Extended llm support (e.g. Llama 3, M8x22b) and synthetic test generation

[ciekawy]

Enhancements for new opensource models with focus on Llama 3 70b instruct

This is a draft PR - it introduces several improvements to RAGAs to enhance its compatibility with Llama 3 70b and improve the quality of generated outputs, especially for synthetic data generation.

Prompt Adjustments

• Modified prompts in prompt.py and prompts.py to increase the quality of generated outputs, which is particularly important for synthetic data generation.
• The updated prompts aim to provide clearer instructions and more relevant examples to guide the language model in generating higher-quality outputs.

Support for Non-Typical LangChain LLM Configurations

• Added support for LLM configurations that require additional parameters and dynamic calculations when invoking LLM generation.
• Specifically, the Together.ai Llama 3 70b instruct model requires extra parameters, including dynamically calculated values in a callback.
• Introduced a new LLMConfig class to encapsulate the custom configuration options.
• Implemented a together_prompt_callback function to handle the dynamic prompt generation for the Together.ai Llama 3 70b instruct model.

Example Configuration

def together_prompt_callback(
    prompt: PromptValue
) -> t.Tuple[t.List[PromptValue], t.Dict[str, t.Any]]:
    empty_prompt = StringPromptValue(text='')
    prompt_str = prompt.prompt_str
    messages = [{"content": prompt_str, "role": "user"}]
    return [empty_prompt], {"messages": messages}

together_llama3instruct_config = LLMConfig(
    stop=["<|eot_id|>"],
    prompt_callback=together_prompt_callback,
    type="chat",
    prompt_format_string="<human>: {prompt}\n<bot>:",
    request_type="language-model-inference",
)