Rebuilt how AI Studio knows what a model can do (#960)
Build and Release / Determine run mode (push) Waiting to run
Build and Release / Sync Flatpak repo (push) Blocked by required conditions
Build and Release / Collect Flatpak artifacts (push) Blocked by required conditions
Build and Release / Verify (push) Waiting to run
Build and Release / Build app (${{ matrix.dotnet_runtime }}) (-aarch64-apple-darwin, osx-arm64, macos-latest, aarch64-apple-darwin, dmg,app,updater, dmg) (push) Blocked by required conditions
Build and Release / Build app (${{ matrix.dotnet_runtime }}) (-aarch64-pc-windows-msvc.exe, win-arm64, windows-latest, aarch64-pc-windows-msvc, nsis,updater, nsis) (push) Blocked by required conditions
Build and Release / Build app (${{ matrix.dotnet_runtime }}) (-aarch64-unknown-linux-gnu, linux-arm64, ubuntu-22.04-arm, aarch64-unknown-linux-gnu, appimage,updater, appimage) (push) Blocked by required conditions
Build and Release / Build app (${{ matrix.dotnet_runtime }}) (-x86_64-apple-darwin, osx-x64, macos-latest, x86_64-apple-darwin, dmg,app,updater, dmg) (push) Blocked by required conditions
Build and Release / Build app (${{ matrix.dotnet_runtime }}) (-x86_64-pc-windows-msvc.exe, win-x64, windows-latest, x86_64-pc-windows-msvc, nsis,updater, nsis) (push) Blocked by required conditions
Build and Release / Build app (${{ matrix.dotnet_runtime }}) (-x86_64-unknown-linux-gnu, linux-x64, ubuntu-22.04, x86_64-unknown-linux-gnu, appimage,updater, appimage) (push) Blocked by required conditions
Build and Release / Prepare & create release (push) Blocked by required conditions
Build and Release / Publish release (push) Blocked by required conditions

This commit is contained in:
Thorsten Sommer authored and GitHub committed 2026-09-13 14:17:25 +02:00
1 parent d21e09dd1e
commit d85b4e71b6
287 files changed
+18341 -3677

No files matched your search

+242
View File
@@ -0,0 +1,242 @@
using AIStudio.Chat;
namespace AIStudio.Tests.Chat;
/// <summary>
/// Checks what a conversation is counted as costing before it is sent.
/// </summary>
/// <remarks>
/// The number under the input field used to count the sentence being typed and nothing else, which
/// answers a question nobody asks: what decides whether the next message fits is everything that
/// travels with it. So what is collected here has to be what the message builder actually sends --
/// no more, because a number which counts something that stays behind is wrong in the direction
/// that makes a person stop writing.
/// </remarks>
[TestFixture]
public sealed class ConversationPartsTests
{
private string directory = string.Empty;
[SetUp]
public void CreateFiles()
{
this.directory = Path.Combine(Path.GetTempPath(), $"ai-studio-tests-{Guid.NewGuid():N}");
Directory.CreateDirectory(this.directory);
}
[TearDown]
public void RemoveFiles()
{
if (Directory.Exists(this.directory))
Directory.Delete(this.directory, true);
}
[Test]
public void TheWholeConversationCountsAndNotOnlyWhatIsBeingTyped()
{
var thread = new ChatThread
{
SystemPrompt = "You are helpful.",
Blocks =
[
Block("What is the capital of France?"),
Block("Paris."),
],
};
var parts = ConversationParts.Of(thread, "You are helpful.", "And of Italy?", null, imagesAreSent: true);
Assert.Multiple(() =>
{
Assert.That(parts.Texts, Is.EqualTo(new[] { "You are helpful.", "What is the capital of France?", "Paris." }));
Assert.That(parts.GrowingTexts, Is.EqualTo(new[] { "And of Italy?" }));
});
}
[Test]
public void WhatIsStillBeingWrittenIsKeptApartFromWhatStands()
{
//
// Both cost the same and both are counted. They are kept apart because of what happens
// afterwards: a message which stands says the same thing forever and its count is worth
// remembering, while the answer being streamed is a different text three seconds later.
