!pip install python-dotenv 
!pip install langchain -U langchain-community 
!pip install PyMuPDF 
!pip install rank-bm25 
!pip install deepeval 
!pip install langchain_ollama 
!pip install pypdf 
!pip install sentence-transformers 
!pip install faiss-gpu

!pip install colab-xterm
%load_ext colabxterm

%xterm

curl -fsSL https://ollama.com/install.sh | sh

ollama serve

ollama pull llama3.1

from concurrent.futures import ThreadPoolExecutor
from langchain.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.embeddings import SentenceTransformerEmbeddings
from langchain.vectorstores import FAISS
from langchain.prompts import PromptTemplate
from langchain_community.llms import Ollama
from langchain.schema import Document
import textwrap
import fitz  # PyMuPDF
import asyncio
import random
import numpy as np
import json
from typing import List
from rank_bm25 import BM25Okapi

def replace_tabs_with_spaces(docs: List[Document]) -> List[Document]:
    for doc in docs:
        doc.page_content = doc.page_content.replace('\t', ' ')
    return docs

def pdf_to_vectorstore(path: str, chunk_size=1000, chunk_overlap=200) -> FAISS:
    loader = PyPDFLoader(path)
    docs = loader.load()

    with ThreadPoolExecutor() as executor:
        text_splitter = RecursiveCharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap, length_function=len)
        chunks = list(executor.map(lambda d: text_splitter.split_documents([d]), docs))

    chunks = [item for sublist in chunks for item in sublist]
    cleaned_chunks = replace_tabs_with_spaces(chunks)

    embeddings = SentenceTransformerEmbeddings(model_name="all-MiniLM-L6-v2")
    vectorstore = FAISS.from_documents(cleaned_chunks, embeddings)
    
    return vectorstore

def initialize_llm_model():
    return Ollama(model="llama3.1", temperature=0)

def create_qa_chain(llm):
    question_answer_prompt_template = """ 
    For the question below, provide a concise but sufficient answer based ONLY on the provided context:
    {context}
    Question:
    {question}
    """

    question_answer_prompt = PromptTemplate(template=question_answer_prompt_template, input_variables=["context", "question"])
    question_answer_chain = question_answer_prompt | llm
    return question_answer_chain

class HyDERetriever:
    def __init__(self, files_path, chunk_size=500, chunk_overlap=100):
        self.llm = initialize_llm_model()
        self.chunk_size = chunk_size
        self.chunk_overlap = chunk_overlap
        self.vectorstore = pdf_to_vectorstore(files_path, chunk_size=self.chunk_size, chunk_overlap=self.chunk_overlap)

        self.hyde_prompt = PromptTemplate(
            input_variables=["query", "chunk_size"],
            template="""Given the question '{query}', generate a hypothetical document that directly answers this question. The document should be detailed and in-depth.
                        The document size should be exactly {chunk_size} characters.""",
        )
        self.hyde_chain = self.hyde_prompt | self.llm

    def generate_hypothetical_document(self, query):
        input_variables = {"query": query, "chunk_size": self.chunk_size}
        result = self.hyde_chain.invoke(input_variables)
        return result if isinstance(result, str) else result.content

def retrieve_context_for_question(question: str, retriever) -> List[str]:
    if question in context_cache:
        return context_cache[question]

    docs = retriever.get_relevant_documents(question)
    context = [doc.page_content for doc in docs]
    context_cache[question] = context
    return context

class HyDERetriever:
    def __init__(self, files_path, chunk_size=500, chunk_overlap=100):
        # (Initialization code as above)

    def retrieve(self, query, k=3):
        hypothetical_doc = self.generate_hypothetical_document(query)
        similar_docs = self.vectorstore.similarity_search(hypothetical_doc, k=k)
        return similar_docs, hypothetical_doc

def correctness_eval(generated_answer: str, ground_truth: str) -> float:
    return 1.0 if generated_answer.strip().lower() == ground_truth.strip().lower() else 0.0

def faithfulness_eval(generated_answer: str, retrieved_documents: List[str]) -> float:
    context_string = " ".join(retrieved_documents).lower()
    return 1.0 if generated_answer.strip().lower() in context_string else 0.0

def relevance_eval(question: str, generated_answer: str) -> float:
    return 1.0 if question.lower() in generated_answer.lower() else 0.0

def run_evaluation(retriever, num_questions: int = 5) -> None:
    llm = initialize_llm_model()
    qa_chain = create_qa_chain(llm)

