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 +1,22 @@
namespace AIStudio.Chat.Embedded;
/// <summary>
/// Decides when an embedded chat sends the message it opens with, see EmbeddedChat.OpeningMessage.
/// </summary>
public static class EmbeddedChatOpening
{
/// <summary>
/// Tells whether an embedded chat sends its opening message now.
/// </summary>
/// <remarks>
/// The message may arrive after the chat loaded, e.g., when the parent learns it only then: it is sent all the
/// same, as long as the user wrote nothing in the chat meanwhile. It is sent once per chat, also when sending
/// fails and it waits in the field instead.
/// </remarks>
/// <param name="openingMessage">The message the chat opens with, or null without one.</param>
/// <param name="conversation">The chat shown, or null while it loads.</param>
/// <param name="wasSent">Whether the chat sent its opening message already.</param>
/// <returns>True when the chat sends the message now.</returns>
public static bool SendsNow(ContentText? openingMessage, EmbeddedConversation? conversation, bool wasSent) =>
openingMessage is not null && !wasSent && conversation is { IsAvailable: true, IsRunning: false, UserMessageCount: 0 };
}
@@ -0,0 +1,39 @@
namespace AIStudio.Chat.Embedded;
/// <summary>
/// A message for an embedded chat, together with what the assistant decides about the request it starts.
/// </summary>
public sealed record EmbeddedChatRequest
{
/// <summary>
/// The provider which answers.
/// </summary>
public required AIStudio.Settings.Provider Provider { get; init; }
/// <summary>
/// The message, with its attachments. The chat stores it as it is.
/// </summary>
public required ContentText Message { get; init; }
/// <summary>
/// Whether the user does not see the message, e.g., a request which lets the model open a conversation.
/// </summary>
public bool IsHidden { get; init; }
/// <summary>
/// Whether messages to the user leave out which provider and which model answer, see IProvider.HidesIdentity.
/// </summary>
public bool HidesModelIdentity { get; init; }
/// <summary>
/// Readies the chat for the request before the message goes in: its system prompt, profile, tools, and data
/// sources, which a chat partly does not store. Returns false to send nothing; it tells the user why itself.
/// </summary>
public Func<ChatThread, Task<bool>> PrepareAsync { get; init; } = _ => Task.FromResult(true);
/// <summary>
/// Gets the chat which the request sends, e.g., with images which the stored chat never holds. The chat itself
/// by default.
/// </summary>
public Func<ChatThread, ChatThread> ThreadForRequest { get; init; } = thread => thread;
}
@@ -0,0 +1,336 @@
using System.Collections.Concurrent;
using AIStudio.Provider;
using AIStudio.Settings;
using AIStudio.Tools.AssistantSessions;
namespace AIStudio.Chat.Embedded;
/// <summary>
/// Runs the chats which assistants show on their own pages, such as the two chats of a task of the Model Comparison
/// assistant.
/// </summary>
/// <remarks>
/// A singleton, so that an answer goes on while the user looks at another page: the page asks for the chat again when
/// it comes back and listens to ConversationChanged. Each request registers in AssistantSessionService, which shows the
/// work in the navigation. An embedded chat lives in a store of its assistant, never in a workspace, see
/// IEmbeddedChatStore. What a request carries beside the conversation, such as the system prompt or the tools, the
/// assistant decides for every request, see EmbeddedChatRequest.
/// </remarks>
/// <param name="settingsManager">The settings.</param>
/// <param name="sessionService">The registry of running assistant sessions.</param>
/// <param name="messageBus">The message bus, for errors.</param>
/// <param name="logger">The logger.</param>
/// <param name="answerAsync">Lets the model answer a request; asks the provider by default, and is replaced in tests.</param>
public sealed class EmbeddedChatService(SettingsManager settingsManager, AssistantSessionService sessionService, MessageBus messageBus, ILogger<EmbeddedChatService> logger, Func<EmbeddedChatRequest, ContentText, ChatThread, CancellationToken, Task>? answerAsync = null)
{
/// <summary>
/// How often a streamed answer renders at most.
/// </summary>
private static readonly TimeSpan STREAMING_RENDER_MIN_TIME = TimeSpan.FromMilliseconds(100);
private static string TA(string fallbackEN) => Tools.PluginSystem.I18N.I.TA(fallbackEN, typeof(EmbeddedChatService).Namespace, nameof(EmbeddedChatService));
/// <summary>
/// The chats loaded so far, by their IDs.
/// </summary>
private readonly ConcurrentDictionary<Guid, EmbeddedConversation> conversations = new();
/// <summary>
/// Serializes loading a chat, so that two pages never load the same chat twice.
