LLM: initial semconv definition

[lmolkova]

trafficstars

Define initial semantic conventions that establishes a foundation with essential attributes, including a vendor-specific example, and basic event definitions:

• [x] pick a namespace: gen_ai (genai?)
• [x] define basic request and response attributes.
  • gen_ai.system - Name of the LLM foundation model system or vendor
  • gen_ai.request.max_tokens - Max number of tokens the LLM to generate per request
  • gen_ai.request.model - Name of the LLM model used for the request
  • gen_ai.request.temperature - Temperature setting for the LLM request
  • gen_ai.request.top_p - Top_p sampling setting for the LLM request
  • gen_ai.response.model - Name of the LLM model used for the response
  • gen_ai.response.finish_reason - Reason why the LLM stopped generating tokens
  • gen_ai.response.id - Unique identifier for the response
• [x] define usage attributes
  • requirement levels, if they belong on spans or events
  • gen_ai.usage.completion_tokens - Number of tokens used in the LLM response
  • gen_ai.usage.prompt_tokens - Number of tokens used in the LLM prompt ~~gen_ai.usage.total_tokens - Total number of tokens used in both prompt and response~~
• [ ] include at least one vendor example (openai) - @drewby - #1385
  • gen_ai.openai.request.logit_bias - The logit_bias used in the request
  • gen_ai.openai.request.presence_penalty - The presence_penalty used in the request
  • gen_ai.openai.request.seed - Seed used in request to improve determinism.
  • gen_ai.openai.request.response_format - Format of the LLM's response, e.g., text or JSON
  • gen_ai.openai.response.created - UNIX timestamp of when the response was created
• [ ] event definitions - @lmolkova - #980
  • sensitivity, requirement level, attributes
  • gen_ai.content.prompt - Captures the full prompt string sent to an LLM
  • gen_ai.content.completion - Captures the full response string from an LLM
• [x] general attributes - @lmolkova - #1297
  • server.address - Address of the server hosting the LLM
  • server.port - Port number used by the server
  • error.type
• [ ] metrics
• [ ] streaming