Attribute Mapping Reference

When MLflow ingests traces from third-party OpenTelemetry instrumentation, it automatically translates framework-specific span attributes into MLflow's standard format. This ensures that key information like span types, inputs/outputs, token usage, and model names display correctly in the MLflow UI regardless of the originating framework.

How Translation Works

When a trace is received via the OTLP endpoint, MLflow runs each span through a series of translators. Each translator understands a specific set of OTel semantic conventions and maps them to MLflow span attributes.

Supported Translations

MLflow includes built-in translators for the following frameworks and conventions:

• OpenTelemetry GenAI Semantic Conventions
• OpenInference (Arize)
• Traceloop / OpenLLMetry
• Langfuse
• Vercel AI SDK
• Spring AI
• Google ADK
• LiveKit Agents
• VoltAgent

Translated Span Attributes

MLflow Attribute	Where It Appears in the UI
Span type (mlflow.spanType)	Determines the icon shown next to each span in the span tree (e.g., a model icon for LLM, a wrench icon for TOOL)
Inputs / Outputs (mlflow.spanInputs, mlflow.spanOutputs)	Displayed in the Inputs / Outputs tab of the span detail panel
Token usage (mlflow.chat.tokenUsage)	Shown in the trace drawer header and aggregated in the trace table's Tokens column
Model name (mlflow.llm.model)	Shown in the span detail panel as the model used for the LLM call
Model provider (mlflow.llm.provider)	Shown alongside the model name (e.g., openai, anthropic)
Message format (mlflow.message.format)	Controls how inputs/outputs are rendered in the chat UI view

Mapping Tables

OpenTelemetry GenAI Semantic Conventions

Based on the OTel GenAI semantic conventions.

Span Type (from gen_ai.operation.name):

Source Value	MLflow Span Type
chat	CHAT_MODEL
create_agent	AGENT
embeddings	EMBEDDING
execute_tool	TOOL
generate_content	LLM
invoke_agent	AGENT
text_completion	LLM
response	LLM

Inputs / Outputs:

Source Attributes	MLflow Attribute
gen_ai.input.messages, gen_ai.tool.call.arguments	mlflow.spanInputs
gen_ai.output.messages, gen_ai.tool.call.result	mlflow.spanOutputs

Inputs and outputs can also be extracted from span events:

• Input events: gen_ai.system.message, gen_ai.user.message, gen_ai.assistant.message
• Output events: gen_ai.choice

Token Usage:

Source Attribute	MLflow Token Field
gen_ai.usage.input_tokens	input_tokens
gen_ai.usage.output_tokens	output_tokens
(calculated from input + output)	total_tokens

Model:

Source Attribute	MLflow Attribute
gen_ai.response.model	mlflow.llm.model
gen_ai.request.model	mlflow.llm.model
gen_ai.provider.name	mlflow.llm.provider

---

OpenInference (Arize)

Based on the OpenInference semantic conventions.

Span Type (from openinference.span.kind): Mapped directly to the matching MLflow span type (e.g., LLM → LLM, TOOL → TOOL, AGENT → AGENT, etc.).

Inputs / Outputs:

Source Attributes	MLflow Attribute
input.value	mlflow.spanInputs
output.value	mlflow.spanOutputs

Token Usage:

Source Attribute	MLflow Token Field
llm.token_count.prompt	input_tokens
llm.token_count.completion	output_tokens
llm.token_count.total	total_tokens

Model:

Source Attribute	MLflow Attribute
llm.model_name	mlflow.llm.model
embedding.model_name	mlflow.llm.model
llm.provider	mlflow.llm.provider

---

Traceloop / OpenLLMetry

Based on OpenLLMetry semantic conventions.

Span Type (from traceloop.span.kind):

Source Value	MLflow Span Type
workflow	WORKFLOW
task	TASK
agent	AGENT
tool	TOOL
unknown	UNKNOWN

Inputs / Outputs:

Source Attributes	MLflow Attribute
traceloop.entity.input, gen_ai.prompt.*.content (pattern), gen_ai.completion.*.tool_calls.*.arguments (pattern)	mlflow.spanInputs
traceloop.entity.output, gen_ai.completion.*.content (pattern)	mlflow.spanOutputs

Token Usage:

Source Attribute	MLflow Token Field
gen_ai.usage.prompt_tokens	input_tokens
gen_ai.usage.completion_tokens	output_tokens
llm.usage.total_tokens	total_tokens

Model Provider:

Source Attribute	MLflow Attribute
gen_ai.system	mlflow.llm.provider

---

Langfuse

Maps Langfuse observation attributes to MLflow.

