[eduhza]

I'm trying to use Google_GenerativeAI.Microsoft to use an IChatClient with Gemini configured but i'm having issues with the tools.

My tools are defined in my MCP Server using ModelContextProtocol

[McpServerToolType]
public class FuncionarioTools(ILogger<FuncionarioTools> logger)
{
    [McpServerTool, Description("Adiciona um novo funcionário")]
    public async Task<string> AdicionarFuncionario([Description("Dados do funcionário a ser adicionado")] FuncionarioDto adicionarFuncionarioDto)
    {
        logger.LogInformation("Adicionando funcionário: {Nome}", adicionarFuncionarioDto.Nome);
        // do stuffs...
    }
}

But when I send the tools to GetResponseAsync I get errors for all DateOnly properties and parameters that has DataAnnotations such as [EmailAdress]:

GenerativeAI.Exceptions.ApiException
HResult=0x80131500
Message=INVALID_ARGUMENT (Code: 400): * GenerateContentRequest.tools[0].function_declarations[1].parameters.properties[dataReferencia].format: only 'enum' and 'date-time' are supported for STRING type

• GenerateContentRequest.tools[0].function_declarations[3].parameters.properties[adicionarExtraAssinaturaDto].properties[cobrarApartirDe].format: only 'enum' and 'date-time' are supported for STRING type
• GenerateContentRequest.tools[0].function_declarations[84].parameters.properties[adicionarFuncionarioDto].properties[emailPessoal].format: only 'enum' and 'date-time' are supported for STRING type
  ....
  Source=GenerativeAI
  StackTrace:
  at GenerativeAI.Core.ApiBase.d__13.MoveNext()
  at GenerativeAI.Core.ApiBase.d__12`2.MoveNext()
  at GenerativeAI.BaseModel.d__2.MoveNext()
  at GenerativeAI.GenerativeModel.d__36.MoveNext()
  at GenerativeAI.Microsoft.GenerativeAIChatClient.d__10.MoveNext()
  at Compartilhado.Aplicacao.Servicos.Compartilhado.AgenteIA.AgenteIAServico.d__4.MoveNext()

Is there a fix to this or another proper way to achieve the communication with gemini correctly exposing my MCP Tools?
Here is my code:

public sealed class AgenteIAServico(
    IChatClient chatClient,
    IClienteMcpServidor mcpClient,
    ILogger<AgenteIAServico> logger
) : IAgenteIAServico
{
    public async Task<RespostaPromptDto> ProcessarPromptAsync(string prompt, List<AIContent> historico)
    {
        try
        {
            await mcpClient.IniciarServidorMcp();
            prompt = "Buscar enderecos com id = 1";
            logger.LogInformation("Processando prompt: {Mensagem}", prompt);

            var mcpTools = await mcpClient.ListarToolsAsync();

            var chatHistory = new List<Microsoft.Extensions.AI.ChatMessage>
            {
                new(ChatRole.User, prompt)
            };

            var chatResponse = await chatClient.GetResponseAsync(chatHistory, new ChatOptions
            {
                Tools = [.. mcpTools]
            });
            return new RespostaPromptDto
            {
                Mensagem = chatResponse.Text,
                HistoricoConversa = [],
                Sucesso = true
            };
        }
        catch (Exception ex)
        {
            logger.LogError(ex, "Erro ao processar prompt");
            return new RespostaPromptDto
            {
                Mensagem = "Erro interno do servidor",
                HistoricoConversa = historico,
                Sucesso = false,
                ErroMensagem = ex.Message
            };
        }
}

Dependency injection:

        var geminiApiKey = configuration["GEMINI_API_KEY"];
        if (!string.IsNullOrWhiteSpace(geminiApiKey))
        {
            services.AddScoped<IChatClient>(provider => new GenerativeAIChatClient(geminiApiKey));
        }
        services.AddSingleton<IClienteMcpServidor, ClienteMcpServidor>();

[gunpal5]

GenerativeAI SDK already has a converter for DateOnly types. It seems like an issue with MEAI or ModelContextProtocol SDK's schema generation. Please create an issue in MEAI or ModelContextProtocol repository.

https://github.com/modelcontextprotocol/csharp-sdk
https://github.com/dotnet/extensions
@stephentoub

https://github.com/gunpal5/Google_GenerativeAI/blob/main/src/GenerativeAI/Types/Converters/TimeOnlyConverter.cs

Google_GenerativeAI/src/GenerativeAI/Constants/DefaultSerializerOptions.cs

Line 137 in 93f36a2

private static void AddConverters(JsonSerializerOptions options)

I will try to add a workaround for this in this SDK.

