LLM: define common/system-specific event structure

[lmolkova]

Area(s)

llm

Is your change request related to a problem? Please describe.

See https://github.com/open-telemetry/semantic-conventions/pull/825#discussion_r1533040974

825 is going to introduce LLM events and (currently) requires to put prompts and completions as json string in OpenAI format. Some LLM vendors might not use json and instrumentation would be forced to use specific structure.

We should decide if:

• we define a common format and all LLM vendor instrumentations are forced to use it.
• we want one event with OpenAI JSON messages format (or another common format) and other vendors are forced to use it.
• vendor-specific events. E.g. gen_ai.{vendor}.prompt and then vendors could document their own structure or document that they emit OpenAI-formatted events
• we make format dynamic with additional attributes representing mime type.

This is closely coupled with https://github.com/open-telemetry/semantic-conventions/issues/829

Describe the solution you'd like

TBD

[lmolkova]

Based on the discussion in this thread:

• we should consider recording individual choices as events rather than completions
• finish_reason should become a string rather than array and appear on events - this way it matches openai response structure
• we can still add more attributes or event payload fields to describe the structure of choice