bug: No Pretty traces in Langfuse sessions

[mathg]

Describe the bug

Hello!

I am using Langfuse @observe decorator integration with Langchain + OpenAI LLM to create my traces on Langfuse. I have access to my traces on Langfuse, but it only offers the json view and there is no button for the Pretty view available.

Thanks for the help!

To reproduce

Could it be because this was built using the old Langchain AgentExecutor instead of LangGraph? Here is how I am doing my integration :

...

from langfuse import Langfuse
from langfuse.decorators import langfuse_context, observe
from langfuse.openai import openai

langfuse = Langfuse()
...

@observe()
def run(self, message: str, uid: str, lang: str = "fr", returnOnlyAgent: str = False) -> str:
    
    # make sure uid is a string
    uid = str(uid)
    
    # Set the session ID using the uid
    langfuse_context.update_current_trace(
        session_id=uid,
    )
    
    # Get the Langchain handler from the context
    langfuse_handler = langfuse_context.get_current_langchain_handler()

    llm = ChatOpenAI(openai_api_key=OPENAI_API_KEY, model=OPENAI_CHAT_MODEL_4, temperature=0, callbacks=[langfuse_handler])
    
    langfuse_prompt = langfuse.get_prompt("guidelines")        
    prompt = langfuse_prompt.get_langchain_prompt(today_full_date=self.today_full_date)
    
    ...

    promptMessages = ChatPromptTemplate.from_messages([
        ("system", prompt),
        MessagesPlaceholder(variable_name='chat_history', optional=True),
        HumanMessagePromptTemplate(prompt=PromptTemplate(input_variables=['input'], template='{input}')),
        MessagesPlaceholder(variable_name='agent_scratchpad')]
    )
    agent = create_openai_tools_agent(llm, tools, promptMessages)
    agent_chain_convo_exec = AgentExecutor(agent=agent, tools=tools, session_id=uid, memory=memory, verbose=self.use_verbose)
    
    agent_chain_convo_exec = agent_chain_convo_exec.with_config(
        run_name="ChatBot Production"
    )
    
    response = agent_chain_convo_exec.invoke(input_dict, config={"callbacks":[langfuse_handler]})

SDK and container versions

langchain==0.3.22
langchain-core==0.3.50
langchain-openai==0.3.12
langfuse==2.60.2

Additional information

No response

Are you interested to contribute a fix for this bug?

No

[dosubot]

Hello, @mathg! I'm here to assist you with any bugs, questions, or contributions. Let's dive into the issue you've encountered.

The absence of the Pretty view in Langfuse sessions when using the @observe decorator with Langchain and OpenAI LLM is likely due to the use of the old Langchain AgentExecutor instead of LangGraph. The Pretty view is only available if the trace is identified as a Langgraph trace, which is determined by the isLanggraphTrace function. This function checks if any of the observations' metadata can be parsed using the LanggraphMetadataSchema ^[1].

If the traces generated by AgentExecutor do not meet the criteria for a Langgraph trace, the Pretty view will not be available. To resolve this, consider updating your integration to use LangGraph instead of AgentExecutor, ensuring that the traces conform to the expected format for Langgraph traces ^[2][3][4].

To continue talking to Dosu, mention @dosu.

---

^{Help Dosu learn! Give it feedback: Great Response | Irrelevant Answer | Incorrect Sources | Too Verbose | Hallucination | Bug Report | Other}

[mathg]

Hello, I went back end updated my system to use Langgraph with create_react_agent instead of AgentExecutor but I still don't have pretty traces.

# Set the session ID using the uid
langfuse_context.update_current_trace(
    session_id=uid,
)

# Get the Langchain handler from the context
langfuse_handler = langfuse_context.get_current_langchain_handler()

...

langgraph_agent_executor = create_react_agent(
    llm,
    tools,
    prompt=promptMessages,
    checkpointer=checkpointer
)

input_dict = {
    "messages": [
        ("human", message)
    ]
}


# Get the response
config = {"callbacks": [langfuse_handler], "configurable": {"thread_id": uid}}
response = langgraph_agent_executor.invoke(
    input_dict,
    config=config
)

[hassiebp]

Thanks for raising this! The pretty view is only available if the ingested input / output follows the ChatML schema. Since your messages is not a dict with keys role and content, this will be rendered as raw messages.

If you need further support, please open a support ticket in our Github discussions 👍🏾