app.py

import os
import re
import lz4
import json
import time
import uuid
import torch
import base64
import asyncio
import msgpack
import validators
import numpy as np
import pandas as pd
import streamlit as st
from vllm import LLM
from numpy import ndarray
from outlines import models
from datetime import datetime
from typing import List, Dict
from pydantic import BaseModel, Field
from dense_embed import embed_text
from ppt_chunker import ppt_chunker
from qdrant_client import QdrantClient
from streamlit_navigation_bar import st_navbar
from vllm.sampling_params import SamplingParams
from fastembed import SparseTextEmbedding, SparseEmbedding
from outlines.fsm.json_schema import build_regex_from_schema
from unstructured.nlp.tokenize import download_nltk_packages
from scipy.sparse import csr_matrix, save_npz, load_npz, vstack
from infinity_emb import AsyncEngineArray, EngineArgs, AsyncEmbeddingEngine
from prompts import (
    transform_query,
    build_prompt_conv,
    route_llm,
    open_query_prompt,
    question_type_prompt,
    idk,
    self_knowledge,
    answer_with_context
)
from qdrant_client.models import (
    NamedSparseVector,
    NamedVector,
    SparseVector,
    PointStruct,
    ScoredPoint,
    Prefetch,
    FusionQuery,
    Fusion,
    SearchRequest,
    Modifier,
    OptimizersConfigDiff,
    HnswConfigDiff,
    Distance,
    VectorParams,
    SparseVectorParams,
    SparseIndexParams,
    Batch,
    Filter,
    HasIdCondition,
    Datatype,
    BinaryQuantization,
    BinaryQuantizationConfig
)

class Question(BaseModel):
    answer: str
schema = json.dumps(Question.model_json_schema())

icon_to_types = {
    'ppt':('data:image/png;base64,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',
          'Powerpoint'),
    'pptx':('data:image/png;base64,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',
          'Powerpoint'),
    'txt':('data:image/png;base64,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',
          'Txt'),
    'doc':('data:image/png;base64,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',
          'Microsoft Word'),
    'docx':('data:image/png;base64,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',
          'Microsoft Word'),
    'xslx':('data:image/png;base64,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',
           'Excel')
}

def query_keywords_search(query: str, client: QdrantClient, collection_name: str, sparse_embeddings):
    return client.query_points(
        collection_name=collection_name,
        prefetch=Prefetch(query=sparse_embeddings, using='title-sparse', limit=25),
        query=FusionQuery(fusion=Fusion.RRF),
        with_vectors=False,
        with_payload=True,
        limit=1
    )

def query_hybrid_search(query: str, client: QdrantClient, collection_name: str, dense_embeddings, sparse_embeddings):
    return client.query_points(
        collection_name=collection_name,
        prefetch=[
            Prefetch(query=sparse_embeddings, using="text-sparse", limit=25),
            Prefetch(query=dense_embeddings[0], using="text-dense", limit=25)
        ],
        query=FusionQuery(fusion=Fusion.RRF),
        with_vectors=False,
        with_payload=True,
        limit=10,
        score_threshold=0.95
    )
    
def generate_answer(query: str, 
                    client: QdrantClient, 
                    collection_name: str, 
                    llm, 
                    dense_model: AsyncEmbeddingEngine, 
                    sparse_model: SparseTextEmbedding, 
                    past_messages: str, 
                    search_strategy: str, 
                   ):
    sparse_embeddings = list(sparse_model.query_embed(query))[0].as_object()
    
    s = time.time()

    if search_strategy == 'Exact Search':
        scored_point = query_keywords_search(query, client, collection_name, sparse_embeddings).points[0]

        text = scored_point.payload['text']
        metadata = scored_point.payload['metadata']

        answer = f"{text}\n\n\nSource :\n\n{metadata}"
    else:
        regex = build_regex_from_schema(schema, r"[\n ]?")
        gen_text = outlines.generate.regex(llm, regex)

        gen_choice = outlines.generate.choice(llm, choices=['Yes', 'No'])
        prompt = route_llm(context, query)
        action = gen_choice(prompt, max_tokens=2, sampling_params=SamplingParams(temperature=0))
        print(f'Choice: {action}')
    
