LangSmith: @traceable and Client Wrappers

Runs, run types and nesting

The @traceable decorator turns a function call into a LangSmith run. Like Langfuse's decorator, it captures arguments as inputs, the return value as outputs, timing and errors, and nested traceable calls become child runs in the same trace. The run_type tells the UI how to display the run: llm runs show messages and token counts, retriever runs render returned documents, tool runs show tool arguments and results, and the default chain is a generic step. You can also pass name, tags and metadata to the decorator, or per call through the special langsmith_extra argument, for example to set a precomputed run_id you will later attach feedback to. For model calls, wrap_openai() wraps an OpenAI client so every completion is logged as an llm run with messages, parameters and token usage, nested under the current traceable function. Because uploads happen in a background thread, the decorator adds little latency to your request path.

Tracing a small pipeline

Requires langsmith and openai with keys set. A retriever run returns documents shaped with page_content and metadata so the UI renders them nicely.

from langsmith import traceable
from langsmith.wrappers import wrap_openai
from openai import OpenAI

client = wrap_openai(OpenAI())  # each call becomes an llm run

@traceable(run_type="retriever")
def retrieve(question: str) -> list[dict]:
    return [{"page_content": "Refunds within 7 days.", "metadata": {"source": "faq.md"}}]

@traceable(name="support-answer", tags=["support"], metadata={"prompt_version": "v4"})
def answer(question: str) -> str:
    docs = retrieve(question)
    context = "\n".join(d["page_content"] for d in docs)
    resp = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {"role": "system", "content": f"Answer only from:\n{context}"},
            {"role": "user", "content": question},
        ],
    )
    return resp.choices[0].message.content

answer("Can I get a refund after 10 days?")

Pick run types deliberately

Leaving everything as the default chain works, but you lose the specialised views: retrieved documents are shown as raw JSON and model calls do not show token counts. Mark retrieval functions retriever, tool functions tool and direct model calls llm.

त्वरित जाँच: You want LangSmith to display retrieved documents nicely for a vector-search function. Which run type fits?

• llm
• parser
• retriever
• prompt