//
var streaming = Block("The answer so far");
((ContentText)streaming.Content!).IsStreaming = true;
var thread = new ChatThread { Blocks = [Block("A question."), streaming] };
var parts = ConversationParts.Of(thread, string.Empty, "a draft", null, imagesAreSent: true);
Assert.Multiple(() =>
{
Assert.That(parts.Texts, Is.EqualTo(new[] { "A question." }));
Assert.That(parts.GrowingTexts, Is.EqualTo(new[] { "The answer so far", "a draft" }));
});
}
[Test]
public void AnAnswerWhichIsFinishedStandsLikeAnyOtherMessage()
{
var finished = Block("The whole answer.");
((ContentText)finished.Content!).IsStreaming = false;
var parts = ConversationParts.Of(new() { Blocks = [finished] }, string.Empty, string.Empty, null, imagesAreSent: true);
Assert.Multiple(() =>
{
Assert.That(parts.Texts, Is.EqualTo(new[] { "The whole answer." }));
Assert.That(parts.GrowingTexts, Is.Empty);
});
}
[Test]
public void TheSystemPromptCountedIsTheOneWhichWouldBeSent()
{
//
// Not the one standing in the thread. A chat template may replace it, retrieved data is
// appended to it, a profile adds a paragraph and the tool policy adds another -- and
// switching a profile while writing has to move the number, which it cannot do if the
// thread's own field is what gets counted.
//
var thread = new ChatThread { SystemPrompt = "What the person typed." };
var parts = ConversationParts.Of(thread, "What the request carries.", string.Empty, null, imagesAreSent: true);
Assert.That(parts.Texts, Is.EqualTo(new[] { "What the request carries." }));
}
[Test]
public void ABlockHiddenFromTheUserStillCosts()
{
//
// Hidden on the screen, not in the request: the message builder sends it like any other
// block, so its tokens are gone whether or not anybody can see where they went.
//
var hidden = Block("An instruction the user does not see.");
var thread = new ChatThread { Blocks = [new() { ContentType = hidden.ContentType, Role = hidden.Role, Content = hidden.Content, HideFromUser = true }] };
var parts = ConversationParts.Of(thread, string.Empty, string.Empty, null, imagesAreSent: true);
Assert.That(parts.Texts, Is.EqualTo(new[] { "An instruction the user does not see." }));
}
[Test]
public void WithoutAConversationOnlyTheDraftCounts()
{
var parts = ConversationParts.Of(null, string.Empty, "Hello", null, imagesAreSent: true);
Assert.Multiple(() =>
{
Assert.That(parts.Texts, Is.Empty);
Assert.That(parts.GrowingTexts, Is.EqualTo(new[] { "Hello" }));
});
}
[TestCase("")]
[TestCase(" ")]
public void NothingWrittenIsNothingToCount(string draft)
{
var parts = ConversationParts.Of(null, string.Empty, draft, null, imagesAreSent: true);
Assert.Multiple(() =>
{
Assert.That(parts.Texts, Is.Empty);
Assert.That(parts.GrowingTexts, Is.Empty);
});
}
[Test]
public void ABlockWithoutTextIsSkippedBecauseItIsNeverSent()
{
//
// The message builder drops a block whose text is empty, whatever else hangs off it. A
// count which added that block's attachments would report tokens for a message which is
// never built.
//
var document = this.WriteFile("notes.txt", "some content");
var empty = Block(string.Empty);
((ContentText)empty.Content!).FileAttachments.Add(FileAttachment.FromPath(document));
var parts = ConversationParts.Of(new() { Blocks = [empty] }, string.Empty, string.Empty, null, imagesAreSent: true);
Assert.Multiple(() =>
{
Assert.That(parts.Texts, Is.Empty);
Assert.That(parts.Documents, Is.Empty);
});
}
[Test]
public void AttachmentsOfTheConversationAndOfTheComposerBothCount()
{
//
// A document attached three messages ago is sent again with every further message, so it
// costs its tokens again every time. That is exactly the thing a person cannot see and
// which this number is for.
