utils.py

import os
import pickle
import re
from typing import List, Union

import faiss
from langchain import OpenAI, LLMChain
from langchain.agents import ConversationalAgent
from langchain.agents import Tool, AgentExecutor, LLMSingleActionAgent, AgentOutputParser
from langchain.chains import ConversationalRetrievalChain
from langchain.document_loaders import DirectoryLoader, TextLoader, UnstructuredHTMLLoader
from langchain.embeddings import OpenAIEmbeddings
from langchain.memory import ConversationBufferWindowMemory
from langchain.prompts import BaseChatPromptTemplate
from langchain.schema import AgentAction, AgentFinish, HumanMessage
from langchain.text_splitter import CharacterTextSplitter
from langchain.vectorstores.faiss import FAISS

pickle_file = "open_ai.pkl"
index_file = "open_ai.index"

gpt_3_5 = OpenAI(model_name='gpt-4',temperature=0)

embeddings = OpenAIEmbeddings()

chat_history = []

memory = ConversationBufferWindowMemory(memory_key="chat_history")

gpt_3_5_index = None

class CustomOutputParser(AgentOutputParser):

    def parse(self, llm_output: str) -> Union[AgentAction, AgentFinish]:
        # Check if agent replied without using tools
        if "AI:" in llm_output:
            return AgentFinish(return_values={"output": llm_output.split("AI:")[-1].strip()},
                               log=llm_output)
        # Check if agent should finish
        if "Final Answer:" in llm_output:
            return AgentFinish(
                # Return values is generally always a dictionary with a single `output` key
                # It is not recommended to try anything else at the moment :)
                return_values={"output": llm_output.split("Final Answer:")[-1].strip()},
                log=llm_output,
            )
        # Parse out the action and action input
        regex = r"Action: (.*?)[\n]*Action Input:[\s]*(.*)"
        match = re.search(regex, llm_output, re.DOTALL)
        if not match:
            raise ValueError(f"Could not parse LLM output: `{llm_output}`")
        action = match.group(1).strip()
        action_input = match.group(2)
        # Return the action and action input
        return AgentAction(tool=action, tool_input=action_input.strip(" ").strip('"'), log=llm_output)

# Set up a prompt template
class CustomPromptTemplate(BaseChatPromptTemplate):
    # The template to use
    template: str
    # The list of tools available
    tools: List[Tool]

    def format_messages(self, **kwargs) -> str:
        # Get the intermediate steps (AgentAction, Observation tuples)
        # Format them in a particular way
        intermediate_steps = kwargs.pop("intermediate_steps")
        thoughts = ""
        for action, observation in intermediate_steps:
            thoughts += action.log
            thoughts += f"\nObservation: {observation}\nThought: "
        # Set the agent_scratchpad variable to that value
        kwargs["agent_scratchpad"] = thoughts
        # Create a tools variable from the list of tools provided
        kwargs["tools"] = "\n".join([f"{tool.name}: {tool.description}" for tool in self.tools])
        # Create a list of tool names for the tools provided
        kwargs["tool_names"] = ", ".join([tool.name for tool in self.tools])
        formatted = self.template.format(**kwargs)
        return [HumanMessage(content=formatted)]

def get_search_index():
    global gpt_3_5_index
    if os.path.isfile(pickle_file) and os.path.isfile(index_file) and os.path.getsize(pickle_file) > 0:
        # Load index from pickle file
        with open(pickle_file, "rb") as f:
            search_index = pickle.load(f)
    else:
        search_index = create_index()

    gpt_3_5_index = search_index


def create_index():
    source_chunks = create_chunk_documents()
    search_index = search_index_from_docs(source_chunks)
    faiss.write_index(search_index.index, index_file)
    # Save index to pickle file
    with open(pickle_file, "wb") as f:
        pickle.dump(search_index, f)
    return search_index


def search_index_from_docs(source_chunks):
    # print("source chunks: " + str(len(source_chunks)))
    # print("embeddings: " + str(embeddings))
    search_index = FAISS.from_documents(source_chunks, embeddings)
    return search_index


def get_html_files():
    loader = DirectoryLoader('docs', glob="**/*.html", loader_cls=UnstructuredHTMLLoader, recursive=True)
    document_list = loader.load()
    return document_list


def fetch_data_for_embeddings():
    document_list = get_text_files()
    document_list.extend(get_html_files())
    print("document list" + str(len(document_list)))
    return document_list


def get_text_files():
    loader = DirectoryLoader('docs', glob="**/*.txt", loader_cls=TextLoader, recursive=True)
    document_list = loader.load()
    return document_list


def create_chunk_documents():
    sources = fetch_data_for_embeddings()

    splitter = CharacterTextSplitter(separator=" ", chunk_size=800, chunk_overlap=0)

