Appreciate your videos man. Langgraph js doesnt have great docs so i had to find for a video demonstrating this well. Yours is the only js video i found. Please continue making different types of ai agents using langgraph and ai sdk. One video suggestion could be a hierarchical agentic application.
thanks for making the video! I've found the roundtrip behavior of generateText not so reliable with a particular use case (at least with small models?) so I was thinking of trying to implement round tripping behavior using langgraph instead and thats where the question came up. Basically, I can't tell how much reasoning is occurring between model turns in generateText based on previous context for a particularly challenging agent/prompt use case. My issue was getting the tool calling to work correctly in generateText through different model turns. Anyways thanks and I find the concept of orchestrating these models together really interesting. Would love to see you explore this space more.
I had the same experience when dealing with the react style agents, that's why I like langgraph so much. And keep in mind that smaller models might not be trained for tooluse, that could also cause problems. Groq did some finetunes for tooluse with smaller models, you might want to check them out.
Wow, I really like your approach. I have a suggestion for you: Why don’t you focus on building a RAG application where you are focusing on retrieving data from a Notion file or some doc.
question - do you know how to get type safety here by passing the graph state as a generic somewhere? const result = await graph.invoke({ keyDoesNotExist: 'i do not exist', // no error })
i don't know of a way to get typesafety inside the invoke function but you could maybe do something like this: const invokeArgs: MyStateType = { // the state } const result = await graph.invoke(invokeArgs) its not perfect but better then without any typing...
Appreciate your videos man. Langgraph js doesnt have great docs so i had to find for a video demonstrating this well. Yours is the only js video i found. Please continue making different types of ai agents using langgraph and ai sdk. One video suggestion could be a hierarchical agentic application.
great tutorial as always.
Thank you so much 🙏
Nice! Keep working 🎉
Thank you so much
Hi can you show how use AI SDK streamText or stream object function with Langgraph?
thanks for making the video! I've found the roundtrip behavior of generateText not so reliable with a particular use case (at least with small models?) so I was thinking of trying to implement round tripping behavior using langgraph instead and thats where the question came up. Basically, I can't tell how much reasoning is occurring between model turns in generateText based on previous context for a particularly challenging agent/prompt use case. My issue was getting the tool calling to work correctly in generateText through different model turns. Anyways thanks and I find the concept of orchestrating these models together really interesting. Would love to see you explore this space more.
I had the same experience when dealing with the react style agents, that's why I like langgraph so much.
And keep in mind that smaller models might not be trained for tooluse, that could also cause problems. Groq did some finetunes for tooluse with smaller models, you might want to check them out.
Wait, it's not exactly agents because it's only one flow like pipeline where multi agent are multi direction feedback...
Wow, I really like your approach.
I have a suggestion for you: Why don’t you focus on building a RAG application where you are focusing on retrieving data from a Notion file or some doc.
Thanks for the great idea, I uploaded a video that goes shows how to do this.
question - do you know how to get type safety here by passing the graph state as a generic somewhere?
const result = await graph.invoke({
keyDoesNotExist: 'i do not exist', // no error
})
i don't know of a way to get typesafety inside the invoke function but you could maybe do something like this:
const invokeArgs: MyStateType = {
// the state
}
const result = await graph.invoke(invokeArgs)
its not perfect but better then without any typing...