//
var older = this.WriteFile("older.txt", "older content");
var draft = this.WriteFile("draft.txt", "draft content");
var block = Block("Please read this.");
((ContentText)block.Content!).FileAttachments.Add(FileAttachment.FromPath(older));
var parts = ConversationParts.Of(new() { Blocks = [block] }, string.Empty, "And this one.", [FileAttachment.FromPath(draft)], imagesAreSent: true);
Assert.That(parts.Documents.Select(document => document.FileName), Is.EqualTo(new[] { "older.txt", "draft.txt" }));
}
[Test]
public void AnAttachmentWhoseFileIsGoneCountsForNothing()
{
//
// It is not sent either: the message builder reports it as unavailable and leaves it out.
//
var attachment = FileAttachment.FromPath(Path.Combine(this.directory, "never-existed.txt"));
var parts = ConversationParts.Of(null, string.Empty, "Here", [attachment], imagesAreSent: true);
Assert.That(parts.Documents, Is.Empty);
}
[Test]
public void ImagesAreCountedSeparatelyFromDocuments()
{
var document = this.WriteFile("notes.txt", "content");
var image = this.WriteFile("photo.png", "not really a png");
var parts = ConversationParts.Of(null, string.Empty, "Look", [FileAttachment.FromPath(document), FileAttachment.FromPath(image)], imagesAreSent: true);
Assert.Multiple(() =>
{
Assert.That(parts.Documents.Select(entry => entry.FileName), Is.EqualTo(new[] { "notes.txt" }));
Assert.That(parts.Images, Is.EqualTo(1));
});
}
[Test]
public void AModelWhichTakesNoImagesIsSentNoneAndIsToldAboutNone()
{
//
// The message builder leaves the pictures out entirely for such a model, so reporting them
// as uncounted would tell a person about a cost which is not there.
//
var image = this.WriteFile("photo.png", "not really a png");
var parts = ConversationParts.Of(null, string.Empty, "Look", [FileAttachment.FromPath(image)], imagesAreSent: false);
Assert.That(parts.Images, Is.Zero);
}
private static ContentBlock Block(string text) => new()
{
ContentType = ContentType.TEXT,
Role = ChatRole.USER,
Content = new ContentText { Text = text },
};
private string WriteFile(string name, string content)
{
var path = Path.Combine(this.directory, name);
File.WriteAllText(path, content);
return path;
}
}
@@ -0,0 +1,210 @@
using AIStudio.Chat;
namespace AIStudio.Tests.Chat;
/// <summary>
/// Checks when the token count is recomputed and when it is not.
/// </summary>
/// <remarks>
/// This is the part which kept going wrong. The number used to be wired to the places which change
/// the conversation -- fifteen of them in the end -- and four review rounds each found another place
/// which had been forgotten. So it is no longer wired to anything: whoever suspects a change nudges,
/// and what is checked here is that the tracker turns those nudges into the right amount of work.
///
/// The waits are generous on purpose. What is asserted is the behaviour, not the clock, so every
/// interval here is far enough apart that a busy build machine cannot turn one into the other.
/// </remarks>
[TestFixture]
public sealed class ConversationTokenTrackerTests
{
/// <summary>
/// A heartbeat which never fires, for the tests which are about nudges alone.
/// </summary>
private static readonly TimeSpan NO_HEARTBEAT = Timeout.InfiniteTimeSpan;
[Test]
public async Task ManyNudgesInARowCostOneCount()
{
//
// Loading a chat touches several things one after the other, and every one of them renders.
// Counting once per render would measure the same conversation half a dozen times.
//
var runs = 0;
var firstRun = new TaskCompletionSource();
await using var tracker = new ConversationTokenTracker(_ =>
{
Interlocked.Increment(ref runs);
firstRun.TrySetResult();
return Task.CompletedTask;
}, () => TimeSpan.FromSeconds(2), NO_HEARTBEAT);
tracker.Start();
for (var i = 0; i < 50; i++)
tracker.Nudge();
await firstRun.Task.WaitAsync(TimeSpan.FromSeconds(5));
await Task.Delay(200);
Assert.That(runs, Is.EqualTo(1));
}
[Test]
public async Task ANudgeArrivingDuringACountLeadsToExactlyOneMore()
{
//
// Something changed while we were reading, so the answer we just worked out may already be
// out of date -- but only one further count can be needed, however many nudges arrived.