/// </summary>
private readonly SemaphoreSlim loadLock = new(1, 1);
/// <summary>
/// Guards whether a chat answers right now, so that two requests never run in the same chat.
/// </summary>
private readonly Lock runningLock = new();
private readonly Func<EmbeddedChatRequest, ContentText, ChatThread, CancellationToken, Task> answerAsync = answerAsync ?? AnswerWithProviderAsync;
/// <summary>
/// Happens whenever a chat changed: a message, a chunk of an answer, or the end of a request. It names the ID of
/// the chat.
/// </summary>
/// <remarks>
/// Called on the thread of the request. A page renders through InvokeAsync.
/// </remarks>
public event Func<Guid, Task>? ConversationChanged;
/// <summary>
/// Gets a chat, loading it on the first call.
/// </summary>
/// <param name="store">Where the chat lives.</param>
/// <param name="component">The assistant which shows the chat.</param>
/// <param name="chatId">The ID of the chat.</param>
/// <param name="createThread">Creates the thread of a chat which has none stored yet; it gets the ID of the chat.</param>
/// <param name="token">The cancellation token.</param>
/// <returns>The chat; it is not available when this version cannot open the stored one.</returns>
public async Task<EmbeddedConversation> GetAsync(IEmbeddedChatStore store, Tools.Components component, Guid chatId, Func<ChatThread> createThread, CancellationToken token = default)
{
if (this.conversations.TryGetValue(chatId, out var conversation))
return conversation;
await this.loadLock.WaitAsync(token);
try
{
if (this.conversations.TryGetValue(chatId, out conversation))
return conversation;
ChatThread? thread;
try
{
thread = await store.LoadThreadAsync(chatId, token);
if (thread is null)
{
thread = createThread();
thread.ChatId = chatId;
}
}
catch (InvalidDataException exception)
{
// The file stays as it is, see IEmbeddedChatStore.LoadThreadAsync:
logger.LogWarning(exception, "Could not open the embedded chat '{ChatId}' of the component '{Component}'.", chatId, component);
thread = null;
}
conversation = new EmbeddedConversation { ChatId = chatId, Component = component, Store = store, Thread = thread };
this.conversations[chatId] = conversation;
return conversation;
}
finally
{
this.loadLock.Release();
}
}
/// <summary>
/// Sends a message and lets the model answer it.
/// </summary>
/// <remarks>
/// The message is stored before the model answers, so that a crash loses at most the answer. What goes wrong is
/// reported through the message bus; the method itself does not fail.
/// </remarks>
/// <param name="chatId">The ID of a chat which GetAsync loaded before.</param>
/// <param name="request">The message and what the request carries beside it.</param>
/// <returns>True when the message went into the chat, even when the model could not answer it.</returns>
public async Task<bool> SendAsync(Guid chatId, EmbeddedChatRequest request)
{
if (string.IsNullOrWhiteSpace(request.Message.Text) || !this.conversations.TryGetValue(chatId, out var conversation) || conversation.Thread is not { } thread)
return false;
lock (this.runningLock)
{
if (conversation.IsRunning)
return false;
conversation.IsRunning = true;
}
var messageIsStored = false;
ContentBlock? answerBlock = null;
var sessionKey = SessionKey(conversation);
var sessionId = Guid.Empty;
var status = AssistantSessionStatus.FAILED;
try
{
if (!await request.PrepareAsync(thread) || !await this.CheckProviderAsync(conversation, thread, request.Provider))
return false;
thread.RuntimeComponent = conversation.Component;
thread.Blocks.Add(new ContentBlock
{
Time = DateTimeOffset.Now,
ContentType = ContentType.TEXT,
Role = ChatRole.USER,
Content = request.Message,
HideFromUser = request.IsHidden,
});
await conversation.Store.SaveThreadAsync(thread);
messageIsStored = true;
var answer = new ContentText { InitialRemoteWait = true };
answerBlock = new ContentBlock
{
Time = DateTimeOffset.Now,
ContentType = ContentType.TEXT,
Role = ChatRole.AI,
Content = answer,
};
var lastRendering = DateTimeOffset.MinValue;
answer.StreamingEvent = async () =>
{
var now = DateTimeOffset.Now;
if (now - lastRendering < STREAMING_RENDER_MIN_TIME)
return;
lastRendering = now;
await this.NotifyAsync(chatId);
};
answer.StreamingDone = () => this.NotifyAsync(chatId);
thread.Blocks.Add(answerBlock);
var cancellation = new CancellationTokenSource();
sessionId = (await sessionService.TryBeginAsync(sessionKey, conversation.Component.Name, cancellation, thread, new(StringComparer.Ordinal))).SessionId;
await this.NotifyAsync(chatId);
await this.answerAsync(request, answer, request.ThreadForRequest(thread), cancellation.Token);
status = cancellation.IsCancellationRequested ? AssistantSessionStatus.CANCELED : AssistantSessionStatus.COMPLETED;
}
catch (OperationCanceledException)
{
status = AssistantSessionStatus.CANCELED;
}
catch (ProviderRequestException exception)
{
logger.LogError(exception, "The provider request of an embedded chat failed. Status={StatusCode}, Reason='{ReasonPhrase}'.", exception.StatusCode, exception.ReasonPhrase);
await messageBus.SendError(new(AppIcons.PROVIDER_ERROR, exception.UserMessage));
}
catch (FileNotFoundException)
{
// What the chat belonged to was deleted before the message went in; nothing is left to store.
