AI Observability reads these attributes from your spans. A panel stays empty until your spans carry the attributes that it uses. Most names come from the OpenTelemetry GenAI semantic conventions.
If a span does not carry a GenAI attribute, SigNoz takes the value from the first default source key in the table that the span has. SigNoz keeps most source keys on the span. For gen_ai.input.messages and gen_ai.output.messages, SigNoz removes the source key after it writes the GenAI attribute. If your instrumentation uses a name that is not in the table, map it in Attribute Mapping.
Span attributes
In the Used by column, None means that no AI Observability panel or column reads the attribute.
LLM spans
The default gen_ai.llm mapping group runs on spans with an attribute key that contains model or llm.input_messages.
| Attribute | Holds | Used by | Default source keys |
|---|---|---|---|
gen_ai.request.model | The model that the call requested. | All LLM panels and the model filter. SigNoz counts a span with this attribute as an LLM call. | reranker.model_name, llm.model_name, llm.request.model, ai.model.id, langfuse.observation.model.name, embedding.model_name, model, gen_ai.response.model |
gen_ai.response.model | The model that served the response. | None | llm.response.model, ai.response.model |
gen_ai.provider.name | The model provider, for example openai. | Cost by provider and the provider filter. | gen_ai.system, llm.vendor, llm.provider, llm.system, ai.model.provider |
gen_ai.operation.name | The type of operation, for example chat or LLM. | The Gen Ai Operation Name filter of the Explorer. | openinference.span.kind, llm.request.type, traceloop.span.kind |
gen_ai.usage.input_tokens | The number of input tokens. | Token panels and cost. | input_tokens, gen_ai.usage.prompt_tokens, llm.usage.prompt_tokens, llm.token_count.prompt, ai.usage.inputTokens, ai.usage.promptTokens |
gen_ai.usage.output_tokens | The number of output tokens. | Token panels and cost. | output_tokens, gen_ai.usage.completion_tokens, llm.usage.completion_tokens, llm.token_count.completion, ai.usage.outputTokens, ai.usage.completionTokens |
gen_ai.usage.cache_read.input_tokens | Input tokens that the provider read from its prompt cache. | Cache token panels, cache cost, and Prompt cache hit ratio. | cache_read_tokens, gen_ai.usage.input_tokens.cached, gen_ai.usage.prompt_token_details.cached_read_tokens, gen_ai.usage.cache_read_input_tokens, llm.token_count.prompt_details.cache_read, ai.usage.cachedInputTokens |
gen_ai.usage.cache_creation.input_tokens | Input tokens that the provider wrote to its prompt cache. | Cache token panels and cache cost. | cache_creation_tokens, gen_ai.usage.cache_write.input_tokens, gen_ai.usage.cache_creation_input_tokens, llm.token_count.prompt_details.cache_write |
gen_ai.usage.reasoning.output_tokens | Output tokens that the model used for reasoning. These tokens are part of gen_ai.usage.output_tokens. | The Usage Breakdown of the trace in Trace Details. | gen_ai.usage.reasoning_tokens, gen_ai.usage.completion_token_details.reasoning_tokens, llm.token_count.completion_details.reasoning, ai.usage.reasoningTokens |
gen_ai.server.ttft | The time to the first token of a streamed response. | TTFT (p95) and Time to first token. Streaming instrumentations, such as OpenLIT, send this attribute in seconds. | None |
gen_ai.response.finish_reasons | The reasons that the model stopped, for example ["stop"]. | Finish reasons. | llm.finish_reason, llm.response.stop_reason, llm.response.finish_reason, gen_ai.response.finish_reason, ai.response.finishReason |
gen_ai.input.messages | The prompt messages. | The input column of the Explorer, from the earliest span in the trace that has it. | gen_ai.prompt, ai.prompt.messages, input.value |
gen_ai.output.messages | The completion messages. | The output column of the Explorer, from the latest span in the trace that has it. | gen_ai.completion, ai.response.toolCalls, ai.response.text, output.value |
gen_ai.conversation.id | The ID of the conversation or session. | None | gen_ai.client.session_id, session.id, langfuse.session.id |
Tool spans
The default gen_ai.tool mapping group runs on spans with an attribute key that contains tool.name, ai.toolCall., tool_call_args, tool_response, or tool_use_id.
| Attribute | Holds | Used by | Default source keys |
|---|---|---|---|
gen_ai.tool.name | The name of the tool. | All Tool calls panels. SigNoz counts a span with this attribute as a tool call. | gen_ai.client.tool_name, tool_name, tool.name, ai.toolCall.name |
gen_ai.tool.call.id | The ID of the tool call. | None | gen_ai.client.tool_use_id, tool.id, ai.toolCall.id |
gen_ai.tool.description | The description of the tool. | None | tool.description |
gen_ai.tool.call.arguments | The arguments of the tool call. | None | ai.toolCall.args, traceloop.entity.input, gcp.vertex.agent.tool_call_args, input.value |
gen_ai.tool.call.result | The result of the tool call. | None | ai.toolCall.result, traceloop.entity.output, gcp.vertex.agent.tool_response, output.value |
gen_ai.operation.name | The type of operation, for example execute_tool or TOOL. | The Gen Ai Operation Name filter of the Explorer. | openinference.span.kind, llm.request.type, traceloop.span.kind |
Agent spans
The default gen_ai.agent mapping group runs on spans with an attribute key that contains agent.
| Attribute | Holds | Used by | Default source keys |
|---|---|---|---|
gen_ai.agent.name | The name of the agent. | Agent spans for the Per-trace usage panels. | crewai.agent.role, agent.name, agent_name |
gen_ai.agent.id | The ID of the agent. | None | agent_id, crewai.agent.id, agent.id |
gen_ai.agent.description | The description of the agent. | None | agent.description |
gen_ai.output.messages | The final result of the agent. | The output column of the Explorer, from the latest span in the trace that has it. | final_result |
gen_ai.operation.name | The type of operation, for example invoke_agent or AGENT. | The Gen Ai Operation Name filter of the Explorer. | openinference.span.kind, llm.request.type, traceloop.span.kind |
An AI trace is a trace with at least one LLM, tool, or agent span. Error panels use the error status of the span, not an attribute.
Resource attributes
| Attribute | Used by |
|---|---|
deployment.environment | The environment filter. |
service.name | The service_name filter and Cost by service. |
Tool spans do not carry a model. As a result, the model and provider filters do not affect the Tool calls panels or Calls per AI trace.
Next steps
- AI Observability overview: see cost, tokens, latency, and tool calls on one dashboard.
- Attribute Mapping: map other attribute names to the GenAI names.
- LLM Observability integrations: instrument OpenAI, Anthropic, LiteLLM, CrewAI, and other frameworks.
Get Help
If you need help with the steps in this topic, please reach out to us on SigNoz Community Slack. If you are a SigNoz Cloud user, please use in product chat support located at the bottom right corner of your SigNoz instance or contact us at cloud-support@signoz.io.