//
var runs = 0;
var firstRunStarted = new TaskCompletionSource();
var releaseFirstRun = new TaskCompletionSource();
await using var tracker = new ConversationTokenTracker(async _ =>
{
if (Interlocked.Increment(ref runs) is not 1)
return;
firstRunStarted.TrySetResult();
await releaseFirstRun.Task;
}, () => TimeSpan.FromMilliseconds(100), NO_HEARTBEAT);
tracker.Start();
tracker.Nudge();
await firstRunStarted.Task.WaitAsync(TimeSpan.FromSeconds(5));
//
// Nudged while the first run is held up, five times over, because a render storm is what
// this has to survive.
//
for (var i = 0; i < 5; i++)
tracker.Nudge();
releaseFirstRun.SetResult();
await Task.Delay(1_000);
Assert.That(runs, Is.EqualTo(2));
}
[Test]
public async Task WithoutAnyNudgeTheHeartbeatStillCounts()
{
//
// For what happens outside AI Studio: an attached file somebody edits in another program
// changes what the next message costs, and nothing here renders because of it.
//
var runs = 0;
await using var tracker = new ConversationTokenTracker(_ =>
{
Interlocked.Increment(ref runs);
return Task.CompletedTask;
}, () => TimeSpan.FromMilliseconds(50), TimeSpan.FromMilliseconds(200));
tracker.Start();
await Task.Delay(1_000);
Assert.That(runs, Is.GreaterThanOrEqualTo(2));
}
[Test]
public async Task TheQuietTimeIsAskedAnewAfterEveryRun()
{
//
// Because the right answer changes with what is going on. Showing a new number renders, and
// a render nudges, so while something moves continuously this delay is the entire cadence
// -- and a chat which is waiting for a model wants a slower one than a chat which is not.
//
var runs = 0;
var asked = 0;
await using var tracker = new ConversationTokenTracker(_ =>
{
Interlocked.Increment(ref runs);
return Task.CompletedTask;
}, () =>
{
Interlocked.Increment(ref asked);
return TimeSpan.FromMilliseconds(50);
}, TimeSpan.FromMilliseconds(100));
tracker.Start();
await Task.Delay(1_000);
Assert.Multiple(() =>
{
Assert.That(runs, Is.GreaterThanOrEqualTo(2));
Assert.That(asked, Is.EqualTo(runs));
});
}
[Test]
public async Task NothingIsCountedAfterTheTrackerIsGone()
{
var runs = 0;
var tracker = new ConversationTokenTracker(_ =>
{
Interlocked.Increment(ref runs);
return Task.CompletedTask;
}, () => TimeSpan.FromMilliseconds(50), TimeSpan.FromMilliseconds(100));
tracker.Start();
await Task.Delay(400);
await tracker.DisposeAsync();
var afterDisposal = runs;
await Task.Delay(400);
Assert.Multiple(() =>
{
Assert.That(afterDisposal, Is.GreaterThan(0), "The tracker never ran, so this proves nothing about stopping it.");
Assert.That(runs, Is.EqualTo(afterDisposal));
});
}
[Test]
public async Task ACountWhichHangsDoesNotHoldUpDisposal()
{
//
// Counting ends in an IPC call to the runtime, and a component going away must not wait for
// one which is not coming back. The token handed to the work is the way out, and this is
// the test that it really is one.
//
var running = new TaskCompletionSource();
var tracker = new ConversationTokenTracker(async token =>
{
running.TrySetResult();
await Task.Delay(Timeout.InfiniteTimeSpan, token);
}, () => TimeSpan.FromMilliseconds(50), NO_HEARTBEAT);
tracker.Start();
tracker.Nudge();
await running.Task.WaitAsync(TimeSpan.FromSeconds(5));
var disposal = tracker.DisposeAsync().AsTask();
var finishedInTime = await Task.WhenAny(disposal, Task.Delay(TimeSpan.FromSeconds(2))) == disposal;
Assert.That(finishedInTime, Is.True);
}
[Test]
public async Task AFailedCountDoesNotEndTheTracker()
{
//
// A tracker which died on one bad answer would leave a stale number standing forever, which
// is the one failure this whole mechanism exists to rule out.
