Mark B.

Thanks! Additionally here is an image of an incorrect point with a green label. Just as extra

Roger Y. Thanks for your reply. Wrapping the ainvoke function with the with line does create a span from which I can get the ID, but it does not really do what I would like. If I keep the LangChainInstrumentor().instrument(tracer_provider=tracer_provider) line in my code, it just creates a wrapper around it and Phoenix shows an empty trace, which can be unfolded and then shows the LangChain traces, but not what I'm looking for, because the UI then doesn't show the input and output. (See attached image) If I remove the LangChainInstrumentor() line, then it is just an empty trace... I'd just like to attach an annotation to the top-most LangChain logged trace. So is it seems like that there is a bug in from openinference.instrumentation.langchain import get_current_span which should be fixed. Otherwise quite a lot of other users of Phoenix that also use LangChain have no way of logging feedback as annotations on their traces... Edit: I'm on version 0.1.29 of openinference-instrumentation-langchain

The config seems to be an empty dictionary

Hi Roger Y., thanks for sharing that snippet of code. Unfortunately the get_current_span returns None, so that doesn't seem to work...

Hey 🔒[private user] , thanks for your reply! That is a bit unfortunate with the autoinstrumentor... Do you possibly know another way to log LangChain traces to Phoenix, but also save the feedback received from the end users? That is basically all I'm looking for. The LangChainInstrumentor made the logging easy, but I find it strange that Phoenix would otherwise have such a cumbersome way of storing the received user feedback

Unfortunately RunLLM couldn't help me answering the first question related to getting the span_id. My generate_answer.py file contains the following code which is called from a FastAPI endpoint in main.py :

async def generate_llm_answer() -> str:
	from opentelemetry import trace
	
    prompt = ChatPromptTemplate.from_messages(
        [
            ("system", prompts.SYSTEM_PROMPT),
            ("human", prompts.EXPANDED_PROMPT),
        ]
    )

    rag_chain = ...

    answer = await rag_chain.ainvoke(summarized_question)
    
	span = trace.get_current_span()
    span_id = span.get_span_context().span_id.to_bytes(8, "big").hex()
	
    return answer

But this doesn't get a valid span_id, because span is a NonRecordingSpan. I've tried the LLM answer, but that seems to generate a new span with a new ID which doesn't match the one already recorded by the LangChainInstrumentor

RunLLM LangChainInstrumentor().instrument(tracer_provider=tracer_provider) sets a tracer. How can I get the current span in a function from this LangChainInstrumentor?

RunLLM 3. Is there a Python package to send feedback to Phoenix? Or is it only possible via HTTP?