}
catch (Exception exception) when (exception is IOException or UnauthorizedAccessException or InvalidDataException)
{
logger.LogError(exception, "Could not store the embedded chat '{ChatId}'.", chatId);
await messageBus.SendError(new(AppIcons.SAVE_FAILED, TA("AI Studio could not save this conversation. Please try again.")));
}
finally
{
await this.FinishAsync(conversation, thread, answerBlock, sessionKey, sessionId, status);
}
return messageIsStored;
}
/// <summary>
/// Stops the answer the model writes in a chat.
/// </summary>
/// <param name="chatId">The ID of the chat.</param>
public Task CancelAsync(Guid chatId) => this.conversations.TryGetValue(chatId, out var conversation)
? sessionService.CancelAsync(SessionKey(conversation))
: Task.CompletedTask;
/// <summary>
/// Stops the answer of a chat and forgets the chat, e.g., because the user deleted what it belonged to.
/// </summary>
/// <param name="chatId">The ID of the chat.</param>
public async Task ForgetAsync(Guid chatId)
{
await this.CancelAsync(chatId);
this.conversations.TryRemove(chatId, out _);
}
/// <summary>
/// Lets the provider of a request answer it.
/// </summary>
/// <param name="request">The request.</param>
/// <param name="answer">The answer, which the stream fills.</param>
/// <param name="thread">The chat which the request sends.</param>
/// <param name="token">The cancellation token.</param>
private static async Task AnswerWithProviderAsync(EmbeddedChatRequest request, ContentText answer, ChatThread thread, CancellationToken token)
{
var provider = request.Provider.CreateProvider();
provider.HidesIdentity = request.HidesModelIdentity;
await answer.CreateFromProviderAsync(provider, request.Provider.Model, request.Message, thread, token);
}
/// <summary>
/// Checks whether a provider may answer in a chat, and tells the user why not.
/// </summary>
/// <param name="conversation">The chat.</param>
/// <param name="thread">The thread of the chat, readied for the request.</param>
/// <param name="provider">The provider.</param>
/// <returns>True when the provider may answer.</returns>
private async Task<bool> CheckProviderAsync(EmbeddedConversation conversation, ChatThread thread, AIStudio.Settings.Provider provider)
{
string? reason = null;
if (provider.UsedLLMProvider is LLMProviders.NONE)
reason = TA("Please select a provider.");
else if (!settingsManager.IsProviderConfident(provider, conversation.Component) || !thread.IsLLMProviderAllowed(provider))
reason = TA("The selected provider may not see this conversation. Please select another provider.");
if (reason is null)
return true;
await messageBus.SendError(new(AppIcons.PROVIDER_ERROR, reason));
return false;
}
/// <summary>
/// Ends a request: removes an empty answer, stores the chat, and closes the session.
/// </summary>
/// <param name="conversation">The chat.</param>
/// <param name="thread">The thread of the chat.</param>
/// <param name="answerBlock">The block of the answer, or null when the request did not get that far.</param>
/// <param name="sessionKey">The key of the session.</param>
/// <param name="sessionId">The ID of the session, or Guid.Empty when none began.</param>
/// <param name="status">How the request ended.</param>
private async Task FinishAsync(EmbeddedConversation conversation, ChatThread thread, ContentBlock? answerBlock, AssistantSessionKey sessionKey, Guid sessionId, AssistantSessionStatus status)
{
if (answerBlock is { Content: ContentText answer } && string.IsNullOrWhiteSpace(answer.Text))
thread.Blocks.Remove(answerBlock);
if (answerBlock is not null)
{
try
{
await conversation.Store.SaveThreadAsync(thread);
}
catch (FileNotFoundException)
{
// What the chat belonged to was deleted while the model answered; nothing is left to store.
