LombdaAgentSDK #58
Replies: 8 comments 11 replies
-
|
Okayyyyyy... Made conversion a breeze! ConvertStringToIntState inputState = new();
ConvertStringToIntState resultState = new();
inputState.AddTransition(_ => true, (result) => result.ToString(), resultState); // Middle Parameter is conversion delegate
resultState.AddTransition(_ => true, new ExitState());Also handles the type Safety even with the converted results which is nice. |
Beta Was this translation helpful? Give feedback.
-
|
Wouldn't it be better to use generics here ConvertState<TIn, TOut>? |
Beta Was this translation helpful? Give feedback.
-
|
It is, but It is completely abstracted in the backend public class StateTransition<TInput, TOutput> : StateTransition<TInput>
{
public ConversionMethod<TInput, TOutput> ConverterMethod { get; set; }Then in the State Because the State Inputs and Outputs are generic you can define a State Transition with a new generic Output T by conversion public abstract class BaseState<TInput, TOutput> : BaseState
{
public void AddTransition<T>(TransitionEvent<TOutput> methodToInvoke, ConversionMethod<TOutput, T> conversionMethod, BaseState nextState)
{
var transition = new StateTransition<TOutput, T>(methodToInvoke, conversionMethod, nextState);
Transitions.Add(transition);
}But because the lambda function result is inferred adding the generics on the functions is implied by the lambda result and verified at runtime. So in my first Example above it should be: ConvertStringToIntState inputState = new();
ConvertStringToIntState resultState = new();
inputState.AddTransition<string>(_ => true, (result) => result.ToString(), resultState); // Middle Parameter is conversion delegate
resultState.AddTransition(_ => true, new ExitState()); |
Beta Was this translation helpful? Give feedback.
-
|
Massive update pushed to master.. Finally made my way to actually making a basic UI to see implementation which required LombdaAgent the master controller for the whole verbose and streaming system as well as a master agent on that to control the conversation. Obsoleted the openai c# lib in favor for yours.. added a very bad UI window form for baby AGI.. and most important I had the AI make me a API for communicating with LombdaAgent which completes most of the work flow .. I want to have the API client stand alone and somehow (possibly another server) that controls the custom Agents people create without having to fork my API project and injecting theirs. I found with the structured Outputs I need to figure out 2 things..
New
Fixed
Known Issues
|
Beta Was this translation helpful? Give feedback.
-
|
Thanks for the update, how would you like to serialize enums? I can look into that. Btw, I'm working on accepting delegates natively as tool inputs / structured schema: |
Beta Was this translation helpful? Give feedback.
-
|
Nice I love just accepting the delegate very much like the python SDK from open ai.. Yeah to be honest that is probably the only code I had AI generated for my project was the json function schema generator.. other than the new API.. I’m new to json serialization but I did figure out how it works pretty and improved it.. simple actually..and Oh wow dictionary.. never considered that type! I actually today accepted a key, value parameter to generate metadata which essentially does dictionary output (for vector db metadata on baby AGI) … and for enums I just want to abstract away having to use the tag.. my current idea would be to use reflections to determine which json converter I should use but would love a cleaner solution. |
Beta Was this translation helpful? Give feedback.
-
New interface for interacting with agents using MAUI and the API
(I created new repo for the maui project) 
New updates to my LLMTornado ForkChallenges
|
Beta Was this translation helpful? Give feedback.
-
@BoHomola implemented smart upcasting of requests from /chat/completions to /responses endpoint. This allows us to use models that are exclusive to responses from the This also works with streaming, and native So basically, you can either continue doing this:
Or you can use just ResponseRequestParameters = new ResponseRequest
{
// config
}To [TornadoTest]
public static async Task StreamResponseSimpleTextUsingChat()
{
Conversation chat = Program.Connect().Chat.CreateConversation(new ChatRequest
{
Model = ChatModel.OpenAi.O4.V4Mini,
MaxTokens = 4000,
// using Responses endpoint is preferred, if the model supports it
ResponseRequestParameters = new ResponseRequest
{
Reasoning = new ReasoningConfiguration
{
Effort = ResponseReasoningEfforts.Medium,
Summary = ResponseReasoningSummaries.Auto
}
}
})
.AppendSystemMessage("You are a helpful assistant")
.AppendUserInput([
new ChatMessagePart("How to explain theory of relativity to a 15 years old student?")
