langgraph/concepts/agentic_concepts/

[giscus[bot]]

langgraph/concepts/agentic_concepts/

Build language agents as graphs

https://langchain-ai.github.io/langgraph/concepts/agentic_concepts/

[Pythonista7]

How can i use Structured Outputs with the builtin ReAct agent? since it takes in a model and not LlmLike , i cant really add llm.with_structured_output() , was just running some experiments and im sure this is possible by defining a custom graph with a node having a runnable to create structured outputs. But would be nice to have it at an agent output level as well with the builtin agent :)

[hwchase17]

with the built in ReAct agent you cant. What you would you want/expect? To be able to specify a schema and then the agent would alway return that schema?

[Pythonista7]

@hwchase17 Yes was pretty much trying to do exactly that

graph = create_react_agent(model=llm,tools=self.tools,messages_modifier=prompt) | RunnableLambda( 
            lambda data: {"text": str(data["messages"][-1].content)}
        ) | parser_chain

And had a parser_chain using the PydanticOutputParser to get the structured output.

[lukajose]

I was trying to do the same, did you find a solution?

[yunhzou]

You can use structured output to generate
class ReACT(BaseModel):
thought: str = Field(description="One thought process.Each thought is only one sentence and should have a objective, you must follow this rule")
action: str = Field(description="An action should be taken to achieve the thought, only one action is allowed. Describe the action in one sentence, best to include the action name you proposed")
terminate: bool = Field(description="When the target has been achieved or the question has been answered, set this to True, otherwise False")

and connect an executor agent(REACT agent or openai tool agent) that equip with tools to look at action description at each step and call the tools if the action can be solved with tools equipped.

[yunhzou]

This implementation lead to something like this.
Assistant: : I need to check the current weather in San Francisco.,
: perform a web search for current weather in San Francisco.,
: False
Assistant: : The current weather in San Francisco is sunny. However, if you're a Gemini, you might want to be cautious! 😈
Assistant: : The current weather in San Francisco is sunny.,
: No action is required.,
: True

[ashantanu]

what can we do for caching?
langchain had LLMCaching but that doesn't seem to be working for me.

[rlancemartin]

Have a look at persistence documentation:
https://langchain-ai.github.io/langgraph/concepts/persistence/#persistence

[yadgire7]

Can I create an agent using LlamaCpp and bind tools to it?

[rlancemartin]

Yes, see here:
https://www.youtube.com/watch?v=rsDlu-9UP00

And here:
https://python.langchain.com/docs/integrations/chat/llamacpp/

If you want to use LlamaCpp tool calling w/ LangChain integration.

You can also use LlamaCpp directly w/o LangChain, as LangGraph does not require you to use LangChain.

Here is an example graph using Ollama (rather than LlamaCpp) to show how local llms + LangGraph can work:
https://github.com/langchain-ai/langgraph/blob/main/examples/rag/langgraph_rag_agent_llama3_local.ipynb

[WesGBrooks]

LOVE the new memory capabilities. For chatbot use-cases, is there anything coming out sometime soon to be able to manage conversation flow across multiple messages with the user? Let's say there's 5 questions you want the user to work through with the agent. Are there any paradigms you've seen for effective graph controller design to manage the flow of the conversation across multiple graph instantiations? i.e. Google's DialogFlow describes the flow of the conversation something like this.

I've been exploring a "conversation flow manager" which keeps track of the state of the conversation via a field in the graph state, but it seems like there has to be a better way to manage conversation state [not just graph run state] reliably. Am I missing a useful paradigm with the current tools to do this effectively?