OpenInference Span Kinds

The twelve OpenInference span kinds — LLM, Tool, Agent, Chain, Retriever, and the rest — with the canonical attributes Phoenix expects on each.

A span kind is the category of operation a span represents. OpenTelemetry has its own span kinds for network calls (SERVER, CLIENT, PRODUCER, …). OpenInference adds an AI-specific set: LLM, Tool, Agent, Chain, Retriever, and more.

Span kind is what drives the span-kind icons in the Phoenix UI. It’s also how Phoenix knows which attributes to expect — an LLM span gets a different visual treatment from a tool span.

Setting the Span Kind

Span kind is set via the openinference.span.kind attribute. The value is ALL CAPS — "TOOL", not "Tool". Casing matters: the canonical OpenInference Python enum is OpenInferenceSpanKindValues.TOOL = "TOOL".

from openinference.semconv.trace import (
    SpanAttributes,
    OpenInferenceSpanKindValues,
)

span.set_attribute(
    SpanAttributes.OPENINFERENCE_SPAN_KIND,
    OpenInferenceSpanKindValues.TOOL.value,
)

Phoenix UI behavior: span-kind icons in Phoenix trace tree come directly from the openinference.span.kind attribute. A span without that attribute renders with a blank icon. Auto-instrumentors set this for you — set it explicitly on every manual span.

The Twelve Span Kinds

OpenInference defines twelve span-kind values. The first four (LLM, Chain, Agent, Tool) are the ones you’ll encounter most often:

Icon	Span kind	Description
	LLM	A call to a large language model.
	CHAIN	A starting point or a link between different LLM application steps.
	AGENT	A span that encompasses calls to LLMs and tools — typically the top-level wrapper for an agent loop.
	TOOL	A span that represents a call to an external tool, API, or function on behalf of an LLM.
	RETRIEVER	A data retrieval query for context from a datastore (e.g., a vector DB query).
	EMBEDDING	An encoding of unstructured data into a vector.
	RERANKER	A relevance-based re-ordering of documents.
	GUARDRAIL	A validation of LLM input or output for safety, policy, or compliance.
	EVALUATOR	An evaluation process — the type, configuration, and results.
	DECISION	A call to a decision model — one that scores or selects among candidate options supplied in the request (routing, classification, rubric scoring) rather than generating free-form text. See DECISION.
	PROMPT	A span representing prompt construction or templating.
	UNKNOWN	Default when no kind is set explicitly.

DECISION spans are new. Phoenix renders them with their own icon and a span view that names the decision model. The OpenInference instrumentors for the OpenAI Decisions API and TypeSafe AI System One emit them. The decision.* attributes are described under DECISION below and may still evolve as more decision-model providers appear.

Common Attributes Across All Kinds

These attributes apply to every span kind. Auto-instrumentors set them automatically; set them yourself on every manual span:

Attribute	Purpose
openinference.span.kind	The span kind (ALL CAPS — e.g., "LLM").
input.value	The span’s input. Free-form string or serialized JSON.
input.mime_type	"text/plain" (default) or "application/json". The TS semconv also defines "audio/wav".
output.value	The span’s output.
output.mime_type	"text/plain" (default) or "application/json". The TS semconv also defines "audio/wav".
metadata	Arbitrary JSON-stringified metadata.
session.id	Groups traces into a session. See Sessions.
user.id	The user the request belongs to.
tag.tags	A list of tags for filtering and search.

Phoenix UI behavior: Phoenix reads the trace-level input and output from the root span’s input.value and output.value. If your root span doesn’t set them, the trace list view shows those columns as blank even if child spans have data.

Per-Kind Attributes

The four most-used kinds, plus the newer DECISION kind, have a richer attribute set. The full list lives in the OpenInference semantic conventions — the tables below are a summary.

LLM

A call to a large language model.

Attribute	Description
llm.model_name	The model identifier (e.g., gpt-4o, claude-3-5-sonnet-20241022).
llm.provider	The provider (e.g. openai, anthropic, mistralai).
llm.system	The underlying LLM system, when distinct from provider.
llm.input_messages	The chat history sent to the model.
llm.output_messages	The model’s response messages.
llm.invocation_parameters	Temperature, top-p, max tokens, etc. (JSON-stringified).
llm.tools	Tool definitions exposed to the model.
llm.function_call	Function call returned by the model (legacy).
llm.prompts	The raw prompts sent (for completion-style APIs).
llm.choices	The choices returned by the model.
llm.prompt_template.template	The prompt template before variable substitution.
llm.prompt_template.variables	The variables substituted into the template.
llm.prompt_template.version	A version identifier for the template.
llm.token_count.total	Total tokens.
llm.token_count.prompt	Prompt tokens.
llm.token_count.completion	Completion tokens.
llm.token_count.prompt_details.*	Breakdown of prompt tokens (cached, audio, etc.).
llm.token_count.completion_details.*	Breakdown of completion tokens.
llm.cost.total	Total cost.
llm.cost.prompt, llm.cost.completion	Per-direction cost.
llm.cost.*_details.*	Breakdown of cost by token type.
llm.finish_reason	Why the model stopped generating (stop, length, tool_calls, …).

TOOL

A call to an external tool, API, or function on behalf of an LLM.

