llama_index/llama_index/prompts/default_...

The main prompts used are the text qa prompt and the refine prompt
https://github.com/jerryjliu/llama_index/blob/1dd6324355142f26e2040864efe64aff74ab285a/llama_index/response_synthesizers/factory.py#L44

https://github.com/jerryjliu/llama_index/blob/1dd6324355142f26e2040864efe64aff74ab285a/llama_index/prompts/default_prompts.py#L81

If you are using gpt-3.5 or gpt-4, then it uses the prompts from here
https://github.com/jerryjliu/llama_index/blob/1dd6324355142f26e2040864efe64aff74ab285a/llama_index/prompts/chat_prompts.py#L7

if response_mode == ACCUMULATE:
I can use refine_templat on responses array, to compose a new answer base on results? If Yes, using refine_templat will trigger a new call to LLM if I understand correctly ?

accumulate does not use the refine template

It just applies the text_qa_template to every chunk, and returns a concated list of all responses

so using ACCUMULATE I can only take responses and perform another query to LLM with new instructions, or there is another way to process all responses to a new result ?

You'd have to take the result and do something else with it.

Accumulate is mostly meant for for tasks where you want to apply the same query to every text chunk, and accumulate the response to each chunk