Log retriever traces

Log retrieval steps in LangSmith traces for document-level visibility into your RAG pipeline.

These steps are optional. If you skip them, your retriever data will still be logged, but LangSmith will not render it with retriever-specific formatting.

Many LLM applications retrieve documents from vector databases, knowledge graphs, or other indexes as part of a retrieval-augmented generation (RAG) pipeline. LangSmith provides dedicated rendering for retriever steps, which makes it easier to inspect retrieved documents and diagnose retrieval issues.

To enable retriever-specific rendering, complete the following two steps.

Set run_type to retriever

Pass run_type="retriever" to the traceable decorator (Python) or traceable wrapper (TypeScript). This tells LangSmith to treat the step as a retrieval run and apply retriever-specific rendering in the LangSmith UI.

from langsmith import traceable

@traceable(run_type="retriever")
def retrieve_docs(query):
    ...

If you are using the RunTree API instead of traceable, pass run_type="retriever" when creating the RunTree object.

Return documents in the expected format

Return a list of dictionaries (Python) or objects (TypeScript) from your retriever function. Each item in the list represents a retrieved document and must contain the following fields:

Field	Type	Description
page_content	string	The text content of the retrieved document.
type	string	Must always be "Document".
metadata	object	Key-value pairs with metadata about the document, such as source URL, chunk ID, or score. This metadata is displayed alongside the document in the trace.

The following examples show a complete retriever implementation with both requirements applied.

Python

from langsmith import traceable

def _convert_docs(results):
    return [
        {
            "page_content": r,
            "type": "Document",
            "metadata": {"foo": "bar"}
        }
        for r in results
    ]

@traceable(run_type="retriever")
def retrieve_docs(query):
    # Returning hardcoded placeholder documents.
    # In production, replace with a real vector database or document index.
    contents = ["Document contents 1", "Document contents 2", "Document contents 3"]
    return _convert_docs(contents)

retrieve_docs("User query")

The UI displays each retrieved document with its contents and metadata.

Related pages

• Annotate code for tracing: Overview of all tracing methods, including traceable, RunTree, and the REST API.
• Log LLM calls: Similar custom logging requirements for LLM steps.

---

Edit this page on GitHub or file an issue.

Connect these docs to Claude, VSCode, and more via MCP for real-time answers.