        if action == 'Yes':
            dense_embeddings, tokens_count = asyncio.run(embed_text(dense_model[0], transform_query(query)))
            
            scored_points = query_hybrid_search(query, client, collection_name, dense_embeddings, sparse_embeddings).points
            print(f'Score : {scored_points[0]}')   
            docs = [(scored_point.payload['text'], scored_point.payload['metadata']) for scored_point in scored_points]
            contents, metadatas = [list(t) for t in zip(*docs)]
            
            context = "\n".join(contents)
            print(f'Context : \n + {context}')
        
            filtered_metadatas = {
                value
                for metadata in metadatas
                if 'url' in metadata
                for value in [metadata['url']]
            }
            result_metadatas = "\n\n".join(f'{value}' for value in filtered_metadatas)
            
            prompt = answer_with_context(context, query)
            answer = json.loads(gen_text(prompt, max_tokens=300, sampling_params=SamplingParams(temperature=0)))['answer']
            answer = f"{answer}\n\n\nSource(s) :\n\n{result_metadatas}"
            
            if search_strategy == 'Documents + LLM Search':
                answer = f'Documents Based :\n\n{answer}'
        else:
            gen_choice = outlines.generate.choice(llm, choices=['Domain-Specific Question', 'General Question'])
            prompt = question_type_prompt(query)
            action = gen_choice(prompt, max_tokens=3, sampling_params=SamplingParams(temperature=0))
            print(f'Choice 2: {action}')
            
            if action == 'General Question':
                prompt = open_query_prompt(past_messages, query)
                answer = json.loads(gen_text(prompt, max_tokens=300, sampling_params=SamplingParams(temperature=0.6, top_p=0.9, top_k=10)))['answer']
            else:
                if search_strategy == 'Documents Only Search':
                    prompt = idk(query)
                    answer = json.loads(gen_text(prompt, max_tokens=128, sampling_params=SamplingParams(temperature=0.6, top_p=0.9, top_k=10)))['answer']
                elif search_strategy == 'Documents + LLM Search':
                    prompt = self_knowledge(query)
                    answer = json.loads(gen_text(prompt, max_tokens=300, sampling_params=SamplingParams(temperature=0.6, top_p=0.9, top_k=10)))['answer']
                    answer = f'Internal Knowledge :\n\n{answer}'

    torch.cuda.empty_cache()

    e = time.time()
    f = e - s
    print(f'SEARCH TIME : {f}')

    return answer

def collect_files(directory, pattern):
    array = []

    for filename in os.listdir(directory):
        if pattern in filename:
            if filename.endswith('.msgpack'):
                with open(os.path.join(directory, filename), "rb") as data_file_payload:
                    decompressed_payload = data_file_payload.read()
                array.extend(msgpack.unpackb(decompressed_payload, raw=False))
            elif filename.endswith('.npz') and (pattern == '_dense'):
                array.extend(list(np.load(os.path.join(directory, filename)).values()))
            elif filename.endswith('.npz') and ((pattern == '_sparse') or (pattern == '_sparse_titles')):
                sparse_embeddings = []
                loaded_sparse_matrix = load_npz(os.path.join(directory, filename))
                
                for i in range(loaded_sparse_matrix.shape[0]):
                    row = loaded_sparse_matrix.getrow(i)
                    values = row.data.tolist()
                    indices = row.indices.tolist()
                    embedding = SparseVector(indices=indices, values=values)
                    sparse_embeddings.append(embedding)
                array.extend(sparse_embeddings)
            elif (filename.endswith('.npy')):
                ids_list = np.load(os.path.join(directory, filename), allow_pickle=True).tolist()
                array.extend(ids_list)
                
    return array

@st.cache_resource(show_spinner=False)
def load_models_and_documents():
    container = st.empty()
    
    with container.status("Load AI Models and Prepare Documents...", expanded=True) as status:
        st.write('Downloading and Loading MixedBread Mxbai Dense Embedding Model with vLLM as backend...')