}
catch (Exception exception) when (exception is IOException or UnauthorizedAccessException or InvalidDataException)
{
logger.LogError(exception, "Could not store the embedded chat '{ChatId}'.", conversation.ChatId);
await messageBus.SendError(new(AppIcons.SAVE_FAILED, TA("AI Studio could not save this conversation. Please try again.")));
}
}
lock (this.runningLock)
conversation.IsRunning = false;
if (sessionId != Guid.Empty)
{
await sessionService.CompleteAsync(sessionKey, sessionId, status, string.Empty, thread, new(StringComparer.Ordinal));
// The assistant keeps the chat itself; the registry only shows that the assistant is at work:
await sessionService.ClearAsync(sessionKey);
}
await this.NotifyAsync(conversation.ChatId);
}
/// <summary>
/// Tells every page which listens that a chat changed.
/// </summary>
/// <param name="chatId">The ID of the chat.</param>
private async Task NotifyAsync(Guid chatId)
{
if (this.ConversationChanged is not { } handlers)
return;
foreach (var handler in handlers.GetInvocationList().Cast<Func<Guid, Task>>())
{
try
{
await handler(chatId);
}
catch (Exception exception)
{
// A page which went away meanwhile must not stop the answer:
logger.LogDebug(exception, "A page could not show a change of an embedded chat.");
}
}
}
/// <summary>
/// Gets the key under which a request of a chat registers in AssistantSessionService.
/// </summary>
/// <param name="conversation">The chat.</param>
private static AssistantSessionKey SessionKey(EmbeddedConversation conversation) => new(conversation.Component, conversation.ChatId.ToString("D"));
}
@@ -0,0 +1,57 @@
namespace AIStudio.Chat.Embedded;
/// <summary>
/// An embedded chat as EmbeddedChatService keeps it, with whether the model answers in it right now.
/// </summary>
public sealed class EmbeddedConversation
{
/// <summary>
/// The ID of the chat.
/// </summary>
public required Guid ChatId { get; init; }
/// <summary>
/// The assistant which shows the chat. Its requests register in AssistantSessionService under it.
/// </summary>
public required Tools.Components Component { get; init; }
/// <summary>
/// Where the chat lives.
/// </summary>
public required IEmbeddedChatStore Store { get; init; }
/// <summary>
/// The thread of the chat, or null when this version cannot open the stored one.
/// </summary>
public ChatThread? Thread { get; internal set; }
/// <summary>
/// Whether the model answers in the chat right now.
/// </summary>
public bool IsRunning { get; internal set; }
/// <summary>
/// Whether the chat could be opened.
/// </summary>
public bool IsAvailable => this.Thread is not null;
/// <summary>
/// Whether the chat holds any message the user sees.
/// </summary>
public bool HasVisibleMessages => this.Thread?.Blocks.Any(block => !block.HideFromUser) ?? false;
/// <summary>
/// How many messages the user wrote in the chat, those the user does not see left aside.
/// </summary>
public int UserMessageCount => this.Thread?.Blocks.Count(block => block is { Role: ChatRole.USER, HideFromUser: false }) ?? 0;
/// <summary>
/// The first message of the user which the user sees, or null before it.
/// </summary>
public ContentBlock? FirstUserMessage => this.Thread?.Blocks.FirstOrDefault(block => block is { Role: ChatRole.USER, HideFromUser: false });
/// <summary>
/// Whether the chat holds an answer of the model which the user sees.
/// </summary>
public bool HasAnswer => this.Thread?.Blocks.Any(block => block is { Role: ChatRole.AI, HideFromUser: false }) ?? false;
}
@@ -0,0 +1,28 @@
namespace AIStudio.Chat.Embedded;
/// <summary>
/// Where the embedded chats of an assistant live, e.g., in a folder of the assistant below the data directory.
/// </summary>
/// <remarks>
/// An embedded chat never lives in a workspace: it belongs to the assistant which shows it, and the assistant decides
/// how long it stays. The store finds a chat by its ID, ChatThread.ChatId.
/// </remarks>
public interface IEmbeddedChatStore
{
/// <summary>
/// Loads a chat.
/// </summary>
/// <param name="chatId">The ID of the chat.</param>
/// <param name="token">The cancellation token.</param>
/// <returns>The chat, or null when it has no stored thread yet.</returns>
/// <exception cref="InvalidDataException">This version cannot open the stored chat; the file stays as it is.</exception>
public Task<ChatThread?> LoadThreadAsync(Guid chatId, CancellationToken token = default);
/// <summary>
/// Stores a chat, in place of what was stored of it before.
/// </summary>
/// <param name="thread">The chat, with its ID.</param>
/// <param name="token">The cancellation token.</param>
/// <exception cref="FileNotFoundException">What the chat belonged to was deleted meanwhile; nothing is left to store.</exception>
public Task SaveThreadAsync(ChatThread thread, CancellationToken token = default);
}