Semantic conventions for generative client AI spans
Status: Development
[!Warning]
Existing GenAI instrumentations that are using v1.36.0 of this document (or prior):
- SHOULD NOT change the version of the GenAI conventions that they emit by default. Conventions include, but are not limited to, attributes, metric, span and event names, span kind and unit of measure.
- SHOULD introduce an environment variable
OTEL_SEMCONV_STABILITY_OPT_INas a comma-separated list of category-specific values. The list of values includes:
gen_ai_latest_experimental- emit the latest experimental version of GenAI conventions (supported by the instrumentation) and do not emit the old one (v1.36.0 or prior).- The default behavior is to continue emitting whatever version of the GenAI conventions the instrumentation was emitting (1.36.0 or prior).
This transition plan will be updated to include stable version before the GenAI conventions are marked as stable.
Spans
Inference
Status: 
This span represents a client call to Generative AI model or service that generates a response or requests a tool call based on the input prompt.
Span name SHOULD be {gen_ai.operation.name} {gen_ai.request.model}.
Semantic conventions for individual GenAI systems and frameworks MAY specify different span name format
and MUST follow the overall guidelines for span names.
Span kind SHOULD be CLIENTand MAY be set to INTERNAL on spans representing
call to models running in the same process. It’s RECOMMENDED to use CLIENT kind
when the GenAI system being instrumented usually runs in a different process than its
client or when the GenAI call happens over instrumented protocol such as HTTP.
Span status SHOULD follow the Recording Errors document.
| Attribute | Type | Description | Examples | Requirement Level | Stability |
|---|---|---|---|---|---|
gen_ai.operation.name | string | The name of the operation being performed. [1] | chat; generate_content; text_completion | Required |  |
gen_ai.provider.name | string | The Generative AI provider as identified by the client or server instrumentation. [2] | openai; gcp.gen_ai; gcp.vertex_ai | Required | |
error.type | string | Describes a class of error the operation ended with. [3] | timeout; java.net.UnknownHostException; server_certificate_invalid; 500 | Conditionally Required if the operation ended in an error |  |
gen_ai.conversation.id | string | The unique identifier for a conversation (session, thread), used to store and correlate messages within this conversation. [4] | conv_5j66UpCpwteGg4YSxUnt7lPY | Conditionally Required when available | |
gen_ai.output.type | string | Represents the content type requested by the client. [5] | text; json; image | Conditionally Required [6] | |
gen_ai.request.choice.count | int | The target number of candidate completions to return. | 3 | Conditionally Required if available, in the request, and !=1 | |
gen_ai.request.model | string | The name of the GenAI model a request is being made to. [7] | gpt-4 | Conditionally Required If available. | |
gen_ai.request.seed | int | Requests with same seed value more likely to return same result. | 100 | Conditionally Required if applicable and if the request includes a seed | |
server.port | int | GenAI server port. [8] | 80; 8080; 443 | Conditionally Required If server.address is set. |  |
gen_ai.request.frequency_penalty | double | The frequency penalty setting for the GenAI request. | 0.1 | Recommended | |
gen_ai.request.max_tokens | int | The maximum number of tokens the model generates for a request. | 100 | Recommended | |
gen_ai.request.presence_penalty | double | The presence penalty setting for the GenAI request. | 0.1 | Recommended | |
gen_ai.request.stop_sequences | string[] | List of sequences that the model will use to stop generating further tokens. | ["forest", "lived"] | Recommended | |
gen_ai.request.temperature | double | The temperature setting for the GenAI request. | 0.0 | Recommended | |
gen_ai.request.top_k | double | The top_k sampling setting for the GenAI request. | 1.0 | Recommended | |
gen_ai.request.top_p | double | The top_p sampling setting for the GenAI request. | 1.0 | Recommended | |
gen_ai.response.finish_reasons | string[] | Array of reasons the model stopped generating tokens, corresponding to each generation received. | ["stop"]; ["stop", "length"] | Recommended | |
gen_ai.response.id | string | The unique identifier for the completion. | chatcmpl-123 | Recommended | |
gen_ai.response.model | string | The name of the model that generated the response. [9] | gpt-4-0613 | Recommended | |
gen_ai.usage.input_tokens | int | The number of tokens used in the GenAI input (prompt). | 100 | Recommended | |
gen_ai.usage.output_tokens | int | The number of tokens used in the GenAI response (completion). | 180 | Recommended | |
server.address | string | GenAI server address. [10] | example.com; 10.1.2.80; /tmp/my.sock | Recommended | |