        dense_model = AsyncEngineArray.from_args(
            [
                EngineArgs(
                    model_name_or_path='mixedbread-ai/mxbai-embed-large-v1', 
                    engine='torch',
                    device='cuda',
                    embedding_dtype='float32',
                    dtype='float16',
                    pooling_method='cls',
                    lengths_via_tokenize=True
                )
            ]
        )

        st.write('Downloading and Loading Qdrant BM42 Sparse Embedding Model under ONNX using the CPU...')
        
        sparse_model = SparseTextEmbedding(
            'Qdrant/bm42-all-minilm-l6-v2-attentions',
            cache_dir=os.getenv('HF_HOME'),
            providers=['CPUExecutionProvider']
        )

        st.write('Downloading and Loading Mistral Nemo quantized with GPTQ and using Outlines + vLLM Engine as backend...')

        llm = LLM(
            model="shuyuej/Mistral-Nemo-Instruct-2407-GPTQ",
            tensor_parallel_size=1,
            enforce_eager=True,
            gpu_memory_utilization=1,
            max_model_len=10240,
            dtype=torch.float16,
            max_num_seqs=64,
            quantization="gptq"
        )
        model = models.VLLM(llm)

        st.write('Downloading NLTK Packages...')

        download_nltk_packages()

        st.write('Creating Collection for our Qdrant Vector Database in RAM memory...')
        
        client = QdrantClient(':memory:')
        collection_name = 'collection_demo'
        
        client.create_collection(
            collection_name,
            {
                'text-dense': VectorParams(
                    size=1024,
                    distance=Distance.COSINE,
                    datatype=Datatype.FLOAT16,
                    on_disk=False
                )
            },
            {
                'text-sparse': SparseVectorParams(
                    index=SparseIndexParams(
                        on_disk=False
                    ),
                    modifier=Modifier.IDF
                ),
                'title-sparse': SparseVectorParams(
                    index=SparseIndexParams(
                        on_disk=False
                    ),
                    modifier=Modifier.IDF
                )
            },
            2,
            optimizers_config=OptimizersConfigDiff(
                indexing_threshold=0,
                default_segment_number=4
            ),
            hnsw_config=HnswConfigDiff(
                on_disk=False, 
                m=32,
                ef_construct=200
            )
        )
        
        name = 'action_rpg'
        embeddings_path = os.path.join(os.getenv('HF_HOME'), 'embeddings')

        payload_path = os.path.join(embeddings_path, name + '_payload.msgpack')
        payload_titles_path = os.path.join(embeddings_path, name + '_payload_titles.npz')
        dense_path = os.path.join(embeddings_path, name + '_dense.npz')
        sparse_path = os.path.join(embeddings_path, name + '_sparse.npz')
        sparse_titles_path = os.path.join(embeddings_path, name + '_sparse_titles.npz')
        ids_path = os.path.join(embeddings_path, name + '_ids.npy')
        ids_titles_path = os.path.join(embeddings_path, name + '_ids_titles.npz')

        if not os.path.exists(embeddings_path):
            os.mkdir(embeddings_path)

            st.write('Downloading and Loading Video Games Dataset coming from Wikipedia...')

            dataset = pd.read_parquet(os.path.join(os.getenv('HOME'),'data', 'train_pages.parquet.zst'), engine='pyarrow')

            for columnName, columnData in dataset.iteritems():
                if columnName == 'text':
                    documents = columnData.values.tolist()
                elif columnName == 'section_title':
                    metadatas_titles = columnData.values.tolist()
                elif columnName == 'url':
                    metadatas_url = columnData.values.tolist()

            st.write('Transforming the Wikipedia Video Games Dataset into ingestable format for our Qdrant Vector Database...')
            
            payload_docs = [{ 'text': text, 'metadata': { 'url': url } } for text, url in zip(documents, metadatas_url)]
                
            start_sparse = time.time()
                
            sparse_embeddings = [SparseVector(indices=s.indices.tolist(), values=s.values.tolist()) for s in sparse_model.embed(metadatas_titles, 32)]
                
            end_sparse = time.time()
            final_sparse = end_sparse - start_sparse
            print(f'SPARSE TIME: {final_sparse}')

            st.write('Saving on disk the Wikipedia Video Games Dataset into quickly ingestable format...')

