Module llmflex.Memory.memory_utils

Functions

def create_prompt_with_history(memory: Type[BaseChatMemory], user_input: str, llm: Type[BaseLLM], prompt_template: Optional[PromptTemplate] = None, system: Optional[str] = None, recent_token_limit: int = 400, relevant_token_limit: int = 300, relevance_score_threshold: float = 0.8, similarity_score_threshold: float = 0.5, tool_selector: Optional[ToolSelector] = None, knowledge_base: Optional[KnowledgeBase] = None, kb_token_limit: int = 500, kb_score_threshold: float = 0.0) ‑> Union[str, List[Dict[str, str]]]

Create the full prompt or a list of messages that can be passed to the prompt template given the conversation memory.

Args

memory : Type[BaseChatMemory]

Conversation memory.

user_input : str

Latest user query.

llm : Type[BaseLLM]

LLM for counting token.

prompt_template : Optional[PromptTemplate], optional

Prompt template to format the prompt. If None is given, the default prompt template of the llm will be used. Defaults to None.

system : Optional[str], optional

System messsage for the conversation. If None is given, the default system message from the chat memory will be used. Defaults to None.

recent_token_limit : int, optional

Token limit for the most recent conversation history. Defaults to 400.

relevant_token_limit : int, optional

Token limit for the relevant contents from older conversation history. Only used if the memory provided allow relevant content extraction or knowledge base is given. Defaults to 300.

relevance_score_threshold : float, optional

Score threshold for the reranker for relevant conversation history content extraction. Only used if the memory provided allow relevant content extraction or knowledge base is given. Defaults to 0.8.

similarity_score_threshold : float, optional

Score threshold for the vector database search for relevant conversation history content extraction. Only used if the memory provided allow relevant content extraction. Defaults to 0.5.

tool_selector : Optional[ToolSelector], optional

Tool selector with all the available tools for function calling. If None is given, function calling is not enabled. Defaults to None.

knowledge_base : Optional[KnowledgeBase], optional

Knowledge base for RAG. Defaults to None.

kb_token_limit : int, optional

Knowledge base search token limit. Defaults ot 500.

kb_score_threshold : float, optional

Knowledge base search relevance score threshold. Defaults to 0.0.

Returns

Union[str, List[Dict[str, str]]]

Return the full prompt as a string if function calling is not applicable. Otherwise a list of messages will be returned for further formatting for function calling.

Expand source code

def create_prompt_with_history(memory: Type[BaseChatMemory], user_input: str, 
        llm: Type[BaseLLM], prompt_template: Optional[PromptTemplate] = None,
        system: Optional[str] = None, recent_token_limit: int = 400, relevant_token_limit: int = 300, relevance_score_threshold: float = 0.8, 
        similarity_score_threshold: float = 0.5, tool_selector: Optional[ToolSelector] = None, 
        knowledge_base: Optional[KnowledgeBase] = None, kb_token_limit: int = 500, kb_score_threshold: float = 0.0) -> Union[str, List[Dict[str, str]]]:
    """Create the full prompt or a list of messages that can be passed to the prompt template given the conversation memory.

    Args:
        memory (Type[BaseChatMemory]): Conversation memory.
        user_input (str): Latest user query.
        llm (Type[BaseLLM]): LLM for counting token.
        prompt_template (Optional[PromptTemplate], optional): Prompt template to format the prompt. If None is given, the default prompt template of the llm will be used. Defaults to None.
        system (Optional[str], optional): System messsage for the conversation. If None is given, the default system message from the chat memory will be used. Defaults to None.
        recent_token_limit (int, optional): Token limit for the most recent conversation history. Defaults to 400.
        relevant_token_limit (int, optional): Token limit for the relevant contents from older conversation history. Only used if the memory provided allow relevant content extraction or knowledge base is given. Defaults to 300.
        relevance_score_threshold (float, optional): Score threshold for the reranker for relevant conversation history content extraction. Only used if the memory provided allow relevant content extraction or knowledge base is given. Defaults to 0.8.
        similarity_score_threshold (float, optional): Score threshold for the vector database search for relevant conversation history content extraction. Only used if the memory provided allow relevant content extraction. Defaults to 0.5.
        tool_selector (Optional[ToolSelector], optional): Tool selector with all the available tools for function calling. If None is given, function calling is not enabled. Defaults to None.
        knowledge_base (Optional[KnowledgeBase], optional): Knowledge base for RAG. Defaults to None.
        kb_token_limit (int, optional): Knowledge base search token limit. Defaults ot 500.
        kb_score_threshold (float, optional): Knowledge base search relevance score threshold. Defaults to 0.0.

    Returns:
        Union[str, List[Dict[str, str]]]: Return the full prompt as a string if function calling is not applicable. Otherwise a list of messages will be returned for further formatting for function calling.
    """
    from .long_short_memory import LongShortTermChatMemory
    import json
    prompt_template = llm.core.prompt_template if prompt_template is None else prompt_template
    system = memory.system if system is None else system
    user_input = user_input.strip()
    if isinstance(memory, LongShortTermChatMemory):
        mem_type = 'longshort'
    else:
        mem_type = 'base'
    
    messages = [dict(role='system', content=system.strip())]
    return_list: bool = False
    # Adding function metadata if tools are available
    if tool_selector is not None:
        if not tool_selector.is_empty:
            if prompt_template.allow_custom_role:
                messages.append(tool_selector.function_metadata)
                return_list = True
    short_mem = memory.get_token_memory(llm=llm, token_limit=recent_token_limit)
    messages.extend(prompt_template.format_history(history=short_mem, return_list=True))
    extra = dict()
    if knowledge_base:
        kb_content = knowledge_base.search(query=user_input, token_limit=kb_token_limit, relevance_score_threshold=kb_score_threshold)
        if kb_content:
            if prompt_template.allow_custom_role:
                kb_content = list(map(lambda x: dict(content=x.index, source=x.metadata['filename']), kb_content))
                messages.append(dict(role='relevant_contents_from_knowledge_base', content=json.dumps(kb_content, indent=4)))
            else:
                extra['relevant_contents_from_knowledge_base'] = list(map(lambda x: dict(content=x.index, source=x.metadata['filename']), kb_content))
    if mem_type =='longshort':
        long_mem = memory.get_long_term_memory(query=user_input, recent_history=short_mem, llm=llm, 
                token_limit=relevant_token_limit, similarity_score_threshold=similarity_score_threshold, relevance_score_threshold=relevance_score_threshold)
        if long_mem:
            if prompt_template.allow_custom_role:
                messages.append(
                    dict(role='relevant_contents_from_previous_conversation', content=json.dumps(long_mem, indent=4))
                )
            else:
                extra['relevant_contents_from_previous_conversation'] = long_mem
    content = json.dumps(extra, indent=4) + '\n\n' + user_input if len(list(extra.keys())) != 0 else user_input
    messages.append(dict(role='user', content=content))
    if return_list:
        return messages
    else:
        return prompt_template.create_custom_prompt(messages=messages)