How to use fine-tuned models on Vertex AI?

[nguyenlamlll]

Hi everyone,
I have fine-tuned models on Vertex AI, exposed via Endpoints already.

Essentially, if I use powershell, it would look like

$API_ENDPOINT = "us-central1-aiplatform.googleapis.com"
$MODEL_ID = "projects/${Project_Id}/locations/us-central1/endpoints/${Endpoint_Id}"
$GENERATE_CONTENT_API = "generateContent"

Invoke-RestMethod -Uri "https://${API_ENDPOINT}/v1/${MODEL_ID}:${GENERATE_CONTENT_API}" -Method POST -Headers $headers -Body $body

Attached is how I implement in C# semantic kernel so far. (But it isn't working).

using Microsoft.SemanticKernel;
using Google.Apis.Auth.OAuth2;
using Microsoft.SemanticKernel.ChatCompletion;

var builder = Kernel.CreateBuilder();

var modelId = "{endpointId}";
var credential = GoogleCredential.GetApplicationDefault()
    .CreateScoped("https://www.googleapis.com/auth/cloud-platform");

var url = "https://us-central1-aiplatform.googleapis.com/v1/";
var idToken = credential
    .GetOidcTokenAsync(OidcTokenOptions.FromTargetAudience(url))
    .Result;
var accessToken = idToken.GetAccessTokenAsync().Result;

builder.AddVertexAIGeminiChatCompletion(
    modelId,
    () =>
    {
        var credential = GoogleCredential.GetApplicationDefault()
            .CreateScoped("https://www.googleapis.com/auth/cloud-platform");
        var token = credential
            .UnderlyingCredential
            .GetAccessTokenForRequestAsync().Result;
        return new ValueTask<string>(token);
    },
    "us-central1",
    "projectId");

var kernel = builder.Build();

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

var chatHistory = new ChatHistory();
chatHistory.AddUserMessage("What is captial of Italy?");

try
{
    var result = await chatCompletionService.GetChatMessageContentAsync(
        chatHistory,
        null,
        kernel);
    Console.WriteLine("Response:");
    Console.WriteLine(result.Content);
}
catch (Exception ex)
{
    Console.WriteLine($"Error: {ex.Message}");
}

Console.ReadLine();

It throws an error like this

{
"error": {
"code": 404,
"message": "Publisher Model projects/{projectId}/locations/us-central1/publishers/google/models/{endpointId} not found.",
"status": "NOT_FOUND"
}
}

Looking at the endpoint (publishers/google/models/...), it seems we can only interact with base models like gemini-2.5-flash, gemini-2.5-pro, etc.?

Meaning, if I change my modelId to this

var modelId = "gemini-2.5-flash";

Then the code snippet works.

Is there currently a way to use fine-tuned models, not the base ones?

Thanks!