John G.

SimpleStudent - apologies for the delay here, that makes sense In order to upload the results to Phoenix, you'd either need to start from a dataset and use an experiment, or to start from traces that you pull down from Phoenix, evaluate, and push back up. We don't have a way to simply upload metrics at this time

As Mikyo said, LLM classify and experiments don't exactly work together super simply right now LLM classify is designed to evaluate a dataframe, and experiments send individual rows through your evaluator methods at one time. What you need to do generally is to create a one-row dataframe in your experiment evaluators, and then send that through LLM Classify. The notebook I shared uses that approach

The key here, and where your issue is coming from, is that your df must have column names that match the variables in your eval_prompt. Based on the error you sent, your prompt is looking for a column named input that doesn't exist in your dataframe

Hey SimpleStudent - I'd actually recommend looking into this example: https://github.com/Arize-ai/phoenix/blob/main/tutorials/experiments/agents-cookbook.ipynb . The "Create an Experiment" section shows an example of using LLM classify in the evaluator of an experiment

Answering in Phoenix Support!

Luca P. - it looks like this duplication is a bug on our end, likely not an issue with your instrumentation. The team is repro'ing today and should pick it up this coming week. Apologies for all the issues here!

Ah okay think I found it - can you upgrade to openinference-instrumentation-smolagents==0.1.13 We added better support for the generate function and base model invocation approaches. I'm seeing that remove the repeat issue - though I only saw it repeat a total of 3 calls

Okay it doesn't look like there's a general problem with duplicate calls on those versions, I'm seeing the correct traces. Going to try and match your setup more closely. Are you using the InferenceClientModel or InferenceClient object within smolagents, and which model(s) are you using?

Okay that's great - trying to replicate this now!