]);
// try calling Serialize() on "chat" with and without "ResponseRequestParameters"
// and observe the url and body of the request
await chat.StreamResponseRich(new ChatStreamEventHandler
{
OnResponseEvent = (evt) =>
{
// "evt" is IResponseEvent
return ValueTask.CompletedTask;
},
MessageTokenHandler = (delta) =>
{
Console.Write(delta);
return ValueTask.CompletedTask;
},
ReasoningTokenHandler = (reasoningDelta) =>
{
Console.Write(reasoningDelta.Content);
return ValueTask.CompletedTask;
}
});
}I've also made good progress on the delegates - now we support almost any type, allow conditionally excluding certain parameters from the request, overriding/adding custom parameters to the request (e.g., in case you have an enum with many options, but want the model to choose only from a subset): [TornadoTest]
public static async Task TornadoFunction()
{
Conversation conversation = Program.Connect().Chat.CreateConversation(new ChatRequest
{
Model = ChatModel.OpenAi.Gpt41.V41,
Tools =
[
new Tool((
string location,
Continents continent,
ComplexClass cls,
List<string> names,
List<Person> people,
string[] popularGames,
string[,] wonGameOfCheckers3x3useXOchars,
Continents[] allContinents,
string[][] rpgInventoryItemsUseXForEmpty,
HashSet<int> setOfUniqueInts,
object someDataAboutGames,
DateTime dateBattleOfVerdunStarted,
ToolArguments args) => // ToolArguments is a special type that allows accessing the inferred arguments via IDictionary interface
{
// manual decoding example, normally not needed
if (args.TryGetArgument("people", out List<Person>? fetchedPeople))
{
foreach (Person person in fetchedPeople)
{
Console.WriteLine(person.Name);
}
}
return "";
}, new ToolMetadata
{
Params = [
new ToolParamDefinition("allContinents", new ToolParamListEnum("continents", true, [ nameof(Continents.Africa), nameof(Continents.Antarctica) ]))
],
Ignore = [ "wonGameOfCheckers3x3useXOchars" ]
})
],
ToolChoice = OutboundToolChoice.Required
});
conversation.AddUserMessage("Fill the provided JSON structure with mock data");
TornadoRequestContent serialized = conversation.Serialize(new ChatRequestSerializeOptions
{
Pretty = true
});
Console.Write(serialized);
var data = await conversation.GetResponseRich();
int z = 0;
}It should be ready to use in the next version 3.8 - I guess in one or two days. |
Beta Was this translation helpful? Give feedback.
-
|
Okay I'm just going to stick with implementing Chat.. what about Sending the response of the computer-use-preview? I'm using that a lot, it's the whole reason I started writing C# code for AI. |
Beta Was this translation helpful? Give feedback.
-
|
@lofcz 100x easier for me now that you handle the delegate.. I am abstracting a couple more attributes to get Description of function and ignoring parameters [Description("Get the current weather in a given location")]
public static string GetCurrentWeather(
[Description("The city and state, e.g. Boston, MA")] string location,
[IgnoreParam] Unit unit = Unit.celsius)
{
// Call the weather API here.
return $"31 C";
}public static Tool ConvertFunctionToTornadoTool(this Delegate function)
{
MethodInfo method = function.Method;
List<string> required_inputs = new List<string>();
string toolDescription = method.Name;
if (method.IsDefined(typeof(DescriptionAttribute), false))
{
toolDescription = method.GetCustomAttributes<DescriptionAttribute>().First().Description;
}
var parameters = method.GetParameters();
ToolMetadata toolMetadata = new ToolMetadata();
if(toolMetadata.Ignore == null) toolMetadata.Ignore = new List<string>();
foreach (ParameterInfo param in parameters)
{
if (param.Name == null) continue;
if (!param.HasDefaultValue)
{
required_inputs.Add(param.Name);
}
if (param.IsDefined(typeof(IgnoreParamAttribute), inherit: false))
{
if(!toolMetadata.Ignore.Contains(param.Name))
{
toolMetadata.Ignore.Add(param.Name);
}
}
}
var strictSchema = required_inputs.Count == parameters.Length;
return new Tool(function, method.Name, toolDescription, toolMetadata, strictSchema);
} |
Beta Was this translation helpful? Give feedback.
-
|
We can add IgnoreParam to the core as SchemaIgnoreAttribute if you open a PR with that. However, ignoring parametrs most often needs to be done runtime, based on context, e.g. if user doesn't require certain information the tool would normally work with, it is dynamically ignored, but we can still have an attribute for cases known at compile time. |
Beta Was this translation helpful? Give feedback.
-
|
@lofcz i accidently send merge request before seeing this. I'll submit second PR but I want to look into the dynamic ignore a bit. I was having a bit of trouble figuring out the ToolMetadata Params i have yet to read that code just yet. |
Beta Was this translation helpful? Give feedback.