            with open(payload_titles_path, "wb") as outfile_texts:
                packed_payload = msgpack.packb(payload_docs, use_bin_type=True)
                outfile_texts.write(packed_payload)
            
            max_index = 0
            for embedding in sparse_embeddings:
                if len(embedding.indices) > 0:
                    max_index = max(max_index, max(embedding.indices))
        
            sparse_matrices = []
            for embedding in sparse_embeddings:
                data = np.array(embedding.values)
                indices = np.array(embedding.indices)
                indptr = np.array([0, len(data)])
                matrix = csr_matrix((data, indices, indptr), shape=(1, max_index + 1))
                sparse_matrices.append(matrix)
            
            combined_sparse_matrix = vstack(sparse_matrices)
            save_npz(sparse_titles_path, combined_sparse_matrix)

            unique_ids = []
            while len(unique_ids) < len(payload_docs):
                new_id = uuid.uuid4()
                while new_id.hex[0] == '0':
                    new_id = uuid.uuid4()
                unique_ids.append(new_id.int)
            
            np.save(ids_titles_path, np.array(unique_ids), allow_pickle=True)

            st.write('Ingesting saved documents on disk into our Qdrant Vector Database...')

            client.upsert(
                collection_name,
                points=Batch(
                    ids=unique_ids,
                    payloads=payload_docs,
                    vectors={
                        'title-sparse': sparse_embeddings,
                    }
                )
            )
            
            client.update_collection(
                collection_name=collection_name,
                optimizer_config=OptimizersConfigDiff(indexing_threshold=20000)
            )
        else:
            st.write('Loading the saved documents on disk')

            patterns = ['_ids', '_ids_titles', '_payload', '_payload_titles', '_dense', '_sparse', '_sparse_titles']

            unique_ids, unique_ids_titles, payload_docs, payload_docs_titles, dense_embeddings, sparse_embeddings, sparse_embeddings_titles = [
                collect_files(embeddings_path, pattern) for pattern in patterns
            ]

            st.write('Ingesting saved documents on disk into our Qdrant Vector Database...')

            client.upsert(
                collection_name,
                points=Batch(
                    ids=unique_ids,
                    payloads=payload_docs,
                    vectors={
                        'text-dense': dense_embeddings,
                        'text-sparse': sparse_embeddings
                    }
                )
            )

            client.upsert(
                collection_name,
                points=Batch(
                    ids=unique_ids_titles,
                    payloads=payload_docs_titles,
                    vectors={
                        'title-sparse': sparse_embeddings_titles
                    }
                )
            )
        
            client.update_collection(
                collection_name=collection_name,
                optimizer_config=OptimizersConfigDiff(indexing_threshold=20000)
            )

        st.write('Building FSM Index for Agentic Behaviour of our AI...')

        answer = generate_answer('aggro', client, collection_name, model, dense_model, sparse_model, '', 'Exact Search')
        
        status.update(
            label="Processing Complete!", state="complete", expanded=False
        )
        
    time.sleep(5)
    container.empty()
    
    return client, collection_name, llm, model, dense_model, sparse_model
    
    
if __name__ == '__main__':
    st.set_page_config(page_title="Multipurpose AI Agent",layout="wide", initial_sidebar_state='auto')

    client, collection_name, llm, model, dense_model, sparse_model, nlp, conn, cursor = load_models_and_documents()

    styles = {
        "nav": {
            "background-color": "rgb(204, 200, 194)",
        },
        "div": {
            "max-width": "32rem",
        },
        "span": {
            "border-radius": "0.5rem",
            "color": "rgb(125, 102, 84)",
            "margin": "0 0.125rem",
            "padding": "0.4375rem 0.625rem",
        },
        "active": {
            "background-color": "rgba(255, 255, 255, 0.25)",
        },
        "hover": {
            "background-color": "rgba(255, 255, 255, 0.35)",
        },
    }

    if 'menu_id' not in st.session_state:
        st.session_state.menu_id = 'ChatBot'
    
    st.session_state.menu_id = st_navbar(
        ['ChatBot', 'Documents'], 
        st.session_state.menu_id,
        options={
            'hide_nav': False,
            'fix_shadow': False,
            'use_padding': False
        },
        styles=styles
    )

