app.py

import os
import streamlit as st
import pickle
import time
import langchain
#from langchain import OpenAI
#from langchain.chains import RetrievalQAWithSourcesChain
#from langchain.chains.qa_with_sources.loading import load_qa_with_sources_chain
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.document_loaders import UnstructuredURLLoader
#from langchain.embeddings import OpenAIEmbeddings
from langchain.vectorstores import FAISS
import requests
import pandas as pd
from langchain_community.llms import HuggingFaceEndpoint
from sentence_transformers import SentenceTransformer
from langchain.document_loaders import TextLoader
from sentence_transformers import SentenceTransformer, util
from langchain.schema import SystemMessage, HumanMessage, AIMessage
import faiss
from dotenv import load_dotenv
load_dotenv()  # take environment variables from .env (especially openai api key)

def querypreprocess(query: str ):
  vec = model.encode(query)  #again embeddings of query by sentencetransformer and able to search the index vector.
  #svec = np.array(vec).reshape(1,-1)  # as 2D needed
  distances, I = index.search(vec, k=2)
  row_indices = I.tolist()[0]
  list1 = [docs[i].page_content for i in row_indices]
  str1 = " "
  str1 = str1.join(list1)
  #str1 = '\n'.join([str(message) for message in list1])
  #results = ' '.join(map(str, list1)) #list to string convert
  return str1
    
def augmented_prompt(query: str):
  messages = querypreprocess(query)
  source_knowledge =''.join([str(message) for message in messages])
  #source_knowledge =results
  augmented_prompt = f"""
  using the contexts below, answer the query.

  Contexts:
  {source_knowledge}
  Question: {query}
  Answer:"""
  return augmented_prompt

   
st.title("RockyBot: News Research Tool 📈")
st.sidebar.title("News Article URLs")

urls = []
for i in range(3):
    url = st.sidebar.text_input(f"URL {i+1}")
    urls.append(url)

process_url_clicked = st.sidebar.button("Process URLs")
file_path = "sentence_embeddings.pkl"

main_placeholder = st.empty()
#llm = OpenAI(temperature=0.9, max_tokens=500)
llm = HuggingFaceEndpoint(repo_id="mistralai/Mistral-7B-Instruct-v0.2")

if process_url_clicked:
    # load data
    loader = UnstructuredURLLoader(urls=urls)
    main_placeholder.text("Data Loading...Started...✅✅✅")
    data = loader.load()
    # split data
    text_splitter = RecursiveCharacterTextSplitter(
        separators=['\n\n', '\n', '.', ','],
        chunk_size=1000,
        chunk_overlap=0
    )
    main_placeholder.text("Text Splitter...Started...✅✅✅")
    docs = text_splitter.split_documents(data)
    # Create an array of text to embed
    sentences = []
    for i, row in enumerate(docs):
        sentences.append(row.page_content)
    # create embeddings and save it to FAISS index
    #embeddings = OpenAIEmbeddings()
    #vectorstore_openai = FAISS.from_documents(docs, embeddings)
    # initialize sentence transformer model
    model = SentenceTransformer('bert-base-nli-mean-tokens')
    # create sentence embeddings
    sentence_embeddings = model.encode(sentences)
    main_placeholder.text("Embedding Vector Started Building...✅✅✅")
    time.sleep(2)

    # Save the FAISS index to a pickle file
    with open(file_path, "wb") as f:
        pickle.dump(sentence_embeddings, f)

query = main_placeholder.text_input("Question: ")
if query:
    if os.path.exists(file_path):
        with open(file_path, "rb") as f:
            query = pickle.load(f)
            import faiss
            d = sentence_embeddings.shape[1]
            index = faiss.IndexFlatL2(d)   # build the index, d=size of vectors
            # here we assume xb contains a n-by-d numpy matrix of type float32
            index.add(sentence_embeddings)                  # add vectors to the index
            #chain = RetrievalQAWithSourcesChain.from_llm(llm=llm, retriever=vectorstore.as_retriever())
            #result = chain({"question": query}, return_only_outputs=True)
            # result will be a dictionary of this format --> {"answer": "", "sources": [] }
            #xq = model.encode([query])
            #k=2
            #D, I = index.search(xq, k=k)
            #result1 = [f'{i}: {sentences[i]}' for i in I[0]]
            messages = [
                SystemMessage(content="You are a helpful assistant."),
                HumanMessage(content=query),
                AIMessage(content="I am Great, Thank You, How Can I Help You.")
                ]
            prompt = augmented_prompt(query)
            messages.append(prompt)
            result = llm.invoke(messages)
            st.header("Answer")
            # Display sources, if available
            sources = result.get("sources", "")
            if sources:
                st.subheader("Sources:")
                sources_list = sources.split("\n")  # Split the sources by newline
                for source in sources_list:
                    st.write(source)