With the removal of the chats.create() span (which was a local object construction, not an LLM call), we lost some span attributes that were previously captured from the config parameter passed to chats.create(). These config values are reused for every subsequent chat.sendMessage() / chat.sendMessageStream() call on that chat instance and should be welded onto those spans instead.
Missing attributes on chat.sendMessage() / chat.sendMessageStream() spans:
- gen_ai.request.model — from model
- gen_ai.request.temperature — from config.temperature
- gen_ai.request.top_p — from config.topP
- gen_ai.request.top_k — from config.topK
- gen_ai.request.max_tokens — from config.maxOutputTokens
- gen_ai.request.frequency_penalty — from config.frequencyPenalty
- gen_ai.request.presence_penalty — from config.presencePenalty
- gen_ai.request.available_tools — from config.tools
- gen_ai.system_instructions — from config.systemInstruction
Solution brainstorm: The relevant parameters should be captured during the chats.create() call and merged onto the LLM invocation spans.
With the removal of the
chats.create()span (which was a local object construction, not an LLM call), we lost some span attributes that were previously captured from the config parameter passed tochats.create(). These config values are reused for every subsequentchat.sendMessage()/chat.sendMessageStream()call on that chat instance and should be welded onto those spans instead.Missing attributes on
chat.sendMessage()/chat.sendMessageStream()spans:Solution brainstorm: The relevant parameters should be captured during the
chats.create()call and merged onto the LLM invocation spans.