    source_chunks = splitter.split_documents(sources)

    print("sources" + str(len(source_chunks)))

    return source_chunks


def get_qa_chain(gpt_3_5_index):
    global gpt_3_5
    return ConversationalRetrievalChain.from_llm(gpt_3_5, chain_type="stuff", get_chat_history=get_chat_history,
            retriever=gpt_3_5_index.as_retriever(), return_source_documents=True, verbose=True)

def get_chat_history(inputs) -> str:
    res = []
    for human, ai in inputs:
        res.append(f"Human:{human}\nAI:{ai}")
    return "\n".join(res)


def generate_answer(question) -> str:
    global chat_history, gpt_3_5_index
    gpt_3_5_chain = get_qa_chain(gpt_3_5_index)
    result = gpt_3_5_chain(
        {"question": question, "chat_history": chat_history, "vectordbkwargs": {"search_distance": 0.6}})
    chat_history = [(question, result["answer"])]
    sources = []

    for document in result['source_documents']:
        source = document.metadata['source']
        sources.append(source.split('/')[-1].split('.')[0])

    source = ',\n'.join(set(sources))
    return result['answer'] + '\nSOURCES: ' + source


def get_agent_chain(prompt, tools):
    global gpt_3_5
    llm_chain = LLMChain(llm=gpt_3_5, prompt=prompt)
    agent = ConversationalAgent(llm_chain=llm_chain, tools=tools, verbose=True)
    agent_chain = AgentExecutor.from_agent_and_tools(agent=agent, tools=tools, verbose=True, memory=memory,
                                                     intermediate_steps=True)
    return agent_chain


def get_prompt_and_tools():
    tools = get_tools()

    prefix = """Have a conversation with a human, answering the following questions as best you can. Always try to use Vectorstore first. Your name is Coursera Bot because your knowledge base is Coursera course. You have access to the following tools:"""
    suffix = """Begin! If you used vectorstore tool, ALWAYS return a "SOURCES" part in your answer"
    
    {chat_history}
    Question: {input}
    {agent_scratchpad}
    sources:"""
    prompt = ConversationalAgent.create_prompt(
        tools,
        prefix=prefix,
        suffix=suffix,
        input_variables=["input", "chat_history", "agent_scratchpad"]
    )
    return prompt, tools


def get_tools():
    tools = [
        Tool(
            name="Vectorstore",
            func=generate_answer,
            description="useful for when you need to answer questions about the coursera course on 3D Printing.",
            return_direct=True
        )]
    return tools

def get_custom_agent(prompt, tools):

    llm_chain = LLMChain(llm=gpt_3_5, prompt=prompt)

    output_parser = CustomOutputParser()
    tool_names = [tool.name for tool in tools]
    agent = LLMSingleActionAgent(
        llm_chain=llm_chain,
        output_parser=output_parser,
        stop=["\nObservation:"],
        allowed_tools=tool_names
    )
    agent_executor = AgentExecutor.from_agent_and_tools(agent=agent, tools=tools, verbose=True, memory=memory,
                                                        intermediate_steps=True)
    return agent_executor

def get_prompt_and_tools_for_custom_agent():
    template = """
    Have a conversation with a human, answering the following questions as best you can. 
    ALWAYS try to use Vectorstore first. 
    You are a teaching assistant for a Coursera Course: The 3D Printing Evolution and can answer any question about that using vectorstore . You have access to the following tools:
    
    {tools}

    ALWAYS use one of the 2 formats listed below to respond. 

    To answer for the new input, use the following format:
    
    New Input: the input question you must answer
    Thought: Do I need to use a tool? Yes
    Action: the action to take, should be one of [{tool_names}]
    Action Input: the input to the action
    Observation: the result of the action
    ... (this Thought/Action/Action Input/Observation can repeat N times)
    Thought: I now know the final answer
    Final Answer: the final answer to the original input question. SOURCES: the sources referred to find the final answer

    
    When you have a response to say to the Human and DO NOT need to use a tool:
    1. DO NOT return "SOURCES" if you did not use any tool.
    2. You MUST use this format:
    ```
    Thought: Do I need to use a tool? No
    AI: [your response here]
    ```

    Begin! Remember to speak as a personal assistant when giving your final answer.
    ALWAYS return a "SOURCES" part in your answer, if you used any tool. 
    
    Previous conversation history:
    {chat_history}
    New input: {input}
    {agent_scratchpad}
    SOURCES:"""
    tools = get_tools()
    prompt = CustomPromptTemplate(
        template=template,
        tools=tools,
        # This omits the `agent_scratchpad`, `tools`, and `tool_names` variables because those are generated dynamically
        # This includes the `intermediate_steps` variable because that is needed
        input_variables=["input", "intermediate_steps", "chat_history"]
    )
    return prompt, tools