app.py

import os
import torch
import streamlit as st
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS
from langchain_core.prompts import PromptTemplate
from langchain.chains import LLMChain
from langchain_community.llms import HuggingFacePipeline
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
from dotenv import load_dotenv

# Set Streamlit page configuration
st.set_page_config(page_title="Chat with Notes and AI", page_icon=":books:", layout="wide")

# Load environment variables
load_dotenv()

# Dolly-v2-3b model pipeline
@st.cache_resource
def load_pipeline():
    model_name = "databricks/dolly-v2-3b"

    # Load tokenizer
    tokenizer = AutoTokenizer.from_pretrained(model_name, padding_side="left", trust_remote_code=True)

    # Load model with offload folder for disk storage of weights
    model = AutoModelForCausalLM.from_pretrained(
        model_name,
        torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,  # Use float16 for GPU, float32 for CPU
        device_map="auto",          # Automatically map model to available devices (e.g., GPU if available)
        trust_remote_code=True,
        offload_folder="./offload_weights"  # Folder to store offloaded weights
    )

    # Return text-generation pipeline
    return pipeline(
        task="text-generation",
        model=model,
        tokenizer=tokenizer,
        torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
        device_map="auto",
        return_full_text=True
    )

# Initialize Dolly pipeline
generate_text = load_pipeline()

# Create a HuggingFace pipeline wrapper for LangChain
hf_pipeline = HuggingFacePipeline(pipeline=generate_text)

# Template for instruction-only prompts
prompt = PromptTemplate(
    input_variables=["instruction"],
    template="{instruction}"
)

# Template for prompts with context
prompt_with_context = PromptTemplate(
    input_variables=["instruction", "context"],
    template="{instruction}\n\nInput:\n{context}"
)

# Create LLM chains
llm_chain = LLMChain(llm=hf_pipeline, prompt=prompt)
llm_context_chain = LLMChain(llm=hf_pipeline, prompt=prompt_with_context)

# Extracting text from .txt files
def get_text_files_content(folder):
    text = ""
    for filename in os.listdir(folder):
        if filename.endswith('.txt'):
            with open(os.path.join(folder, filename), 'r', encoding='utf-8') as file:
                text += file.read() + "\n"
    return text

# Converting text to chunks
def get_chunks(raw_text):
    from langchain.text_splitter import CharacterTextSplitter
    text_splitter = CharacterTextSplitter(
        separator="\n",
        chunk_size=1000,  # Reduced chunk size for faster processing
        chunk_overlap=200,  # Smaller overlap for efficiency
        length_function=len
    )
    chunks = text_splitter.split_text(raw_text)
    return chunks

# Using Hugging Face embeddings model and FAISS to create vectorstore
def get_vectorstore(chunks):
    embeddings = HuggingFaceEmbeddings(
        model_name="sentence-transformers/all-MiniLM-L6-v2",
        model_kwargs={'device': 'cpu'}  # Ensure embeddings use CPU
    )
    vectorstore = FAISS.from_texts(texts=chunks, embedding=embeddings)
    return vectorstore

# Generating response from user queries
def handle_question(question, vectorstore=None):
    if vectorstore:
        # Reduce the number of retrieved chunks for faster processing
        documents = vectorstore.similarity_search(question, k=2)
        context = "\n".join([doc.page_content for doc in documents])

        # Limit context to 1000 characters to speed up model inference
        context = context[:1000]

        if context:
            result_with_context = llm_context_chain.invoke({"instruction": question, "context": context})
            return result_with_context

    # Fallback to instruction-only chain if no context is found
    return llm_chain.invoke({"instruction": question})

def main():
    st.title("Chat with Notes :books:")

    # Initialize session state
    if "vectorstore" not in st.session_state:
        st.session_state.vectorstore = None

    # Folder for subject data
    data_folder = "data"

    # Subject selection
    subjects = [
        "A Trumped World", "Agri Tax in Punjab", "Assad's Fall in Syria", "Elusive National Unity", "Europe and Trump 2.0",
        "Going Down with Democracy", "Indonesia's Pancasila Philosophy", "Pakistan in Choppy Waters",
        "Pakistan's Semiconductor Ambitions", "Preserving Pakistan's Cultural Heritage", "Tackling Informal Economy",
        "Technical Education in Pakistan", "The Case for Solidarity Levies", "The Decline of the Sole Superpower",
        "The Power of Big Oil", "Trump 2.0 and Pakistan's Emerging Foreign Policy", "Trump and the World 2.0",
        "Trump vs BRICS", "US-China Trade War", "War on Humanity", "Women's Suppression in Afghanistan"
    ]
    subject_folders = {subject: os.path.join(data_folder, subject.replace(' ', '_')) for subject in subjects}
    selected_subject = st.sidebar.selectbox("Select a Subject:", subjects)

    # Process data folder for vectorstore
    subject_folder_path = subject_folders[selected_subject]
    raw_text = ""
    if os.path.exists(subject_folder_path):
        raw_text = get_text_files_content(subject_folder_path)
        if raw_text:
            text_chunks = get_chunks(raw_text)
            vectorstore = get_vectorstore(text_chunks)
            st.session_state.vectorstore = vectorstore
        else:
            st.warning("No content found for the selected subject.")
    else:
        st.error(f"Folder not found for {selected_subject}.")

    # Display preview of notes
    if raw_text:
        st.subheader("Preview of Notes")
        st.text_area("Preview Content:", value=raw_text[:2000], height=300, disabled=True)  # Show a snippet of the notes

    # Chat interface
    st.subheader("Ask Your Question")
    question = st.text_input("Ask a question about your selected subject:")
    if question:
        if st.session_state.vectorstore:
            response = handle_question(question, st.session_state.vectorstore)
            st.subheader("Answer:")
            st.write(response.get("text", "No response found."))
        else:
            st.warning("Please load the content for the selected subject before asking a question.")

if __name__ == '__main__':
    main()