.Net: Bug: VertexAIGeminiChatCompletion 400 bad request when model makes multiple function calls in one message

[Klok-e]

Describe the bug
When the model makes multiple function calls in one message, the gemini API returns the following:

{
  "error" : {
    "code" : 400,
    "message" : "Please ensure that the number of function response parts is equal to the number of function call parts of the function call turn.",
    "status" : "INVALID_ARGUMENT"
  }
}

The request looks like this:

{
  "contents" : [ {
    "parts" : [ {
      "text" : "Using File Search find information on semantic kernel. Invoke the function exactly two times."
    } ],
    "role" : "user"
  }, {
    "parts" : [ {
      "functionCall" : {
        "name" : "FileSearch_GetTextSearchResults",
        "args" : {
          "query" : "semantic kernel"
        }
      }
    }, {
      "functionCall" : {
        "name" : "FileSearch_GetTextSearchResults",
        "args" : {
          "count" : 1,
          "query" : "semantic kernel"
        }
      }
    } ],
    "role" : "model"
  }, {
    "parts" : [ {
      "functionResponse" : {
        "name" : "FileSearch_GetTextSearchResults",
        "response" : {
          "content" : [ {
            "Name" : "0",
            "Link" : "file3.txt",
            "Value" : "Semantic kernel stuff 4"
          } ]
        }
      }
    } ],
    "role" : "function"
  }, {
    "parts" : [ {
      "functionResponse" : {
        "name" : "FileSearch_GetTextSearchResults",
        "response" : {
          "content" : [ {
            "Name" : "0",
            "Link" : "file3.txt",
            "Value" : "Semantic kernel stuff 4"
          } ]
        }
      }
    } ],
    "role" : "function"
  } ],
  "generationConfig" : { },
  "tools" : [ {
    "functionDeclarations" : [ {
      "name" : "FileSearch_GetTextSearchResults",
      "description" : "Perform a search for content related to the specified query. The search will return the name, value and link for the related content.",
      "parameters" : {
        "type" : "object",
        "required" : [ "query" ],
        "properties" : {
          "query" : {
            "description" : "What to search for",
            "type" : "string"
          },
          "count" : {
            "description" : "Number of results (default value: 2)",
            "type" : "integer"
          },
          "skip" : {
            "description" : "Number of results to skip (default value: 0)",
            "type" : "integer"
          }
        }
      }
    } ]
  } ]
}

To Reproduce

#pragma warning disable SKEXP0070
var builder = Kernel.CreateBuilder();

builder = builder
    .AddVertexAIGeminiChatCompletion(
        modelId: options.ModelName,
        bearerKey: accessToken,
        location: options.Location,
        projectId: options.ProjectId
    )
    .AddVertexAIEmbeddingGenerator(
        modelId: options.EmbeddingModel,
        bearerKey: accessToken,
        location: options.Location,
        projectId: options.ProjectId
    )
    .AddVectorStoreTextSearch<TextEmbeddingDataModel>();
#pragma warning restore SKEXP0070

var vectorStore = new InMemoryVectorStore();
        
builder.Services.AddScoped<IEmbeddingGenerator>(x => x.GetRequiredService<IEmbeddingGenerator<string, Embedding<float>>>());
builder.Services.AddSingleton(vectorStore);

builder.Services.AddSingleton<IVectorSearchable<TextEmbeddingDataModel>>(services =>
    vectorStore.GetCollection<string, TextEmbeddingDataModel>(
        "1", new VectorStoreCollectionDefinition { EmbeddingGenerator = services.GetRequiredService<IEmbeddingGenerator<string, Embedding<float>>>(), })
);

var kernel = builder.Build();

var collection = vectorStore.GetCollection<string, TextEmbeddingDataModel>("1", new VectorStoreCollectionDefinition()
{
    EmbeddingGenerator = kernel.GetRequiredService<IEmbeddingGenerator<string, Embedding<float>>>(),
});
await collection.EnsureCollectionExistsAsync();

List<(string, string)> documents =
[
    ("Semantic kernel stuff 1", "file1.txt"),
    ("Semantic kernel stuff 2", "file1.txt"),
    ("Semantic kernel stuff 3", "file2.txt"),
    ("Semantic kernel stuff 4", "file3.txt"),
];

await collection.UpsertAsync(documents.Select((textChunk, i) => new TextEmbeddingDataModel
{
    Key = i.ToString(),
    Text = (textChunk.Item1),
    FileName = textChunk.Item2,
}));

#pragma warning disable SKEXP0001
var textSearch = kernel.Services.GetRequiredService<VectorStoreTextSearch<TextEmbeddingDataModel>>();
#pragma warning restore SKEXP0001

kernel.Plugins.Add(KernelPluginFactory.CreateFromFunctions(
        "FileSearch",
        "Can search in documents",
        [textSearch.CreateGetTextSearchResults(searchOptions: new TextSearchOptions { Top = 3 }),]
    )
);

#pragma warning disable SKEXP0070

GeminiPromptExecutionSettings settings = new()
{
    ToolCallBehavior = GeminiToolCallBehavior.AutoInvokeKernelFunctions,
};
#pragma warning restore SKEXP0070

var chatCompletionService = kernel.GetRequiredService<IChatCompletionService>();

var history = new ChatHistory();

history.AddUserMessage("Using File Search find information on semantic kernel. Invoke the function exactly two times.");

var result = await chatCompletionService.GetChatMessageContentsAsync(history, settings, kernel);
Console.WriteLine(result);

Expected behavior
Gemini answers the question.

Platform

• Language: C#
• Source: Microsoft.SemanticKernel 1.57.0
• AI model: gemini-2.5-flash-preview-05-20
• OS: Windows

[JanainaLudwig]

This seems to be the same issue in GoogleAIGeminiChatCompletionService: #12528

[abbottdev]

Also +1 on this - and I think it's similar to the Parallel bug that's referenced on the linked Python issue - It seems that if I add > 1 allowed tool, this error is thrown, but with a single tool, the error does not appear.