Added the Model Comparison assistant for blind comparisons of two models (#1009)

Co-authored-by: Thorsten Sommer <SommerEngineering@users.noreply.github.com>
This commit is contained in:
Dominic NeuburgandThorsten Sommer authored and GitHub committed 2026-11-14 14:11:34 +01:00
1 parent 12f26445ba
commit b7255694cd
221 files changed
+19989 -90

No files matched your search

@@ -0,0 +1,89 @@
using AIStudio.Assistants.ModelComparison;
using AIStudio.Chat;
using AIStudio.Settings;
using AIStudio.Settings.DataModel;
namespace AIStudio.Tests.Assistants.ModelComparison;
/// <summary>
/// Checks that both chats of a task of a comparison series get the same setup.
/// </summary>
[TestFixture]
public sealed class ModelComparisonSeriesThreadsTests
{
[Test]
public void ANewChatHasTheSetupOfTheSeries()
{
var thread = ModelComparisonSeriesThreads.Create(Series(), "Weekly reports");
using (Assert.EnterMultipleScope())
{
Assert.That(thread.Name, Is.EqualTo("Weekly reports"));
Assert.That(thread.WorkspaceId, Is.EqualTo(Guid.Empty));
Assert.That(thread.Blocks, Is.Empty);
Assert.That(thread.SystemPrompt, Is.EqualTo(ModelComparisonSeriesThreads.SYSTEM_PROMPT));
Assert.That(thread.SelectedProfile, Is.EqualTo("profile"));
Assert.That(thread.SelectedChatTemplate, Is.EqualTo(ChatTemplate.NO_CHAT_TEMPLATE.Id));
Assert.That(thread.SelectedToolIds, Is.EquivalentTo(new[] { "web_search" }));
Assert.That(thread.DataSourceOptions.DisableDataSources, Is.False);
Assert.That(thread.DataSourceOptions.PreselectedDataSourceIds, Is.EquivalentTo(new[] { "handbook" }));
Assert.That(thread.DataSourceOptions.RetrievalMode, Is.EqualTo(DataSourceRetrievalMode.EVERY_MESSAGE));
}
}
[Test]
public void TheSystemPromptAsksTheModelNotToNameItself()
{
Assert.That(ModelComparisonSeriesThreads.SYSTEM_PROMPT, Does.StartWith(SystemPrompts.DEFAULT).And.Contains("Do not state your name"));
}
[Test]
public void ARequestGetsTheSetupOfTheSeriesBack()
{
var series = Series();
var thread = ModelComparisonSeriesThreads.Create(series, "chat");
// What a request could have changed, e.g., a stored chat of an older version:
thread.SystemPrompt = "something else";
thread.SelectedProfile = "another profile";
thread.DataSourceOptions = new DataSourceOptions { DisableDataSources = true, AutomaticDataSourceSelection = true };
ModelComparisonSeriesThreads.PrepareForRequest(thread, series, AIStudio.Settings.Provider.NONE, ["web_search"]);
using (Assert.EnterMultipleScope())
{
Assert.That(thread.SystemPrompt, Is.EqualTo(ModelComparisonSeriesThreads.SYSTEM_PROMPT));
Assert.That(thread.SystemPromptOrigin, Is.Null);
Assert.That(thread.SelectedProvider, Is.EqualTo(AIStudio.Settings.Provider.NONE.Id));
Assert.That(thread.SelectedProfile, Is.EqualTo("profile"));
Assert.That(thread.IncludeDateTime, Is.True);
Assert.That(thread.RuntimeSelectedToolIds, Is.EquivalentTo(new[] { "web_search" }));
Assert.That(thread.RuntimeToolsAreAssistantManaged, Is.True);
Assert.That(thread.DataSourceOptions.DisableDataSources, Is.False);
Assert.That(thread.DataSourceOptions.AutomaticDataSourceSelection, Is.False);
Assert.That(thread.DataSourceOptions.AutomaticValidation, Is.False);
}
}
[Test]
public void ASeriesWithoutDataSourcesSearchesNone()
{
var thread = ModelComparisonSeriesThreads.Create(Series() with { DataSourceIds = [], RetrievalMode = null }, "chat");
using (Assert.EnterMultipleScope())
{
Assert.That(thread.DataSourceOptions.DisableDataSources, Is.True);
Assert.That(thread.DataSourceOptions.PreselectedDataSourceIds, Is.Empty);
}
}
private static ModelComparisonSeries Series() => new()
{
Id = Guid.NewGuid(),
Name = "Weekly reports",
ProfileId = "profile",
ToolIds = ["web_search"],
DataSourceIds = ["handbook"],
RetrievalMode = DataSourceRetrievalMode.EVERY_MESSAGE,
};
}