//
var runs = 0;
await using var tracker = new ConversationTokenTracker(_ =>
{
Interlocked.Increment(ref runs);
throw new InvalidOperationException("The tokenizer did not answer.");
}, () => TimeSpan.FromMilliseconds(50), TimeSpan.FromMilliseconds(100));
tracker.Start();
await Task.Delay(1_000);
Assert.That(runs, Is.GreaterThanOrEqualTo(2));
}
}
+96
View File
@@ -0,0 +1,96 @@
using AIStudio.Chat;
using AIStudio.Models;
namespace AIStudio.Tests.Chat;
/// <summary>
/// Checks what the chat says about the images a conversation carries.
/// </summary>
/// <remarks>
/// The visual briefing refuses a build with too many pictures, because that build is expensive and
/// fails late. A chat cannot refuse anything: the pictures are already in the conversation, and
/// taking them back out means deleting messages. So the chat says so instead, and what is checked
/// here is that it says so at the right moment -- and, more importantly, that it stays quiet when
/// nobody wrote a limit down.
/// </remarks>
[TestFixture]
public sealed class ConversationTokensTests
{
[TestCase(1, 100, false)]
[TestCase(100, 100, false, Description = "Exactly the limit still fits. It is a maximum, not a threshold.")]
[TestCase(101, 100, true)]
[TestCase(3_601, 3_600, true)]
public void TooManyPicturesIsAQuestionOfTheNumberTheVendorStated(int images, int allowed, bool tooMany)
{
var counted = new ConversationTokens
{
IsKnown = true,
UncountedImages = images,
ImageLimits = new ImageLimits(null, allowed),
};
Assert.That(counted.TooManyImages, Is.EqualTo(tooMany));
}
[Test]
public void WithoutAStatedLimitThereIsNoSuchThingAsTooMany()
{
//
// The common case. Most models are served at whatever their operator configured, and an app
// which warned about the seventh picture would be inventing a ceiling nobody wrote.
//
var counted = new ConversationTokens
{
IsKnown = true,
UncountedImages = 500,
ImageLimits = ImageLimits.UNKNOWN,
};
Assert.That(counted.TooManyImages, Is.False);
}
[Test]
public void TheSmallerOfTwoStatedLimitsIsTheOneWhichDecides()
{
//
// A message is part of a request, so a conversation which fits the request limit can still
// be too much for one message. Both are compared against the same number of pictures,
// because a chat sends all of them in one message.
//
var counted = new ConversationTokens
{
IsKnown = true,
UncountedImages = 20,
ImageLimits = new ImageLimits(8, 100),
};
Assert.Multiple(() =>
{
Assert.That(counted.TooManyImages, Is.True);
Assert.That(counted.ImageLimits.MaxInOneMessage, Is.EqualTo(8));
});
}
[Test]
public void AConversationWithoutPicturesNeverComplainsAboutThem()
{
var counted = new ConversationTokens
{
IsKnown = true,
UncountedImages = 0,
ImageLimits = new ImageLimits(null, 0),
};
Assert.That(counted.TooManyImages, Is.False, "Not even against a model which takes none at all.");
}
[Test]
public void AnUnavailableCountClaimsNothingAboutPictures()
{
//
// Nothing could be counted, so nothing is known -- including how many pictures travel. A
// warning built on that would be made up.
//
Assert.That(ConversationTokens.UNAVAILABLE.TooManyImages, Is.False);
}
}
@@ -0,0 +1,65 @@
using AIStudio.Chat;
using AIStudio.Provider;
using AIStudio.Provider.OpenAI;
namespace AIStudio.Tests.Chat;
/// <summary>
/// Checks that writing a message asks the same question as attaching the picture did.
/// </summary>
/// <remarks>
/// These were two questions until now. Attaching a file asked the configured provider, so a person's
/// expert settings counted; building the message asked the automatic answer alone, so they did not.
/// Somebody who switched image input on for their own installation watched the picture attach and
/// then watched it disappear on the way to the model -- every chat round and every tool round, with
/// nothing anywhere saying why.
