mirror of
https://github.com/MindWorkAI/AI-Studio.git
synced 2026-11-21 02:33:48 +00:00
Fixed a misleading tool calling error after a stopped or interrupted answer (#1202)
Co-authored-by: Thorsten Sommer <SommerEngineering@users.noreply.github.com>
This commit is contained in:
1 parent
fb8f51dec4
commit
ca7413d84c
18 files changed
+272
-76
No files matched your search
@@ -101,7 +101,9 @@ public sealed class ChatCompletionToolCallAccumulator(Func<ServerSentEvent, ILis
|
||||
/// </returns>
|
||||
/// <remarks>
|
||||
/// The end of the stream is the end of the message. There is nothing else to wait for: a
|
||||
/// "[DONE]" is not sent by every gateway, and a finish reason not by every one either.
|
||||
/// "[DONE]" is not sent by every gateway, and a finish reason not by every one either. That
|
||||
/// holds for a stream which the provider closed only. A stream which was cut off never gets
|
||||
/// here: its adapter ends the round before, cf. <see cref="ServerSentEventStreamEnd"/>.
|
||||
/// </remarks>
|
||||
public ChatCompletionResponseMessage? Build()
|
||||
{
|
||||
|
||||
@@ -18,7 +18,7 @@ public sealed class ChatCompletionToolCallingAdapter<TRequest>(
|
||||
TextMessage systemPrompt, IDictionary<string, object> apiParameters,
|
||||
IList<object> providerTools, bool mayAskForSequentialToolCalls,
|
||||
IReadOnlyList<(ToolDefinition Definition, IToolImplementation Implementation)> runnableTools,
|
||||
Func<ChatCompletionAPIRequest, CancellationToken, IAsyncEnumerable<ServerSentEvent>> streamRequestAsync,
|
||||
Func<ChatCompletionAPIRequest, ServerSentEventStreamEnd, CancellationToken, IAsyncEnumerable<ServerSentEvent>> streamRequestAsync,
|
||||
Func<ServerSentEvent, IList<ISource>> readSources,
|
||||
ILogger logger)
|
||||
: IToolCallingProviderAdapter where TRequest : ChatCompletionAPIRequest
|
||||
@@ -71,14 +71,19 @@ public sealed class ChatCompletionToolCallingAdapter<TRequest>(
|
||||
// behind it, fragment by fragment. The usage goes out with every round: which of them
|
||||
// describes the conversation is the loop's decision, which knows which round this is.
|
||||
//
|
||||
var streamEnd = new ServerSentEventStreamEnd();
|
||||
var accumulator = new ChatCompletionToolCallAccumulator(readSources);
|
||||
await foreach (var serverSentEvent in streamRequestAsync(requestDto, token))
|
||||
await foreach (var serverSentEvent in streamRequestAsync(requestDto, streamEnd, token))
|
||||
{
|
||||
var part = accumulator.Process(serverSentEvent);
|
||||
if (part.HasContent || part.Usage.IsKnown)
|
||||
yield return ToolCallingStreamEvent.TextDelta(new ContentStreamChunk(part.TextDelta, part.Sources, Usage: part.Usage));
|
||||
}
|
||||
|
||||
// A stream which was cut off ends the round without an outcome, cf. IToolCallingProviderAdapter.ExecuteRoundAsync:
|
||||
if (!streamEnd.IsReached)
|
||||
yield break;
|
||||
|
||||
var message = accumulator.Build();
|
||||
if (message is null)
|
||||
yield break;
|
||||
|
||||
@@ -207,7 +207,7 @@ public sealed class ProviderOpenAI() : BaseProvider(LLMProviders.OPEN_AI, new Ur
|
||||
baseInput,
|
||||
additionalApiParameters,
|
||||
runnableTools,
|
||||
(requestDto, requestToken) => this.StreamResponsesRequest(requestDto, requestedSecret, requestToken));
|
||||
(requestDto, streamEnd, requestToken) => this.StreamResponsesRequest(requestDto, requestedSecret, streamEnd, requestToken));
|
||||
|
||||
var loop = Program.SERVICE_PROVIDER.GetRequiredService<IToolCallingLoop>();
|
||||
var loopContext = new ToolCallingLoopContext
|
||||
@@ -296,9 +296,9 @@ public sealed class ProviderOpenAI() : BaseProvider(LLMProviders.OPEN_AI, new Ur
|
||||
/// classification come from the shared stream reader, which the tool calling rounds used to
|
||||
/// go without; reading the events is the adapter's business.