gen_ai.input.messages | any | The chat history provided to the model as an input. [11] | [ { “role”: “user”, “parts”: [ { “type”: “text”, “content”: “Weather in Paris?" } ] }, { “role”: “assistant”, “parts”: [ { “type”: “tool_call”, “id”: “call_VSPygqKTWdrhaFErNvMV18Yl”, “name”: “get_weather”, “arguments”: { “location”: “Paris” } } ] }, { “role”: “tool”, “parts”: [ { “type”: “tool_call_response”, “id”: " call_VSPygqKTWdrhaFErNvMV18Yl”, “result”: “rainy, 57°F” } ] } ] | Opt-In | |
gen_ai.output.messages | any | Messages returned by the model where each message represents a specific model response (choice, candidate). [12] | [ { “role”: “assistant”, “parts”: [ { “type”: “text”, “content”: “The weather in Paris is currently rainy with a temperature of 57°F." } ], “finish_reason”: “stop” } ] | Opt-In | |
gen_ai.system_instructions | any | The system message or instructions provided to the GenAI model separately from the chat history. [13] | [ { “type”: “text”, “content”: “You are an Agent that greet users, always use greetings tool to respond” } ]; [ { “type”: “text”, “content”: “You are a language translator." }, { “type”: “text”, “content”: “Your mission is to translate text in English to French." } ] | Opt-In | |
[1] gen_ai.operation.name: If one of the predefined values applies, but specific system uses a different name it’s RECOMMENDED to document it in the semantic conventions for specific GenAI system and use system-specific name in the instrumentation. If a different name is not documented, instrumentation libraries SHOULD use applicable predefined value.
[2] gen_ai.provider.name: The attribute SHOULD be set based on the instrumentation’s best
knowledge and may differ from the actual model provider.
Multiple providers, including Azure OpenAI, Gemini, and AI hosting platforms are accessible using the OpenAI REST API and corresponding client libraries, but may proxy or host models from different providers.
The gen_ai.request.model, gen_ai.response.model, and server.address
attributes may help identify the actual system in use.
The gen_ai.provider.name attribute acts as a discriminator that
identifies the GenAI telemetry format flavor specific to that provider
within GenAI semantic conventions.
It SHOULD be set consistently with provider-specific attributes and signals.
For example, GenAI spans, metrics, and events related to AWS Bedrock
should have the gen_ai.provider.name set to aws.bedrock and include
applicable aws.bedrock.* attributes and are not expected to include
openai.* attributes.
[3] error.type: The error.type SHOULD match the error code returned by the Generative AI provider or the client library,
the canonical name of exception that occurred, or another low-cardinality error identifier.
Instrumentations SHOULD document the list of errors they report.
[4] gen_ai.conversation.id: Instrumentations SHOULD populate conversation id when they have it readily available
for a given operation, for example:
when client framework being instrumented manages conversation history (see LlamaIndex chat store)
when instrumenting GenAI client libraries that maintain conversation on the backend side (see AWS Bedrock agent sessions, OpenAI Assistant threads)
Application developers that manage conversation history MAY add conversation id to GenAI and other spans or logs using custom span or log record processors or hooks provided by instrumentation libraries.
[5] gen_ai.output.type: This attribute SHOULD be used when the client requests output of a specific type. The model may return zero or more outputs of this type.
This attribute specifies the output modality and not the actual output format. For example, if an image is requested, the actual output could be a URL pointing to an image file.
Additional output format details may be recorded in the future in the gen_ai.output.{type}.* attributes.
[6] gen_ai.output.type: when applicable and if the request includes an output format.
[7] gen_ai.request.model: The name of the GenAI model a request is being made to. If the model is supplied by a vendor, then the value must be the exact name of the model requested. If the model is a fine-tuned custom model, the value should have a more specific name than the base model that’s been fine-tuned.
[8] server.port: When observed from the client side, and when communicating through an intermediary, server.port SHOULD represent the server port behind any intermediaries, for example proxies, if it’s available.
[9] gen_ai.response.model: If available. The name of the GenAI model that provided the response. If the model is supplied by a vendor, then the value must be the exact name of the model actually used. If the model is a fine-tuned custom model, the value should have a more specific name than the base model that’s been fine-tuned.
[10] server.address: When observed from the client side, and when communicating through an intermediary, server.address SHOULD represent the server address behind any intermediaries, for example proxies, if it’s available.
[11] gen_ai.input.messages: Instrumentations MUST follow Input messages JSON schema.
When the attribute is recorded on events, it MUST be recorded in structured
form. When recorded on spans, it MAY be recorded as a JSON string if structured
format is not supported and SHOULD be recorded in structured form otherwise.