/// </remarks>
[TestFixture]
public sealed class ListContentBlockExtensionsTests
{
[Test]
public async Task ImageInputSwitchedOnByHandReachesTheMessageAsWell()
{
var provider = new AIStudio.Settings.Provider(0, "test", "Test", LLMProviders.SELF_HOSTED, new Model("llama3.3:70b", null))
{
// The rules say this model reads text only, which is what makes it the right model here:
CapabilityOverrides = new() { MultipleImageInput = true },
};
var messages = await BlocksWithAPicture().BuildMessagesAsync(provider, _ => "user", Text, Picture);
Assert.That(messages.Single(), Is.InstanceOf<MultimodalMessage>(), "The picture is part of the message because the person said this model can read one.");
}
[Test]
public async Task WithoutThatSwitchThePictureStaysOut()
{
var provider = new AIStudio.Settings.Provider(0, "test", "Test", LLMProviders.SELF_HOSTED, new Model("llama3.3:70b", null));
var messages = await BlocksWithAPicture().BuildMessagesAsync(provider, _ => "user", Text, Picture);
Assert.That(messages.Single(), Is.InstanceOf<TextMessage>(), "Nothing says this model reads pictures, so the text goes on its own.");
}
/// <summary>
/// One block of text with a picture hanging on it.
/// </summary>
/// <returns>The blocks.</returns>
private static List<ContentBlock> BlocksWithAPicture() =>
[
new()
{
Role = ChatRole.USER,
ContentType = ContentType.TEXT,
Content = new ContentText
{
Text = "What is in this picture?",
FileAttachments = [new FileAttachmentImage("picture.png", "/tmp/picture.png", 1_024)],
},
},
];
private static ISubContent Text(string text) => new SubContentText { Text = text };
private static Task<ISubContent> Picture(FileAttachmentImage image) => Task.FromResult<ISubContent>(new SubContentText { Text = image.FileName });
}
+54
View File
@@ -0,0 +1,54 @@
using System.Globalization;
using AIStudio.Chat;
namespace AIStudio.Tests.Chat;
/// <summary>
/// Checks how a number of tokens is written under the input field.
/// </summary>
/// <remarks>
/// The culture is an argument rather than something taken from the machine, and that is the point
/// being checked as much as the digits are: AI Studio's language is chosen in its own settings, so
/// the thread's culture says nothing about which separators a person expects to read.
/// </remarks>
[TestFixture]
public sealed class TokenAmountTests
{
private static readonly CultureInfo AMERICAN = CultureInfo.GetCultureInfo("en-US");
private static readonly CultureInfo GERMAN = CultureInfo.GetCultureInfo("de-DE");
[TestCase(0, "0")]
[TestCase(7, "7")]
[TestCase(847, "847")]
[TestCase(999, "999", Description = "The last number written out in full.")]
[TestCase(1_000, "1.00k")]
[TestCase(1_234, "1.23k")]
[TestCase(12_347, "12.35k")]
[TestCase(128_000, "128.00k")]
[TestCase(400_000, "400.00k")]
[TestCase(999_499, "999.50k")]
[TestCase(999_999, "1.00M", Description = "Rounded before the unit is chosen, so it does not read as 1,000.00k.")]
[TestCase(1_000_000, "1.00M")]
[TestCase(1_048_576, "1.05M")]
[TestCase(1_050_000, "1.05M", Description = "Which is how OpenAI writes it themselves.")]
[TestCase(2_000_000, "2.00M")]
public void ANumberOfTokensIsWrittenTheWayItIsRead(int tokens, string wanted)
{
Assert.That(TokenAmount.Format(tokens, AMERICAN), Is.EqualTo(wanted));
}
[TestCase(999, "999")]
[TestCase(1_234, "1,23k")]
[TestCase(400_000, "400,00k")]
[TestCase(1_048_576, "1,05M")]
public void TheSeparatorsAreTheOnesTheUserKnows(int tokens, string wanted)
{
//
// A German reads 1,23k where an American reads 1.23k. Writing either of them the other way
// around reads as a number a thousand times off.
//
Assert.That(TokenAmount.Format(tokens, GERMAN), Is.EqualTo(wanted));
}
}