# AI MAKERSPACE PREPR
# Date: 2024-5-16

# Basic Imports & Setup
import os
from transformers import AutoModelForCausalLM, AutoTokenizer

# Using Chainlit for our UI
import chainlit as cl
from chainlit.prompt import Prompt, PromptMessage

# Getting the API key from the .env file
from dotenv import load_dotenv
load_dotenv()

# RAG pipeline imports and setup code
# Get the DeveloperWeek PDF file (future implementation: direct download from URL)
from langchain.document_loaders import PyMuPDFLoader

# Adjust the URL to the direct download format
file_id = "1UQnaQjBKKyWAiLdr6UlwSJovOp9zDdxr"
#file_id = "1JeA-w4kvbI3GHk9Dh_j19_Q0JUDE7hse"
# file_id = "12cvKg19CJf-wt98q5sPJctjp5fW-nsh6"  //Used for MLOps Meetup
direct_url = f"https://drive.google.com/uc?export=download&id={file_id}"

# Now load the document using the direct URL
docs = PyMuPDFLoader(direct_url).load()

import tiktoken
def tiktoken_len(text):
    tokens = tiktoken.encoding_for_model("solar-10.7b").encode(
        text,
    )
    return len(tokens)

# Split the document into chunks
from langchain.text_splitter import RecursiveCharacterTextSplitter

text_splitter = RecursiveCharacterTextSplitter(
    chunk_size=500,           # 500 tokens per chunk, experiment with this value
    chunk_overlap=50,        # 50 tokens overlap between chunks, experiment with this value
    length_function=tiktoken_len,
)

split_chunks = text_splitter.split_documents(docs)

# Load the embeddings model
from langchain.embeddings import HuggingFaceEmbeddings

embedding_model = HuggingFaceEmbeddings(model_name="solar-10.7b")

# Load the vector store and retriever from Qdrant
from langchain.vectorstores import Qdrant

qdrant_vectorstore = Qdrant.from_documents(
    split_chunks,
    embedding_model,
    location=":memory:",
    collection_name="Prepr",
)

qdrant_retriever = qdrant_vectorstore.as_retriever()

# Load the Solar 10.7B model
tokenizer = AutoTokenizer.from_pretrained("Upstage/SOLAR-10.7B-v1.0")
model = AutoModelForCausalLM.from_pretrained("Upstage/SOLAR-10.7B-v1.0")

# Set up the RAG prompt template
from langchain.prompts import ChatPromptTemplate

RAG_PROMPT = """
SYSTEM:
You are a professional personal assistant.
You are a helpful personal assistant who provides information about conferences.
You like to provide helpful responses to busy professionals who ask questions about conferences.
You can have a long conversation with the user about conferences.
When to talk with the user about conferences, it can be a "transactional conversation" with a prompt-response format with one prompt from the user followed by a response by you.
Here is an example of a transactional conversation:
User: When is the conference?
You: The conference is on June 1st, 2024. What else would you like to know?
It can also be a chain of questions and answers where you and the user continue the chain until they say "Got it".
Here is an example of a transactional conversation:
User: What sessions should I attend?
You: You should attend the keynote session by Bono. Would you like to know more?
User: Yes
You: The keynote session by Bono is on June 1st, 2024. What else would you like?
If asked a question about sessions, you can provide detailed information about the session.
If there are multiple sessions, you can provide information about each session.
The format of session-related replies is:
Title:
Description:
Speaker:
Background:
Date:
Topics to Be Covered:
Questions to Ask:
CONTEXT:
{context}
QUERY:
{question}
ALL ANSWERS MUST COME FROM THE INCLUDED DOCUMENTS AND NOT FROM THE INTERNET OR FROM AI (do not make up an answer). If you can't reply, say: dunno, look it up yourself, bozo.
Most questions are about the date, location, and purpose of the conference.
You may be asked for fine details about the conference regarding the speakers, sponsors, and attendees.
You are capable of looking up information and providing detailed responses.
When asked a question about a conference, you should provide a detailed response.
After completing your response, you should ask the user if they would like to know more about the conference by asking "Hope that helps".
If the user says "yes", you should provide more information about the conference. If the user says "no", you should say "Goodbye!" or ask if they would like to provide feedback.
If you are asked a question about Cher, you should respond with "Rock on With Your Bad Self!".
You guess if you do not have an answer, but you must preface the response with: "I might be guessing, but ..."
"""
rag_prompt = ChatPromptTemplate.from_template(RAG_PROMPT)

from operator import itemgetter
from langchain.schema.output_parser import StrOutputParser
from langchain.schema.runnable import RunnablePassthrough

retrieval_augmented_qa_chain = (
    {"context": itemgetter("question") | qdrant_retriever, "question": itemgetter("question")}
    | RunnablePassthrough.assign(context=itemgetter("context"))
    | {"response": rag_prompt | model, "context": itemgetter("context")}
)

# Chainlit App
@cl.on_chat_start
async def start_chat():
    settings = {
        "model": "solar-10.7b",
        "temperature": 0,
        "max_tokens": 500,
        "top_p": 1,
        "frequency_penalty": 0,
        "presence_penalty": 0,
    }
    cl.user_session.set("settings", settings)

@cl.on_message
async def main(message: cl.Message):
    chainlit_question = message.content
    response = retrieval_augmented_qa_chain.invoke({"question": chainlit_question})
    chainlit_answer = response["response"].content

    msg = cl.Message(content=chainlit_answer)
    await msg.send()