Messages MUST be provided in the order they were sent to the model. Instrumentations MAY provide a way for users to filter or truncate input messages.
[!Warning] This attribute is likely to contain sensitive information including user/PII data.
See Recording content on attributes section for more details.
[12] gen_ai.output.messages: Instrumentations MUST follow Output messages JSON schema
Each message represents a single output choice/candidate generated by the model. Each message corresponds to exactly one generation (choice/candidate) and vice versa - one choice cannot be split across multiple messages or one message cannot contain parts from multiple choices.
When the attribute is recorded on events, it MUST be recorded in structured form. When recorded on spans, it MAY be recorded as a JSON string if structured format is not supported and SHOULD be recorded in structured form otherwise.
Instrumentations MAY provide a way for users to filter or truncate output messages.
[!Warning] This attribute is likely to contain sensitive information including user/PII data.
See Recording content on attributes section for more details.
[13] gen_ai.system_instructions: This attribute SHOULD be used when the corresponding provider or API
allows to provide system instructions or messages separately from the
chat history.
Instructions that are part of the chat history SHOULD be recorded in
gen_ai.input.messages attribute instead.
Instrumentations MUST follow System instructions JSON schema.
When recorded on spans, it MAY be recorded as a JSON string if structured format is not supported and SHOULD be recorded in structured form otherwise.
Instrumentations MAY provide a way for users to filter or truncate system instructions.
[!Warning] This attribute may contain sensitive information.
See Recording content on attributes section for more details.
error.type has the following list of well-known values. If one of them applies, then the respective value MUST be used; otherwise, a custom value MAY be used.
| Value | Description | Stability |
|---|---|---|
_OTHER | A fallback error value to be used when the instrumentation doesn’t define a custom value. | |
gen_ai.operation.name has the following list of well-known values. If one of them applies, then the respective value MUST be used; otherwise, a custom value MAY be used.
| Value | Description | Stability |
|---|---|---|
chat | Chat completion operation such as OpenAI Chat API | |
create_agent | Create GenAI agent | |
embeddings | Embeddings operation such as OpenAI Create embeddings API | |
execute_tool | Execute a tool | |
generate_content | Multimodal content generation operation such as Gemini Generate Content | |
invoke_agent | Invoke GenAI agent | |
text_completion | Text completions operation such as OpenAI Completions API (Legacy) | |
gen_ai.output.type has the following list of well-known values. If one of them applies, then the respective value MUST be used; otherwise, a custom value MAY be used.
| Value | Description | Stability |
|---|---|---|
image | Image | |
json | JSON object with known or unknown schema | |
speech | Speech | |
text | Plain text | |
gen_ai.provider.name has the following list of well-known values. If one of them applies, then the respective value MUST be used; otherwise, a custom value MAY be used.
| Value | Description | Stability |
|---|---|---|
anthropic | Anthropic | |
aws.bedrock | AWS Bedrock | |
azure.ai.inference | Azure AI Inference | |
azure.ai.openai | Azure OpenAI | |
cohere | Cohere | |
deepseek | DeepSeek | |
gcp.gemini | Gemini [14] | |
gcp.gen_ai | Any Google generative AI endpoint [15] | |
gcp.vertex_ai | Vertex AI [16] | |
groq | Groq | |
ibm.watsonx.ai | IBM Watsonx AI | |
mistral_ai | Mistral AI | |
openai | OpenAI | |
perplexity | Perplexity | |
x_ai | xAI | |
[14]: Used when accessing the ‘generativelanguage.googleapis.com’ endpoint. Also known as the AI Studio API.
[15]: May be used when specific backend is unknown.
[16]: Used when accessing the ‘aiplatform.googleapis.com’ endpoint.
Embeddings
Status:
Describes GenAI embeddings span - a request to a Generative AI model or service that generates an embeddings based on the input.
The gen_ai.operation.name SHOULD be embeddings.
Span name SHOULD be {gen_ai.operation.name} {gen_ai.request.model}.
Span kind SHOULD be CLIENT.
