LangSmith makes it easy to attach feedback to traces.
This feedback can come from users, annotators, automated evaluators, etc., and is crucial for monitoring and evaluating applications.
Use create_feedback() / createFeedback
Here we’ll walk through how to log feedback using the SDK.
Child runs
You can attach user feedback to ANY child run of a trace, not just the trace (root run) itself.
This is useful for critiquing specific steps of the LLM application, such as the retrieval step or generation step of a RAG pipeline.
Non-blocking creation (Python only)
The Python client will automatically background feedback creation if you pass trace_id= to create_feedback().
This is essential for low-latency environments, where you want to make sure your application isn’t blocked on feedback creation. from langsmith import trace, traceable, Client
@traceable
def foo(x):
return {"y": x * 2}
@traceable
def bar(y):
return {"z": y - 1}
client = Client()
inputs = {"x": 1}
with trace(name="foobar", inputs=inputs) as root_run:
result = foo(**inputs)
result = bar(**result)
root_run.outputs = result
trace_id = root_run.id
child_runs = root_run.child_runs
# Provide feedback for a trace (a.k.a. a root run)
client.create_feedback(
key="user_feedback",
score=1,
trace_id=trace_id,
comment="the user said that ..."
)
# Provide feedback for a child run
foo_run_id = [run for run in child_runs if run.name == "foo"][0].id
client.create_feedback(
key="correctness",
score=0,
run_id=foo_run_id,
# trace_id= is optional but recommended to enable batched and backgrounded
# feedback ingestion.
trace_id=trace_id,
)
create_feedback() / createFeedback. See this guide for how to get the run ID of an in-progress run.
To learn more about how to filter traces based on various attributes, including user feedback, see this guide.
