app.py

from flask import Flask, render_template, request, redirect, url_for, session, flash
import os
from werkzeug.utils import secure_filename
#from retrival import generate_data_store
from retrival import generate_data_store #,add_document_to_existing_db, delete_chunks_by_source
from langchain_community.vectorstores import Chroma
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.prompts import ChatPromptTemplate
from langchain_core.prompts import PromptTemplate, ChatPromptTemplate
from langchain_huggingface import HuggingFaceEndpoint
from huggingface_hub import InferenceClient
from langchain.schema import Document 
from langchain_core.documents import Document
from dotenv import load_dotenv
import re
import numpy as np
import glob
import shutil
from werkzeug.utils import secure_filename
import asyncio

import nltk
nltk.download('punkt_tab')

import nltk
nltk.download('averaged_perceptron_tagger_eng')

app = Flask(__name__)

# Set the secret key for session management
app.secret_key = os.urandom(24)

# Configurations
UPLOAD_FOLDER = "uploads/"
VECTOR_DB_FOLDER = "VectorDB/"
TABLE_DB_FOLDER = "TableDB/"

app.config['UPLOAD_FOLDER'] = UPLOAD_FOLDER
app.config['DEBUG'] = True
app.config['ENV'] = 'development'


os.makedirs(UPLOAD_FOLDER, exist_ok=True)
os.makedirs(VECTOR_DB_FOLDER, exist_ok=True)
os.makedirs(TABLE_DB_FOLDER, exist_ok=True)

# Global variables
CHROMA_PATH = None
TABLE_PATH = None

PROMPT_TEMPLATE_DOC = """
<s>[INST] You are a retrieval-augmented generation (RAG) assistant. Your task is to generate a response strictly based on the given context. Follow these instructions:

- Use only the provided context; do not add external information.
- The context contains multiple retrieved chunks separated by "###". Choose only the most relevant chunks to answer the question and ignore unrelated ones.
- If available, use the provided source information to support the response.
- Answer concisely and factually.

Context:
{context}

---

Question:
{question}

Response:
[/INST]

"""
# prompt if the document having the tables
PROMPT_TEMPLATE_TAB = """
<s>[INST] You are a retrieval-augmented generation (RAG) assistant. Your task is to generate a response strictly based on the given context. Follow these instructions:

- Use only the provided context; do not add external information.
- The context contains multiple retrieved chunks separated by "###". Choose only the most relevant chunks to answer the question and ignore unrelated ones.
- If available, use the provided source information to support the response.
- If a table is provided as html, incorporate its relevant details into the response while maintaining a structured format.
- Answer concisely and factually.

Context:
{context}

---

Table:
{table}  

---

Question:
{question}

Response:
[/INST]

"""


#HFT = os.getenv('HF_TOKEN')
#client = InferenceClient(api_key=HFT)

@app.route('/', methods=['GET'])
def home():
    return render_template('home.html')

@app.route('/chat', methods=['GET', 'POST'])
def chat():
    
    if 'history' not in session:
        session['history'] = []
    print("sessionhist1",session['history'])
    
    global CHROMA_PATH
    global TABLE_PATH
    
    old_db = session.get('old_db', None) 
    print(f"Selected DB: {CHROMA_PATH}")
    
    # if old_db != None:
    #     if CHROMA_PATH != old_db:
    #         session['history'] = [] 
        
    #print("sessionhist1",session['history'])
    
    if request.method == 'POST':
        query_text = request.form['query_text']
        if CHROMA_PATH is None:
            flash("Please select a database first!", "error")
            return redirect(url_for('list_dbs'))

        
        # Load the selected Document Database 
        embedding_function = HuggingFaceEmbeddings(model_name="all-MiniLM-L6-v2")
        #embedding_function = HuggingFaceEmbeddings(model_name="mixedbread-ai/mxbai-embed-large-v1")
        db = Chroma(persist_directory=CHROMA_PATH, embedding_function=embedding_function)
        # Convert the query to its embedding vector
        query_embedding = embedding_function.embed_query(query_text)
        if isinstance(query_embedding, float):
            query_embedding = [query_embedding]
        # print(f"Query embedding: {query_embedding}")
        # print(f"Type of query embedding: {type(query_embedding)}")
        # print(f"Length of query embedding: {len(query_embedding) if isinstance(query_embedding, (list, np.ndarray)) else 'Not applicable'}")
        results_document = db.similarity_search_by_vector_with_relevance_scores(
            embedding=query_embedding,  # Pass the query embedding
            k=3,
            #filter=filter_condition         # Pass the filter condition
        )
        
        print("results------------------->",results_document)  
        print("============================================")  
        print("============================================")
           
        context_text_document = "   \n\n###\n\n   ".join(
            [f"Source: {doc.metadata.get('source', '')} Page_content:{doc.page_content}\n" for doc, _score in results_document]
        )
        
        # Loading Table Database only if available
        if TABLE_PATH is not None:
            #embedding_function = HuggingFaceEmbeddings(model_name="all-MiniLM-L6-v2")
            embedding_function = HuggingFaceEmbeddings(model_name="mixedbread-ai/mxbai-embed-large-v1")
            tdb = Chroma(persist_directory=TABLE_PATH, embedding_function=embedding_function)
            results_table = tdb.similarity_search_by_vector_with_relevance_scores(
                embedding=query_embedding,  # Pass the query embedding
                k=2
                #filter=filter_condition         # Pass the filter condition
            )
            print("results------------------->",results_table)       
            context_text_table = "\n\n---\n\n".join([doc.page_content for doc, _score in results_table])