Span status SHOULD follow the Recording Errors document.
| Attribute | Type | Description | Examples | Requirement Level | Stability |
|---|---|---|---|---|---|
gen_ai.operation.name | string | The name of the operation being performed. [1] | chat; generate_content; text_completion | Required | |
error.type | string | Describes a class of error the operation ended with. [2] | timeout; java.net.UnknownHostException; server_certificate_invalid; 500 | Conditionally Required if the operation ended in an error | |
gen_ai.request.model | string | The name of the GenAI model a request is being made to. [3] | gpt-4 | Conditionally Required If available. | |
server.port | int | GenAI server port. [4] | 80; 8080; 443 | Conditionally Required If server.address is set. | |
gen_ai.request.encoding_formats | string[] | The encoding formats requested in an embeddings operation, if specified. [5] | ["base64"]; ["float", "binary"] | Recommended | |
gen_ai.usage.input_tokens | int | The number of tokens used in the GenAI input (prompt). | 100 | Recommended | |
server.address | string | GenAI server address. [6] | example.com; 10.1.2.80; /tmp/my.sock | Recommended | |
[1] gen_ai.operation.name: If one of the predefined values applies, but specific system uses a different name it’s RECOMMENDED to document it in the semantic conventions for specific GenAI system and use system-specific name in the instrumentation. If a different name is not documented, instrumentation libraries SHOULD use applicable predefined value.
[2] error.type: The error.type SHOULD match the error code returned by the Generative AI provider or the client library,
the canonical name of exception that occurred, or another low-cardinality error identifier.
Instrumentations SHOULD document the list of errors they report.
[3] gen_ai.request.model: The name of the GenAI model a request is being made to. If the model is supplied by a vendor, then the value must be the exact name of the model requested. If the model is a fine-tuned custom model, the value should have a more specific name than the base model that’s been fine-tuned.
[4] server.port: When observed from the client side, and when communicating through an intermediary, server.port SHOULD represent the server port behind any intermediaries, for example proxies, if it’s available.
[5] gen_ai.request.encoding_formats: In some GenAI systems the encoding formats are called embedding types. Also, some GenAI systems only accept a single format per request.
[6] server.address: When observed from the client side, and when communicating through an intermediary, server.address SHOULD represent the server address behind any intermediaries, for example proxies, if it’s available.
error.type has the following list of well-known values. If one of them applies, then the respective value MUST be used; otherwise, a custom value MAY be used.
| Value | Description | Stability |
|---|---|---|
_OTHER | A fallback error value to be used when the instrumentation doesn’t define a custom value. | |
gen_ai.operation.name has the following list of well-known values. If one of them applies, then the respective value MUST be used; otherwise, a custom value MAY be used.
| Value | Description | Stability |
|---|---|---|
chat | Chat completion operation such as OpenAI Chat API | |
create_agent | Create GenAI agent | |
embeddings | Embeddings operation such as OpenAI Create embeddings API | |
execute_tool | Execute a tool | |
generate_content | Multimodal content generation operation such as Gemini Generate Content | |
invoke_agent | Invoke GenAI agent | |
text_completion | Text completions operation such as OpenAI Completions API (Legacy) | |
Execute tool span
Status:
Describes tool execution span.
gen_ai.operation.name SHOULD be execute_tool.
Span name SHOULD be execute_tool {gen_ai.tool.name}.
GenAI instrumentations that are able to instrument tool execution call SHOULD do so. However, it’s common for tools to be executed by the application code. It’s recommended for the application developers to follow this semantic convention for tools invoked by the application code.
Span kind SHOULD be INTERNAL.
Span status SHOULD follow the Recording Errors document.
| Attribute | Type | Description | Examples | Requirement Level | Stability |
|---|---|---|---|---|---|
gen_ai.operation.name | string | The name of the operation being performed. [1] | chat; generate_content; text_completion | Required | |
error.type | string | Describes a class of error the operation ended with. [2] | timeout; java.net.UnknownHostException; server_certificate_invalid; 500 | Conditionally Required if the operation ended in an error | |
gen_ai.tool.call.id | string | The tool call identifier. | call_mszuSIzqtI65i1wAUOE8w5H4 | Recommended if available | |
gen_ai.tool.description | string | The tool description. | Multiply two numbers | Recommended if available | |
gen_ai.tool.name | string | Name of the tool utilized by the agent. | Flights | Recommended | |
gen_ai.tool.type | string | Type of the tool utilized by the agent [3] | function; extension; datastore | Recommended if available | |
[1] gen_ai.operation.name: If one of the predefined values applies, but specific system uses a different name it’s RECOMMENDED to document it in the semantic conventions for specific GenAI system and use system-specific name in the instrumentation. If a different name is not documented, instrumentation libraries SHOULD use applicable predefined value.
[2] error.type: The error.type SHOULD match the error code returned by the Generative AI provider or the client library,
the canonical name of exception that occurred, or another low-cardinality error identifier.
Instrumentations SHOULD document the list of errors they report.
[3] gen_ai.tool.type: Extension: A tool executed on the agent-side to directly call external APIs, bridging the gap between the agent and real-world systems.
Agent-side operations involve actions that are performed by the agent on the server or within the agent’s controlled environment.
