byron mackay•6mo ago

My local instance of langfuse is showing

My local instance of langfuse is showing null for input, output, and metadata on the top trace. If I dig down, I see the outputs showing. I'm using a simple LangChain implementation to run an LLM (code in thread). Am I missing something?

6 Replies

byron mackay•6mo ago

template = PromptTemplate(template="{query} Context: {results}", input_variables=["query", "results"])

chain = LLMChain(llm=llm, prompt=template)



print(chain.invoke({"query": "What are the largest cost drivers in 2023?", "results": results}, 
                config={
                        "callbacks": [langfuse_callback(["rag"])]
                    }
                )
    )

template = PromptTemplate(template="{query} Context: {results}", input_variables=["query", "results"])

chain = LLMChain(llm=llm, prompt=template)



print(chain.invoke({"query": "What are the largest cost drivers in 2023?", "results": results}, 
                config={
                        "callbacks": [langfuse_callback(["rag"])]
                    }
                )
    )

Marc•6mo ago

how did you create the langfuse_callback?

Marc•6mo ago

new docs on this here: https://langfuse.com/docs/tracing-features/tags

Tagging traces - Langfuse

Tags help to filter and organize traces in Langfuse based on use case, functions/apis used, environment and other criteria.

byron mackay•6mo ago

def langfuse_callback(tags):
    langfuse = Langfuse()
    trace = langfuse.trace(
        tags=tags
    )
    return trace.get_langchain_handler()

def langfuse_callback(tags):
    langfuse = Langfuse()
    trace = langfuse.trace(
        tags=tags
    )
    return trace.get_langchain_handler()

Marc•6mo ago

ok, this is the expected behavior. if you only want once trace for each langchain run and then include I/O, go for the other implementation

handler = CallbackHandler(
  tags=["tag-1", "tag-2"]
)

handler = CallbackHandler(
  tags=["tag-1", "tag-2"]
)

Without creating a trace manually

byron mackay•6mo ago

Working now. Thanks, @Marc!