            # Prepare the prompt and query the model
            prompt_template = ChatPromptTemplate.from_template(PROMPT_TEMPLATE_TAB)
            prompt = prompt_template.format(context=context_text_document,table=context_text_table,question=query_text)
            #prompt = prompt_template.format(context=context_text_document,table=context_text_table, question=query_text)
            print("results------------------->",prompt)
        else:
            # Prepare the prompt and query the model
            prompt_template = ChatPromptTemplate.from_template(PROMPT_TEMPLATE_DOC)
            prompt = prompt_template.format(context=context_text_document,question=query_text)
            #prompt = prompt_template.format(context=context_text_document,table=context_text_table, question=query_text)
            print("results------------------->",prompt)
            
        #Model Defining and its use 
        repo_id = "mistralai/Mistral-7B-Instruct-v0.3"
        HFT = os.environ["HF_TOKEN"]
        llm = HuggingFaceEndpoint(
            repo_id=repo_id,
            #max_tokens=3000,
            max_new_tokens=2000,
            temperature=0.8,
            huggingfacehub_api_token=HFT,
        )

        data= llm(prompt)
        #data = response.choices[0].message.content

        # filtering the uneccessary context.
        if re.search(r'\bmention\b|\bnot mention\b|\bnot mentioned\b|\bnot contain\b|\bnot include\b|\bnot provide\b|\bdoes not\b|\bnot explicitly\b|\bnot explicitly mentioned\b', data, re.IGNORECASE):
            data = "We do not have information related to your query on our end."
            
        # Save the query and answer to the session history
        session['history'].append((query_text, data))
        
        # Mark the session as modified to ensure it gets saved
        session.modified = True
        print("sessionhist2",session['history'])

        return render_template('chat.html', query_text=query_text, answer=data, history=session['history'],old_db=CHROMA_PATH)

    return render_template('chat.html', history=session['history'], old_db=CHROMA_PATH)

@app.route('/create-db', methods=['GET', 'POST'])
def create_db():
    if request.method == 'POST':
        db_name = request.form.get('db_name', '').strip()
        if not db_name:
            return "Database name is required", 400

        # Get uploaded files
        files = request.files.getlist('folder')  # Folder uploads (multiple files)
        single_files = request.files.getlist('file')  # Single file uploads

        print("==================folder==>", files)
        print("==================single_files==>", single_files)

        # Ensure at least one valid file is uploaded
        if not any(file.filename.strip() for file in files) and not any(file.filename.strip() for file in single_files):
            return "No files uploaded", 400

        # Create upload directory
        upload_base_path = os.path.join(app.config['UPLOAD_FOLDER'], secure_filename(db_name))
        print(f"Base Upload Path: {upload_base_path}")
        os.makedirs(upload_base_path, exist_ok=True)

        # Process single file uploads first (if any exist)
        if any(file.filename.strip() for file in single_files):
            for file in single_files:
                if file.filename.strip():  # Ensure the file is valid
                    file_name = secure_filename(file.filename)
                    file_path = os.path.join(upload_base_path, file_name)
                    print(f"Saving single file to: {file_path}")
                    file.save(file_path)

            # If single file is uploaded, skip folder processing
            print("Single file uploaded, skipping folder processing.")
            asyncio.run(generate_data_store(upload_base_path, db_name))
            return redirect(url_for('list_dbs'))

        # Process folder files only if valid files exist
        if any(file.filename.strip() for file in files):
            for file in files:
                if file.filename.strip():  # Ensure it's a valid file
                    file_name = secure_filename(file.filename)
                    file_path = os.path.join(upload_base_path, file_name)
                    print(f"Saving folder file to: {file_path}")
                    file.save(file_path)

        # Generate datastore
        asyncio.run(generate_data_store(upload_base_path, db_name))
        return redirect(url_for('list_dbs'))

    return render_template('create_db.html')


@app.route('/list-dbs', methods=['GET'])
def list_dbs():
    vector_dbs = [name for name in os.listdir(VECTOR_DB_FOLDER) if os.path.isdir(os.path.join(VECTOR_DB_FOLDER, name))]
    return render_template('list_dbs.html', vector_dbs=vector_dbs)

@app.route('/select-db/<db_name>', methods=['POST'])
def select_db(db_name):
    flash(f"{db_name} Database has been selected", "table_selected")
    #Selecting the Documnet Vector DB
    global CHROMA_PATH
    global TABLE_PATH
    print(f"Selected DB: {CHROMA_PATH}")
    print("-----------------------------------------------------1----")
    CHROMA_PATH = os.path.join(VECTOR_DB_FOLDER, db_name)
    CHROMA_PATH = CHROMA_PATH.replace("\\", "/")
    print(f"Selected DB: {CHROMA_PATH}")
    print("-----------------------------------------------------2----")
    
    # Selecting the Table Vector DB
    table_db_path = os.path.join(TABLE_DB_FOLDER, db_name)
    table_db_path = table_db_path.replace("\\", "/")
    TABLE_PATH = table_db_path if os.path.exists(table_db_path) else None
    print(f"Selected Table DB: {TABLE_PATH}")
    
    return redirect(url_for('chat'))

@app.route('/update-dbs/<db_name>', methods=['GET','POST'])
def update_db(db_name):
    if request.method == 'POST':
        db_name = request.form['db_name']

        # Get all files from the uploaded folder
        files = request.files.getlist('folder')
        if not files:
            return "No files uploaded", 400
        print(f"Selected DB: {db_name}")
        DB_PATH = os.path.join(VECTOR_DB_FOLDER, db_name)
        DB_PATH = DB_PATH.replace("\\", "/")
        print(f"Selected DB: {DB_PATH}")

        generate_data_store(DB_PATH, db_name)

        return redirect(url_for('list_dbs'))
    return render_template('update_db.html')

if __name__ == "__main__":
    app.run(debug=False, use_reloader=False)