Function: A tool executed on the client-side, where the agent generates parameters for a predefined function, and the client executes the logic.
Client-side operations are actions taken on the user’s end or within the client application.
Datastore: A tool used by the agent to access and query structured or unstructured external data for retrieval-augmented tasks or knowledge updates.
error.type has the following list of well-known values. If one of them applies, then the respective value MUST be used; otherwise, a custom value MAY be used.
| Value | Description | Stability |
|---|---|---|
_OTHER | A fallback error value to be used when the instrumentation doesn’t define a custom value. | |
gen_ai.operation.name has the following list of well-known values. If one of them applies, then the respective value MUST be used; otherwise, a custom value MAY be used.
| Value | Description | Stability |
|---|---|---|
chat | Chat completion operation such as OpenAI Chat API | |
create_agent | Create GenAI agent | |
embeddings | Embeddings operation such as OpenAI Create embeddings API | |
execute_tool | Execute a tool | |
generate_content | Multimodal content generation operation such as Gemini Generate Content | |
invoke_agent | Invoke GenAI agent | |
text_completion | Text completions operation such as OpenAI Completions API (Legacy) | |
Capturing instructions, inputs, and outputs
Full (buffered) content
Model instructions, user messages, and model outputs are considered sensitive and are often large in size.
Recording large or sensitive content in telemetry may be problematic due to high storage costs, regulatory requirements, or the need to enforce different access models for operational and user data.
OpenTelemetry instrumentations SHOULD NOT capture them by default, but SHOULD provide an option for users to opt in.
Application developers should choose an appropriate usage pattern based on application needs and maturity:
[Default] Don’t record instructions, inputs, or outputs.
Record instructions, inputs, and outputs on the GenAI spans using corresponding attributes (
gen_ai.system_instructions,gen_ai.input.messages,gen_ai.output.messages).This approach is best suited for situations where telemetry volume is manageable and either privacy regulations do not apply or the telemetry storage complies with them, for example, in pre-production environments.
See Recording content on attributes section for more details.
Store content externally and record references on the spans.
This pattern is recommended in production environments where telemetry volume is a concern or sensitive data needs to be handled securely. Using external storage enables separate access controls.
See Uploading content to external storage section for more details.
Recording content on attributes
The content captured in gen_ai.system_instructions, gen_ai.input.messages,
and gen_ai.output.messages attributes is likely to be large.
It may contain media, and even in the text form, it may be larger than observability backend limits for telemetry envelopes or attribute values.
The inputs and outputs attributes follow common structure formally defined in inputs JSON schema and outputs JSON schema. See also their representation in Python code.
[!NOTE]
Recording structured attributes is supported on events (or logs) and may not yet be supported on spans. See OTEP: Extending attributes to support complex values for the details.
If structured attributes are not yet supported on spans in a given language, the corresponding attribute value SHOULD be serialized to JSON string on spans and recorded in its structured form on events.
Instrumentation MAY provide a configuration option allowing to truncate properties such as individual message contents, preserving JSON structure.
Uploading content to external storage
Instrumentations MAY support user-defined in-process hooks to handle content upload.
The hook SHOULD operate independently of the opt-in flags that control capturing of
gen_ai.system_instructions, gen_ai.input.messages, and gen_ai.output.messages.
If such a hook is supported and configured, instrumentations SHOULD invoke it regardless of the span sampling decision with:
- the instructions, inputs, and outputs object using formats defined in this convention and before they are serialized to JSON string;
- the span instance
The hook implementation SHOULD be able to enrich and modify provided span, instructions, and message objects.
If instrumentation is configured to also record gen_ai.system_instructions,
gen_ai.input.messages, and gen_ai.output.messages attributes, it SHOULD do it
after calling the hook and SHOULD record values that were potentially modified within
the hook implementation.
The hook API SHOULD be generic. The application or distro is responsible for the hook implementation including
- the uploading process either in synchronous or asynchronous way,
- recording references to the uploaded content on the span,
- handling content in a different way.
Application or OpenTelemetry distributions MAY also implement content uploading
in the telemetry processing pipeline (in-process or via a collector), based on the
gen_ai.system_instructions, gen_ai.input.messages, and gen_ai.output.messages
attributes. Given the potential data volume, it is RECOMMENDED to tune batching
and export settings accordingly in the OpenTelemetry SDK pipeline.
TODO: document a common approach to record references to externally stored content.
Check out LLM call examples.
Streaming chunks
TODO
Feedback
Was this page helpful?
Thank you. Your feedback is appreciated!
Please let us know how we can improve this page. Your feedback is appreciated!