add_span_attributes

add_span_attributes

Adds custom span-level attributes to a specific runnable component’s metadata.

Span attributes are key-value pairs that apply to a specific component (LLM, tool, or retriever) and are included in the OpenTelemetry spans created when that component executes. Use this to add metadata that describes the component’s configuration, purpose, or operational context.

Unlike session attributes (which apply to all spans in a context), span attributes are scoped to individual components. This is useful for:

• Identifying which model or tool is being used

• Tagging components by purpose or category

• Adding version information or deployment metadata

• Tracking A/B test variants or experimental configurations

The attributes are stored in the component’s metadata dictionary under the key ‘_fiddler_attributes’ and will be automatically included in spans when the component executes. Attributes persist for the lifetime of the component instance.

Supported component types:

• BaseLanguageModel: LLM calls (ChatOpenAI, ChatAnthropic, etc.)

• BaseRetriever: Document retrieval operations

• BaseTool: Tool/function calls in agent workflows

Parameters

Parameter	Type	Required	Default	Description
node	BaseLanguageModel | BaseRetriever | BaseTool	✓	-	The LangChain component to annotate with custom attributes. The component’s metadata will be modified in place.

Returns

None

Return type: None

Examples

Tagging an LLM with model information:

from langchain_openai import ChatOpenAI
from fiddler_langgraph.tracing.instrumentation import add_span_attributes

llm = ChatOpenAI(model="gpt-4")
add_span_attributes(
    llm,
    model_name="gpt-4",
    provider="openai",
    purpose="summarization"
)

Adding version and environment metadata:

add_span_attributes(
    llm,
    version="v2.1.0",
    environment="production",
    region="us-west-2"
)

Tagging tools in a multi-tool agent:

from langchain.tools import Tool

search_tool = Tool(
    name="search",
    func=search_function,
    description="Search the web"
)
add_span_attributes(
    search_tool,
    tool_category="external_api",
    rate_limit="100/min",
    cost_per_call=0.001
)

A/B testing different retrievers:

from langchain_community.vectorstores import FAISS

retriever_a = FAISS.from_documents(docs, embeddings).as_retriever()
add_span_attributes(
    retriever_a,
    variant="semantic_search",
    experiment_id="exp_2024_q1",
    retrieval_strategy="similarity"
)

retriever_b = FAISS.from_documents(docs, embeddings).as_retriever(
    search_type="mmr"
)
add_span_attributes(
    retriever_b,
    variant="mmr_search",
    experiment_id="exp_2024_q1",
    retrieval_strategy="maximum_marginal_relevance"
)

Combining with session attributes:

from fiddler_langgraph.tracing.instrumentation import (
    add_session_attributes,
    add_span_attributes,
    set_conversation_id
)

# Session-level: applies to all spans
set_conversation_id("conv_12345")
add_session_attributes("user_id", "user_789")

# Span-level: applies only to this LLM's spans
llm = ChatOpenAI(model="gpt-4-turbo")
add_span_attributes(
    llm,
    model_tier="premium",
    use_case="customer_support"
)

• Attributes are stored in the component’s metadata dictionary, which persists for the lifetime of the component instance

• If the component doesn’t have a metadata attribute, one will be created

• Multiple calls to add_span_attributes on the same component will merge attributes

• Later calls with the same key will overwrite previous values

• This modifies the component in place - no need to reassign the return value