-
|
I will send you links to some commented demos tommorow so you understand how it works. I believe the dynamic ToolMetadata is the most cool thing about the entire way we handle delegates and enables some powerful patterns beyond what SK can do. |
Beta Was this translation helpful? Give feedback.
-
|
@Johnny2x2 FYI, we now have our own page: https://llmtornado.ai |
Beta Was this translation helpful? Give feedback.
-
|
Website is looking great! I was wondering what the DOC project was.. I don't have .net10 installed so it breaks my build so I typically get rid of it lol. |
Beta Was this translation helpful? Give feedback.
-
|
Texts on the website are lorem ipsum, final textation is TBD. I currently have a semi-working Monaco (VS Code frontend) with Roslyn-based intellisense implemented. This can consume the "LlmTornado" assembly and provide real-time intellisense in your browser (via web workers, so it doesn't lag the UI thread). The run button takes the code from Monaco, compiles it via Roslyn, packs it via WebCIL into wasm, and executes it dynamically. Still a lot of work to do, but it's coming along. The final vision is for the users to simply enter their API key to whatever provider they want to toy with, and experiment with LlmTornado from their browser. The Vitepress part of the docs will have code examples with a "play" button, which redirects you to the playground and sets the initial code to the example. 
|
Beta Was this translation helpful? Give feedback.
-
LlmTornado.Agents@lofcz I have completed the LLMTornado Agent Chat Integration from my lib to yours!! LlmTornado.Agents I'm still working out the Response Lib Testing.. Check it out let me know what you think! I added demos as well so you can try it out. also added a research agent demo for the state machine. Lombda.StateMachineState Machine now has it's own Repo and Nuget Demo CodeBasic [TornadoTest]
public static async Task BasicTornadoRun()
{
TornadoAgent agent = new(
Program.Connect(),
ChatModel.OpenAi.Gpt41.V41Mini,
"You are a useful assistant.");
var result = await TornadoRunner.RunAsync(agent, "What is 2+2?");
Console.WriteLine(result.Messages.Last().Content);
}Streaming[TornadoTest]
public static async Task RunHelloWorldStreaming()
{
TornadoAgent agent = new(
Program.Connect(),
ChatModel.OpenAi.Gpt41.V41Mini,
"Have fun");
// Enhanced streaming callback to handle the new ModelStreamingEvents system
void StreamingHandler(ModelStreamingEvents streamingEvent)
{
switch (streamingEvent.EventType)
{
case ModelStreamingEventType.OutputTextDelta:
if (streamingEvent is ModelStreamingOutputTextDeltaEvent deltaEvent)
{
Console.Write(deltaEvent.DeltaText); // Write the text delta directly
}
break;
default:
break;
}
}
var result = await TornadoRunner.RunAsync(agent, "Hello Streaming World!", streaming: true, streamingCallback: StreamingHandler);
}Tools and Text formating[TornadoTest]
public async Task RunBasicTornadoToolUse()
{
TornadoAgent agent = new TornadoAgent(Program.Connect(),
ChatModel.OpenAi.Gpt41.V41Mini,
"You are a useful assistant.",
tools: [GetCurrentWeather],
outputSchema: typeof(WeatherReport));
var result = await TornadoRunner.RunAsync(agent, "What is the weather in boston?");
Console.WriteLine(result.Messages.Last().Content.ParseJson<WeatherReport>());
} |
Beta Was this translation helpful? Give feedback.
-
|
Looks good! Can you please open a PR so I can review this? At first glance, can we move
Thanks for working on this. |
Beta Was this translation helpful? Give feedback.
-
|
@lofcz I can work on looking at the hand offs and visualization. For the hand offs:if we have an AgentState and lets say 2 other AgentStates as hand offs I can have an AI response either post input or pre input to select another state based off the current input or output. Almost just like a state before or after a certain state to choose where to go next. VisualizationI designed a node graph in windows form for designing opencv workflows which I've been considering adding in as a way to visualize which state you are in.. but windows form is windows only UI so that is why ive been working on making everything send messages via API so i can use any UI |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Hello Again! Figured I'd keep you up to date on my SDK LombdaAgentSDK here instead of on the Issue thread. I have pushed a Beta BabyAGI project with the self build principle. Just finished wiring it up tonight still playing around with it but it is "working".
Because of the typesafe nature of the state transitions sometimes switching between states can be a pain because you need to transition to a state to convert the data to the correct type.. I came up with a new primative to convert the data a bit easier without having to create a whole new state. I guess it is more flexible than just for conversion.
Added in a couple new things to get BabyAGI to work
Beta Was this translation helpful? Give feedback.
All reactions