splunk-otel-util-genai-translator-langsmith 0.1.5

Langsmith -> GenAI translator emitter for OpenTelemetry GenAI

This package provides automatic translation of Langsmith-specific span attributes to OpenTelemetry GenAI semantic convention compliant format.

Overview

Langsmith is the observability platform for LangChain applications, providing tracing, evaluation, and monitoring capabilities. This translator bridges Langsmith’s attribute format to the standardized GenAI semantic conventions.

Installation

pip install splunk-otel-util-genai-translator-langsmith

Usage

The translator automatically enables when the package is installed. No additional configuration is required.

To explicitly enable or configure:

from opentelemetry.util.genai.langsmith import enable_langsmith_translator

# Enable with default configuration
enable_langsmith_translator()

# Enable with custom attribute transformations
enable_langsmith_translator(
    attribute_transformations={
        "rename": {
            "my.custom.attr": "gen_ai.custom.attr"
        }
    }
)

Disabling Auto-Enable

Set the environment variable OTEL_INSTRUMENTATION_GENAI_LANGSMITH_DISABLE=true to disable automatic registration.

Attribute Mappings

The translator performs the following attribute mappings:

Content & Messages

• langsmith.entity.input → gen_ai.input.messages

• langsmith.entity.output → gen_ai.output.messages

• gen_ai.prompt → gen_ai.input.messages

• gen_ai.completion → gen_ai.output.messages

Model & System

• langsmith.metadata.ls_provider → gen_ai.system

• langsmith.metadata.ls_model_name → gen_ai.request.model

• langsmith.metadata.ls_model_type → gen_ai.operation.name

Request Parameters

• langsmith.metadata.ls_temperature → gen_ai.request.temperature

• langsmith.metadata.ls_max_tokens → gen_ai.request.max_tokens

• langsmith.metadata.ls_top_p → gen_ai.request.top_p

• langsmith.metadata.ls_top_k → gen_ai.request.top_k

Token Usage

• langsmith.token_usage.prompt_tokens → gen_ai.usage.input_tokens

• langsmith.token_usage.completion_tokens → gen_ai.usage.output_tokens

• langsmith.token_usage.total_tokens → gen_ai.usage.total_tokens

Tool Calling

• langsmith.tool.name → gen_ai.tool.call.name

• langsmith.tool.id → gen_ai.tool.call.id

• langsmith.tool.arguments → gen_ai.tool.call.arguments

• langsmith.tool.output → gen_ai.tool.call.result

Session & Run Tracking

• langsmith.session_id → gen_ai.conversation.id

• langsmith.thread_id → gen_ai.conversation.id

• langsmith.run_id → gen_ai.run.id

Known Limitations

• Root agent/workflow span duplicates the full conversation. LangSmith serializes the entire MessagesState accumulator as the root span’s output, so gen_ai.output.messages on the top-level invoke_agent span contains every turn of the conversation rather than just the final reply. Child chat and nested agent spans carry the correct per-turn messages.

• Agent vs. workflow classification depends on naming. LangSmith reports every orchestration span with run_type=chain regardless of whether it represents an agent or a higher-level workflow. The translator infers the intent from the span name: roots whose name contains agent or executor (or that already carry gen_ai.agent.name) are classified as invoke_agent; other root chains are classified as invoke_workflow; inner chains become step. To get correct classification, name your agents accordingly — e.g. create_agent(name="weather_agent") or compile(name="research_agent") for single-agent graphs, and a neutral name like compile(name="weather_assistant") for multi-agent workflows.

References

• Langsmith Documentation

• GenAI Semantic Conventions