Attribute	Description
tool.id	A unique identifier for the tool invocation (often correlated with the LLM’s tool-call ID).
tool.name	The tool’s name.
tool.description	The tool’s description (often matches what was shown to the model).
tool.parameters	The JSON-serialized parameters passed to the tool.
tool.json_schema	The tool’s full JSON schema as exposed to the model.

Tool calls are one of the clearest cases where you must instrument manually when making an LLM calls directly, instead of via a framework. When you call OpenAI directly, the auto-instrumentor traces the model’s request for a tool call — but the actual execution of the tool function happens in your Python code, where the instrumentor can’t see it. Wrap the function in a tool span yourself.

AGENT

A span that encompasses calls to LLMs and tools — typically the top-level wrapper for an agent loop.

Attribute	Description
agent.name	The agent’s name.
graph.node.id	Node ID for graph-based agent frameworks (e.g., LangGraph).
graph.node.name	Node display name.
graph.node.parent_id	Parent node ID, for graph traversal visualization.

CHAIN

A starting point or a link between different application steps. Use a chain span to group pre/post-processing logic that doesn’t fit any of the more specific kinds.

Chain spans use only the common attributes — input.value, output.value, metadata, session.id, etc.

DECISION

A call to a decision model. A decision model takes some state plus a fixed set of typed questions and returns one typed, probabilistic answer per question, instead of generating free-form text. Examples are the OpenAI Decisions API, which answers predicate, choice, and score questions over text or images, and TypeSafe AI’s System One / Jev, which answers Noul, Choice, and Score questions. Decision models suit the “smart if-statements” inside an application: routing a request to the right agent or department, classifying content, judging whether a condition holds, or scoring an item against a rubric.

Decision spans use a decision.* namespace instead of llm.*, so decision-model usage stays out of LLM model, token, and cost reporting. The decision.* identification attributes mirror their llm.* counterparts and share the same well-known values (openai, typesafe, and so on).

Attribute	Description
decision.system	The decision API the call speaks, which fixes the request and response shape and the question types: openai for the OpenAI Decisions API, typesafe for System One. Required.
decision.provider	Who hosts the model that answered. Usually the same as the system; differs for a Decisions call through Azure OpenAI (azure) or a self-hosted deployment serving a TypeSafe-compatible API.
decision.model_name	The decision model that answered (e.g. gpt-6-luna, jev-1.13.0). Mirrors llm.model_name.
decision.request.model_name	The model requested by the caller, when it can differ from the model that served the response (e.g. a floating alias such as jev-latest).
decision.response.model_name	The model that actually produced the decision, as reported by the provider.
decision.token_count.input	Input tokens consumed: the state, the questions, and the candidate options.
decision.token_count.output	Output tokens produced. Usually a handful, since the answers are typed values rather than text, and some providers do not bill for them.
input.value / input.mime_type	The raw request as JSON, including the candidate options.
output.value / output.mime_type	The raw response as JSON, including the selection, probabilities, or scores.

Token counts have no prompt/completion split and no total, since a decision call has no chat-style prompt and completion. A decision call is not a chat exchange either, so decision spans carry no llm.input_messages or llm.output_messages, and they should not set llm.system, llm.provider, llm.model_name, or llm.token_count.*.

Phoenix UI behavior: a DECISION span renders with the decision icon in the trace tree. Its details view shows the input and output cards, with the decision model named in the input card header. The name comes from decision.model_name, then decision.response.model_name, then decision.request.model_name, and is led by the provider’s icon when Phoenix has one for decision.provider. Decision spans use decision.token_count.* rather than llm.token_count.*, so their usage does not appear in LLM token and cost totals.

The full convention, including the system-versus-provider rules and a worked example, is in the OpenInference decision spans specification.

Other Kinds

The remaining seven span kinds — RETRIEVER, EMBEDDING, RERANKER, GUARDRAIL, EVALUATOR, PROMPT, UNKNOWN — each have their own attribute schemas. Consult the canonical source for the full per-kind attribute list.

Multimodal Message Content

Inputs and outputs in modern AI applications aren’t always plain text. OpenInference defines a message_content attribute family for multimodal content:

Attribute	Description
message_content.type	One of text, image, audio, reasoning, tool_use.
message_content.text	The text payload (when type is text or reasoning).
message_content.image	The image payload — URL or inline base64 (when type=image).
message_content.id	Provider-assigned identifier for the message content item (e.g., OpenAI Responses reasoning IDs).
message_content.signature	Opaque vendor-issued signature (e.g., Gemini thoughtSignature).
message_content.data	Opaque vendor-issued data (e.g., Anthropic redacted_thinking.data).
message_content.encrypted_content	Encrypted payload for sensitive content (e.g., OpenAI encrypted_content).

For audio specifically, OpenInference exposes a separate set of attributes — audio.url, audio.mime_type, audio.transcript — see the canonical source.

How a Trace Tree Uses Span Kinds

A typical agent trace nests span kinds inside each other:

AGENT span                         (agent.name = "research-assistant")
├── LLM span                       "decide next step"
├── TOOL span                      "search knowledge base"
│   └── RETRIEVER span             "vector search"
│       └── EMBEDDING span         "embed query"
└── LLM span                       "answer with retrieved docs"

Each level has its own kind, its own attributes, and its own rendering in the Phoenix UI. The combination is what makes agent behavior legible.

---

Next step

You know what attributes go on a span. Next, the three ways to actually emit spans — auto, manual, and hybrid:

Next: Instrumentation Approaches