[stephentoub]

How is FuncionarioDto defined? Can you share a minimal repro? It may be that some of the schemas we're emitting aren't supported by gemini. If that's the case, we can address it with a schema transformer in the IChatClient impl. I can help with that once I know more about what is the actual cause.

[gunpal5]

@stephentoub Basically Gemini API doesn't support 'date' or 'time' format for 'type' = 'string', it only supports format as 'enum' and 'date-time'. Its a limitation from Google's APIs.

System.Text.Json generates schema for DateOnly ('type' = 'string' and 'format' = 'date') and for TimeOnly ('type'='string', format='time'). I added a converter to explicitly convert DateOnly and TimeOnly to ('type'='string' and format='date-time') in my SDK and convert back returned date-time value to DateOnly and TimeOnly.

I am not sure if it would be good idea to explicitly change the 'date' and 'time' format to 'date-time' for MEAI or MCP. I don't know how MEAI will deserialize full date and time values to DateOnly and TimeOnly.

[stephentoub]

Untested, but I think it'd look like this:
#63

[gunpal5]

@stephentoub Thank you very much, I will add some test cases and do proper testing with it.

[eduhza]

Thank you @gunpal5 and @stephentoub

It is a very simple Dto, it happens to every DateOnly such as a direct parameter or in a dto. Also for GUID and DataAnnotation EmailAdress. I managed to get it working creatings Dtos every tools using DateTime or string and then converting to the correct dto inside the tool.

[McpServerToolType]
public class FuncionarioTools(ILogger<FuncionarioTools> logger)
{
    [McpServerTool, Description("Adiciona um novo funcionário")]
    public async Task<string> AdicionarFuncionario([Description("Dados do funcionário a ser adicionado")] FuncionarioDto adicionarFuncionarioDto)
    {
        logger.LogInformation("Adicionando funcionário: {Nome}", adicionarFuncionarioDto.Nome);
        // do stuffs...
    }

    [McpServerTool, Description("Buscar funcionário por data de nacimento")]
    public async Task<string> BuscarFuncionarioPorDataDeNacimento([Description("Data de nacimento")] DateOnly dataNacimento)
    {
        logger.LogInformation("Buscando funcionário pora data de nacimento: {DataNacimento}", dataNacimento);
        // do stuffs...
    }
}

How is FuncionarioDto defined? Can you share a minimal repro? It may be that some of the schemas we're emitting aren't supported by gemini. If that's the case, we can address it with a schema transformer in the IChatClient impl. I can help with that once I know more about what is the actual cause.

FuncionarioDto:

using System.ComponentModel.DataAnnotations;
public class FuncionarioDto
{
    public required string Nome { get; set; }

    public string? Matricula { get; set; }

    public DateOnly? DataNascimento { get; set; }

    public required string DocumentoIdentificacaoSocial { get; set; }

    public required DateTime DataAdmissao { get; set; }

    public DateTime? DataDesligamento { get; set; }

    public string? RegistroGeral { get; set; }

    public required EstadoCivil EstadoCivil { get; set; }

    public required Escolaridade Escolaridade { get; set; }

    public int? EnderecoId { get; set; }

    public string? Telefone { get; set; }

    [EmailAddress]
    public string? EmailPessoal { get; set; }
    
    public Guid? KeyCloakId { get; set; }

    public bool Ativo { get; set; } = true;
}

[gunpal5]

@eduhza I am trying to get it fixed, can you create a sample project with your MCP tool for me to debug?

[gunpal5]

@eduhza another work around for now is to change your property types from DateOnly and TimeOnly to DateTime and parse it at your end.

[gunpal5]

Please check latest update: v3.1.1

[eduhza]

Thank you, everything is working now.