langfuse

Langfuse is an LLM observability tool used to monitor LLM behavior in our application. It uses decorators and callbacks to record LLM inference. Always use it in LLM-related modules to ensure traceability.

Langfuse v4 API Guide

As of Langfuse SDK v4.x, the low-level observation API uses start_observation().

Decorator-based usage (unchanged from v3):

from langfuse import observe, get_client

@observe()
def function_name(...):
    get_client().update_current_span(name=f"descriptive_name_{id}")
    trace_id = get_client().get_current_trace_id()
    observation_id = get_client().get_current_observation_id()

Low-level observation creation (v4 pattern):

from langfuse import Langfuse, LangfuseGeneration, LangfuseTool
from langfuse.types import TraceContext

langfuse = Langfuse()

# Create a generation observation
trace_context: TraceContext = {"trace_id": trace_id}
if parent_observation_id:
    trace_context["parent_span_id"] = parent_observation_id

span = langfuse.start_observation(
    name="model_name",
    as_type="generation",  # or "span", "tool", "agent", etc.
    input=user_input,
    trace_context=trace_context,
    metadata={...},
)

# End an observation: update first, then end (v4 .end() only accepts end_time)
span.update(output=result, model=model_name, level="DEFAULT")
span.end()

Key changes from v3:

1. .generation() and .span() removed - use start_observation(as_type=...).
2. .end(**kwargs) no longer accepts output/model/level - split into .update() + .end().
3. Use TraceContext dict with trace_id and optional parent_span_id instead of positional args.
4. New observation types available: LangfuseTool, LangfuseAgent, etc.