Span Type (from langfuse.observation.type):

Source Value	MLflow Span Type
generation	LLM
embedding	EMBEDDING
tool	TOOL
retriever	RETRIEVER
agent	AGENT
chain	CHAIN
evaluator	EVALUATOR
guardrail	GUARDRAIL
span	UNKNOWN

Inputs / Outputs:

Source Attributes	MLflow Attribute
langfuse.observation.input	mlflow.spanInputs
langfuse.observation.output	mlflow.spanOutputs

---

Vercel AI SDK

Based on Vercel AI SDK telemetry.

Span Type (from ai.operationId):

Source Value	MLflow Span Type
ai.generateText	LLM
ai.generateText.doGenerate	LLM
ai.toolCall	TOOL
ai.streamText	LLM
ai.streamText.doStream	LLM
ai.generateObject	LLM
ai.generateObject.doGenerate	LLM
ai.streamObject	LLM
ai.streamObject.doStream	LLM
ai.embed	EMBEDDING
ai.embed.doEmbed	EMBEDDING
ai.embedMany	EMBEDDING

Inputs / Outputs:

Source Attributes	MLflow Attribute
ai.prompt, ai.toolCall.args, ai.value / ai.values	mlflow.spanInputs
ai.response.text, ai.toolCall.result, ai.response.object, ai.embedding / ai.embeddings	mlflow.spanOutputs

For chat spans (ai.generateText.doGenerate, ai.streamText.doStream), MLflow unpacks all ai.prompt.* attributes as inputs and all ai.response.* attributes as outputs.

---

Spring AI

Based on Spring AI using OTel GenAI semantic conventions.

Span Type (from gen_ai.operation.name):

Source Value	MLflow Span Type
chat	CHAT_MODEL
embeddings	EMBEDDING

Inputs / Outputs:

Spring AI stores inputs and outputs in span events rather than attributes:

Event Name	Event Attribute	MLflow Attribute
gen_ai.content.prompt	gen_ai.prompt	mlflow.spanInputs
gen_ai.content.completion	gen_ai.completion	mlflow.spanOutputs

Token Usage:

Source Attribute	MLflow Token Field
gen_ai.usage.input_tokens	input_tokens
gen_ai.usage.output_tokens	output_tokens
(calculated from input + output)	total_tokens

---

Google ADK

Based on Google Agent Development Kit telemetry.

Google ADK uses OTel GenAI semantic conventions for span types and token usage. The following additional attributes are mapped for inputs and outputs:

Inputs / Outputs:

Source Attributes	MLflow Attribute
gcp.vertex.agent.llm_request, gcp.vertex.agent.tool_call_args	mlflow.spanInputs
gcp.vertex.agent.llm_response, gcp.vertex.agent.tool_response	mlflow.spanOutputs

---

LiveKit Agents

Based on LiveKit Agents telemetry. Inherits GenAI semantic conventions for token usage and event-based message extraction.

Span Type:

LiveKit uses a combination of heuristics to determine span type:

Condition	MLflow Span Type
gen_ai.request.model present	LLM
gen_ai.operation.name matches GenAI mapping	(see GenAI table above)
lk.retry_count present	LLM
lk.function_tool* prefix	TOOL
lk.agent_name, lk.instructions, lk.generation_id	AGENT
lk.tts*, lk.input_text	UNKNOWN
lk.user_transcript, lk.transcript_confidence, lk.transcription_delay	UNKNOWN

Inputs / Outputs:

Source Attributes	MLflow Attribute
lk.user_input, lk.user_transcript, lk.chat_ctx, lk.input_text	mlflow.spanInputs
lk.response.text, lk.response.function_calls	mlflow.spanOutputs

Also supports GenAI event-based input/output extraction (see GenAI table above).

Token Usage: Same as GenAI semantic conventions (gen_ai.usage.input_tokens, gen_ai.usage.output_tokens).

---

VoltAgent

Based on VoltAgent telemetry.

Span Type (from span.type, with fallback to entity.type):

Source Value	MLflow Span Type
agent	AGENT
llm	LLM
tool	TOOL
memory	MEMORY

note

span.type is checked first because child spans inherit entity.type from their parent agent, which would incorrectly classify LLM/tool spans as AGENT.

Inputs / Outputs:

Source Attributes	MLflow Attribute
agent.messages, llm.messages, input	mlflow.spanInputs
output	mlflow.spanOutputs

Token Usage:

Source Attribute	MLflow Token Field
usage.prompt_tokens or llm.usage.prompt_tokens	input_tokens
usage.completion_tokens or llm.usage.completion_tokens	output_tokens
usage.total_tokens or llm.usage.total_tokens	total_tokens

Session and User Mapping

In addition to span-level attributes, MLflow also maps the following OTel span attributes to trace-level metadata:

Span Attribute	Trace Metadata Key	Description
session.id	mlflow.trace.session	Session or conversation identifier
user.id	mlflow.trace.user	Application end-user identifier

These attributes follow the OTel semantic conventions for sessions and users.

For more details, see Track Users and Sessions.