|
||||
/// </remarks>
|
||||
private IAsyncEnumerable<ServerSentEvent> StreamResponsesRequest(ResponsesAPIRequest requestDto, RequestedSecret requestedSecret, CancellationToken token)
|
||||
private IAsyncEnumerable<ServerSentEvent> StreamResponsesRequest(ResponsesAPIRequest requestDto, RequestedSecret requestedSecret, ServerSentEventStreamEnd streamEnd, CancellationToken token)
|
||||
{
|
||||
return this.ReadServerSentEventsAsync("OpenAI", "responses call", RequestBuilder, token);
|
||||
return this.ReadServerSentEventsAsync("OpenAI", "responses call", RequestBuilder, streamEnd, token);
|
||||
|
||||
async Task<HttpRequestMessage> RequestBuilder()
|
||||
{
|
||||
|
||||
@@ -106,8 +106,10 @@ public sealed class ResponsesStreamAccumulator
|
||||
return null;
|
||||
|
||||
//
|
||||
// No closing event came, so the round is put back together from the items which did.
|
||||
// Reasoning items are among them, which is what the next request needs to continue.
|
||||
// No closing event came, although the provider closed the stream: some gateways never send
|
||||
// it. So the round is put back together from the items which did come. Reasoning items are
|
||||
// among them, which is what the next request needs to continue. A stream which was cut off
|
||||
// never gets here: its adapter ends the round before, cf. ServerSentEventStreamEnd.
|
||||
//
|
||||
return new ResponsesResponse
|
||||
{
|
||||
|
||||
@@ -15,7 +15,7 @@ namespace AIStudio.Provider.OpenAI;
|
||||
/// </remarks>
|
||||
public sealed class ResponsesToolCallingAdapter(Model chatModel, IList<object> baseInput, IDictionary<string, object> apiParameters,
|
||||
IReadOnlyList<(ToolDefinition Definition, IToolImplementation Implementation)> runnableTools,
|
||||
Func<ResponsesAPIRequest, CancellationToken, IAsyncEnumerable<ServerSentEvent>> streamRequestAsync) : IToolCallingProviderAdapter
|
||||
Func<ResponsesAPIRequest, ServerSentEventStreamEnd, CancellationToken, IAsyncEnumerable<ServerSentEvent>> streamRequestAsync) : IToolCallingProviderAdapter
|
||||
{
|
||||
private readonly List<object> internalItems = [];
|
||||
private readonly List<string> recordedRequestTexts = [];
|
||||
@@ -59,14 +59,19 @@ public sealed class ResponsesToolCallingAdapter(Model chatModel, IList<object> b
|
||||
// goes out with every round: which of them describes the conversation is the loop's
|
||||
// decision, which knows which round this is.
|
||||
//
|
||||
var streamEnd = new ServerSentEventStreamEnd();
|
||||
var accumulator = new ResponsesStreamAccumulator();
|
||||
await foreach (var serverSentEvent in streamRequestAsync(request, token))
|
||||
await foreach (var serverSentEvent in streamRequestAsync(request, streamEnd, token))
|
||||
{
|
||||
var part = accumulator.Process(serverSentEvent);
|
||||
if (part.HasContent || part.Usage.IsKnown)
|
||||
yield return ToolCallingStreamEvent.TextDelta(new ContentStreamChunk(part.TextDelta, part.Sources, Usage: part.Usage));
|
||||
}
|
||||
|
||||
// A stream which was cut off ends the round without an outcome, cf. IToolCallingProviderAdapter.ExecuteRoundAsync:
|
||||
if (!streamEnd.IsReached)
|
||||
yield break;
|
||||
|
||||
var response = accumulator.Build();
|
||||
if (response is null)
|
||||
yield break;
|
||||
|
||||
Reference in new issue
Block a user