Types of Runs

What we're building

A small pipeline of four traced steps, a fake document search, a word counter, a Gemini call, and a chain that wires them together, each labeled with a run_type and carrying metadata and tags. This gives every run in the tree from Lesson 2 an actual category, instead of everything looking the same.

What this reveals

@traceable accepts a run_type argument: "llm", "chain", "tool", "retriever", and a few others. This isn't just labeling, the LangSmith UI renders each type differently: an "llm" run shows a prompt and completion side by side, a "retriever" run shows the documents it returned, a "tool" run shows the function call and result.

You never had to set run_type="llm" for the model.invoke(...) call in answer_question. LangChain's ChatGoogleGenerativeAI is already instrumented, so when LANGSMITH_TRACING=true, every model call it makes is automatically recorded as an "llm" run and nested under whatever @traceable function called it. Manual @traceable and LangChain's built-in tracing compose together in the same run tree without any extra work.

metadata and tags are separate from run_type: they're your own labels. Metadata is a dictionary of arbitrary key/value pairs. Tags are a flat list of short strings, meant for quick filtering. Lesson 18 uses both to filter and build dashboards.

The code, piece by piece

@traceable(run_type="retriever")
def fake_search(query: str) -> list[str]:

Marks this run as a retriever, so the UI shows its return value (a list of document strings) the way it shows real vector-store results.

@traceable(run_type="tool")
def word_count(text: str) -> int:

Marks this as a tool call, distinct from a chain or an LLM call.

@traceable(
    run_type="chain",
    metadata={"lesson": 3, "topic": "run types"},
    tags=["beginner", "langsmith-course"],
)
def answer_question(question: str) -> dict:

The top-level run. run_type="chain" is actually the default, it's written explicitly here just to show the parameter exists. metadata and tags are attached only to this run, not automatically inherited by its children, each child run can set its own.