Span Kinds
The openinference.span.kind attribute is required on every OpenInference span. It names the role the operation
plays in an AI application, which tells a tracing backend how to assemble and present the trace. It is distinct from
OpenTelemetry’s own SpanKind; see Traces and Spans.
| Value | Represents | Conventions |
|---|---|---|
LLM |
A call to a large language model | LLM Spans |
EMBEDDING |
Generation of vector embeddings | Embedding Spans |
CHAIN |
A link between application steps, or the start of a request | |
RETRIEVER |
A query to a vector store, search engine, or database | |
RERANKER |
Reordering of candidate documents by relevance | |
TOOL |
Execution of a tool or function | Tool Calling |
AGENT |
A reasoning step that acts on tools using an LLM’s guidance | |
GUARDRAIL |
An input or output moderation check | |
EVALUATOR |
An evaluation of a model’s output | Annotations and Evaluations |
PROMPT |
The rendering of a prompt template | |
DECISION |
A call to a decision model | Decision Spans |
Each section below lists the attributes that characterize spans of that kind. Any span may also carry the
general attributes input.value and output.value and the
context attributes.
LLM
A call to a Large Language Model, for example a call to OpenAI or Llama for chat completions or text generation.
Characteristic attributes: llm.*, including llm.input_messages, llm.output_messages, llm.tools, and
llm.token_count.*. See LLM Spans.
EMBEDDING
A call to an LLM or embedding service to generate embeddings, for example a call to OpenAI for an ada embedding used in retrieval.
Characteristic attributes: embedding.*. See Embedding Spans.
CHAIN
A starting point or a link between different LLM application steps, for example the beginning of a request to an LLM application, or the glue code that passes context from a retriever to an LLM call.
Characteristic attributes: input.value, output.value.
RETRIEVER
A data retrieval step, for example a call to a vector store or a database to fetch documents or information.
Characteristic attributes: retrieval.documents, with each document’s document.* attributes. input.value usually
holds the query.
RERANKER
The reranking of a set of input documents. For example, a cross-encoder computes the input documents’ relevance scores with respect to a user query, and the top K documents with the highest scores are returned.
Characteristic attributes: reranker.query, reranker.model_name, reranker.top_k, reranker.input_documents,
reranker.output_documents.
TOOL
A call to an external tool such as a calculator or a weather API, or any function execution invoked by an LLM or agent.
Characteristic attributes: tool.name, tool.description, tool.parameters, tool.id. See
Tool Calling for how the LLM side of a tool call is recorded.
AGENT
A span that encompasses calls to LLMs and tools. An agent describes a reasoning block that acts on tools using the guidance of an LLM.
Characteristic attributes: agent.name, and graph.node.* when the agent is part of an execution graph.
GUARDRAIL
A call to a component that protects against jailbreak prompts or undesirable output, by modifying or rejecting an LLM’s input or response. For example, a guardrail span could check whether an LLM’s response contains inappropriate language, through a custom or external guardrail library, and amend the response to remove it.
Characteristic attributes: input.value, output.value.
EVALUATOR
A call to a function or process that evaluates a language model’s outputs, for example assessing the relevance, correctness, or helpfulness of its answers.
Characteristic attributes: annotations or evaluations. An EVALUATOR span is the recommended carrier for
post-hoc feedback.
PROMPT
The rendering of a prompt template, for example substituting variables into a template to produce the final prompt sent to an LLM.
Characteristic attributes: llm.prompt_template.*, and prompt.vendor, prompt.id, and prompt.url when the prompt
comes from a prompt management system.
DECISION
A call to a decision model, which scores or selects among candidate options provided in the request rather than generating free-form text. For example, a decision span could represent choosing a route, judging a condition, or scoring an item against a rubric. Decision models include TypeSafe AI’s Jev.
Characteristic attributes: decision.*. See Decision Spans.