    with open("../data/q_a.json", "r", encoding="utf-8") as file:
        q_a_data = json.load(file)

    questions = [qa["question"] for qa in q_a_data][:num_questions]
    ground_truth_answers = [qa["answer"] for qa in q_a_data][:num_questions]

    retrieved_docs_batch = batch_retrieve_context(questions, retriever)
    retrieved_contexts = [" ".join(doc) for doc in retrieved_docs_batch]

    loop = asyncio.get_event_loop()
    generated_answers = loop.run_until_complete(batch_generate_answers(questions, retrieved_contexts, qa_chain))

    correctness_scores = np.array([correctness_eval(g["answer"], gt) for g, gt in zip(generated_answers, ground_truth_answers)])
    faithfulness_scores = np.array([faithfulness_eval(g["answer"], r) for g, r in zip(generated_answers, retrieved_docs_batch)])
    relevance_scores = np.array([relevance_eval(q, g["answer"]) for q, g in zip(questions, generated_answers)])

    print(f"Avg Correctness: {correctness_scores.mean()}\nAvg Faithfulness: {faithfulness_scores.mean()}\nAvg Relevance: {relevance_scores.mean()}")

if __name__ == "__main__":
    # Define the path to the PDF document
    path = "/path/to/your/Facebook.pdf"
    
    # Initialize the HyDERetriever with the path to the PDF document
    retriever = HyDERetriever(path)

    # Define a test query
    test_query = "What were the primary reasons for Facebook's rebranding to Meta Platforms, Inc. in 2021?"
    
    # Retrieve similar documents and generate a hypothetical document based on the test query
    results, hypothetical_doc = retriever.retrieve(test_query)

    # Extract the content from the retrieved documents
    docs_content = [doc.page_content for doc in results]

    # Print the generated hypothetical document
    print("Hypothetical Document:\n")
    print(wrap_text_with_width(hypothetical_doc) + "\n")

    # Print the content of the retrieved documents
    for i, doc in enumerate(docs_content):
        print(f"Context {i+1}:\n{wrap_text_with_width(doc)}\n")

Hypothetical Document:

**Meta Platforms, Inc. Rebranding Report**

**Executive Summary:**
In October 2021, Facebook, Inc. rebranded to Meta Pla
tforms, Inc., marking a significant shift in the company's identity. This report outlines the primary reasons behind thi
s transformation.

**Reasons for Rebranding:**

1. **Expansion of Services:** The rebranding reflects Facebook's evoluti
on into a comprehensive metaverse platform, encompassing virtual reality (VR), augmented reality (AR), and online social
 interactions.
2. **Diversification of Business:** By separating the company name from its primary product, Meta Platfor
ms, Inc. aims to distance itself from controversies surrounding Facebook, while emphasizing its broader technological am
bitions.
3. **Preparation for Future Growth:** The rebranding positions the company for future growth, as it prepares to
 expand into new markets and industries, such as virtual reality and online commerce.

**Conclusion:**
The rebranding of
 Facebook, Inc. to Meta Platforms, Inc. represents a strategic move towards a more comprehensive and forward-thinking id
entity, reflecting the company's commitment to innovation and expansion.

Context 1:
and advertising drove its growth, culminating in a highly anticipated IPO on May 18, 2012. The 
IPO raised $16 billion, 
making it one of the largest tech IPOs in history and valuing Facebook at 
$104 billion.
1.4 Evolution and Re branding
I
n October 2021, Facebook announced its rebranding to Meta Platforms, Inc., signaling a strategic 
shift towards building
 the “metaverse” — a collective virtual shared space created by the

Context 2:
Facebook: A Comprehensive
Introduction
Facebook, now known as Meta Platforms, Inc., has dramatically transformed the way
 people 
interact and engage with digital content. Since its creation in 2004, the platform has evolved from a 
college 
project into a leading global technology company. This detailed overview covers 
Facebook’s history, services, leadershi
p, location, workforce, and more.
1. History and Foundation
1.1 Origins and Early Development

Context 3:
operations and business strategy helped Facebook become one of the most profitable tech 
companies. Sandberg’s leadershi
p extended to advocating for gender equality in the workplace 
through her book “Lean In” and various public speaking en
gagements.
3.3 Recent Leadership Changes
In recent years, there have been several changes in Facebook’s leadership struc
ture. In 2021, the 
rebranding to Meta Platforms, Inc. marked a shift in focus towards developing the metaverse. The