Hey everyone. I found the multimodal support page for Phoenix in the docs: https://arize.com/docs/phoenix/tracing/how-to-tracing/advanced/multimodal-tracing But it only covers images. Does that mean we cannot view audio samples in the traces?
Hey John G.. Thanks for your comment. Here is the dataset format I'm working with:
chat_history = [
{
"role": "user",
"content": "What's the capital of France?",
"created_at": "2023-01-01T00:00:00Z",
},
{
"role": "assistant",
"content": "The capital of France is Paris.",
"created_at": "2023-01-01T00:00:01Z",
},
{
"role": "user",
"content": "What's the population?",
"created_at": "2023-01-01T00:00:02Z",
},
{
"role": "assistant",
"content": "About 2.1 million as of 2023.",
"created_at": "2023-01-01T00:00:03Z",
},
]here's a script I wrote to attempt to do the conversion:
def make_span(
name, kind, start, end, span_id, trace_id, parent_id=None, attributes=None
):
return Span(
name=name,
context=SpanContext(trace_id=trace_id, span_id=span_id),
span_kind=kind,
start_time=start,
end_time=end,
parent_id=None,
status_code=StatusCode.OK,
status_message=None,
attributes=attributes or {},
events=[], # optional structured log messages like function call, exception, etc.
conversation=None, # or: SpanConversationAttributes(conversation_id=some_uuid)
)
def chat_to_trace(chat_history):
spans = []
for i in range(0, len(chat_history), 2):
user_msg = chat_history[i]
assistant_msg = (
chat_history[i + 1] if i + 1 < len(chat_history) else {"content": ""}
)
span_id = str(uuid.uuid4())
trace_id = str(uuid.uuid4())
llm_model_name = "gpt-4"
llm_provider = "openai"
# Read timestamps from chat_history created_at fields
user_created_at = user_msg.get("created_at")
assistant_created_at = assistant_msg.get("created_at")
# Use created_at timestamps if available, otherwise generate them
start = datetime.fromisoformat(user_created_at.replace("Z", "+00:00")).replace(
tzinfo=None
)
end = datetime.fromisoformat(
assistant_created_at.replace("Z", "+00:00")
).replace(tzinfo=None)
span = {
"name": "LLM Response",
"span_kind": "LLM",
"parent_id": None,
"start_time": start.isoformat() + "Z",
"end_time": end.isoformat() + "Z",
"status_code": "OK",
"status_message": "",
"events": [],
"context.span_id": span_id,
"context.trace_id": trace_id,
# Core attributes
"attributes.input.value": user_msg["content"],
"attributes.output.value": assistant_msg["content"],
"attributes.input.mime_type": "text/plain",
"attributes.output.mime_type": "text/plain",
"attributes.openinference.span.kind": "llm",
# Optional LLM-specific attributes
"attributes.llm.model_name": llm_model_name,
"attributes.llm.provider": llm_provider,
"attributes.llm.input_messages": [user_msg],
"attributes.llm.output_messages": [assistant_msg],
"attributes.llm.token_count.prompt": 10,
"attributes.llm.token_count.completion": 15,
"attributes.llm.token_count.total": 25,
# Mocked or empty values for now
"attributes.llm.invocation_parameters": {},
"attributes.llm.system": "",
"attributes.llm.tools": [],
"attributes.session.id": str(uuid.uuid4()),
"attributes.user.id": "anonymous",
"attributes.metadata": {},
# Deep token details (placeholders)
"attributes.llm.token_count.prompt_details.audio": 0,
"attributes.llm.token_count.prompt_details.cache_read": False,
"attributes.llm.token_count.completion_details.reasoning": 10,
"attributes.llm.token_count.completion_details.audio": 0,
}
spans.append(span)
return spans
spans = chat_to_trace(chat_history)
spans_df = pd.DataFrame(spans)When I attempt to log this, I don't see any errors but I don't see the logs appearing on Phoenix as well:
px.Client().log_traces(
trace_dataset=px.TraceDataset(spans_df), project_name="my_project_name"
)Could you share why that might be happening? If you already have a conversion script, that would be super helpful.
Hi everyone. I have recently moved to a self-hosted instance of Phoenix and wondering how can I move my previous data (user conversations with an LLM) to the traces for annotation. I see that I can upload it as a dataset and edit the rows individually later on if I am not satisfied with the response. But I would prefer uploading it to Traces and have multiple team members leave their annotations on it instead. I would highly prefer to store it as a ChatCompletion object where the inputs and outputs are nicely formatted. It feels like a very common use-case that anyone moving to Phoenix might have but I couldn't find any resource for it. Any guidance on what can be done here?