    st.title('Multipurpose AI Agent')
    #st.markdown("<h1 style='position: fixed; top: 0; left: 0; width: 100%; padding: 10px; text-align: left; color: black;'>Multipurpose AI Agent</h1>", unsafe_allow_html=True)

    data_editor_path = os.path.join(os.getenv('HF_HOME'), 'documents')
    
    if 'df' not in st.session_state:
        if os.path.exists(data_editor_path):
            st.session_state.df = pd.read_parquet(os.path.join(data_editor_path, 'data_editor.parquet.lz4'), engine='pyarrow')
        else:
            st.session_state.df = pd.DataFrame()
            os.mkdir(data_editor_path)
            st.session_state.df.to_parquet(
                os.path.join(
                    data_editor_path, 
                    'data_editor.parquet.lz4'
                ), 
                compression='lz4', 
                engine='pyarrow'
            )

    def on_change_data_editor(client, collection_name):
        print(f'Check : {st.session_state.key_data_editor}')
        
        if st.session_state.key_data_editor['deleted_rows']:
            st.toast('Wait for deletion to complete...')
            embeddings_path = os.path.join(os.getenv('HF_HOME'), 'embeddings')
            
            for deleted_idx in st.session_state.key_data_editor['deleted_rows']:
                name = st.session_state.df.loc[deleted_idx, 'document']
                print(f'WHAT IS THAT : {name}')
                
                os.remove(os.path.join(embeddings_path, name + '_ids.npy'))
                os.remove(os.path.join(embeddings_path, name + '_payload.msgpack'))
                os.remove(os.path.join(embeddings_path, name + '_dense.npz'))
                os.remove(os.path.join(embeddings_path, name + '_sparse.npz'))
                
                client.delete(
                    collection_name=collection_name,
                    points_selector=Filter(
                        must=[
                            FieldCondition(
                                key='url',
                                match=MatchValue(value=st.session_state.df.loc[deleted_idx, 'path'])
                            )
                        ]
                    )
                )
                
                st.session_state.df.drop(deleted_idx)
                
            st.toast('Deletion Completed !', icon='🎉')
            
    if st.session_state.menu_id == 'Documents':       
        st.session_state.df = st.data_editor(
            st.session_state.df, 
            num_rows="dynamic",
            use_container_width=True,
            hide_index=True,
            on_change=on_change_data_editor,
            args=(client, collection_name),
            key='key_data_editor',
            column_config={
                'icon': st.column_config.ImageColumn(
                    'Document'
                ),
                "document": st.column_config.TextColumn(
                    "Name",
                    help="Name of the document",
                    required=True,
                    disabled=True
                ),
                "type": st.column_config.SelectboxColumn(
                    'File type',
                    help='The file format extension of this document',
                    required=True,
                    options=[
                        'Powerpoint',
                        'Microsoft Word',
                        'Excel'
                    ],
                    disabled=True
                ),
                "path": st.column_config.TextColumn(
                    'Path',
                    help='Path to the document',
                    required=False,
                    disabled=True
                ),
                "time": st.column_config.DatetimeColumn(
                    'Date and hour',
                    help='When this document has been ingested here for the last time',
                    format="D MMM YYYY, h:mm a",
                    required=True,
                    disabled=True
                ),
                "exact_search": st.column_config.CheckboxColumn(
                    'Exact Search',
                    help='Wether the Exact Search is available for this document or not',
                    required=True,
                    default=False,
                    disabled=True
                )
            }
        )

    conversations_path = os.path.join(os.getenv('HF_HOME'), 'conversations')
    try:
        with open(conversations_path, 'rb') as fp:
            packed_bytes = fp.read()
        conversations: Dict[str, list] = msgpack.unpackb(packed_bytes, raw=False)
    except:
        conversations = {}

    if st.session_state.menu_id == 'ChatBot':  
        if 'id_chat' not in st.session_state:
            st.session_state.id_chat = 'New Conversation'

        if 'search_strategy' not in st.session_state:
            st.session_state.search_strategy = 'Documents Only Search'
            st.session_state.tooltip = 'The AI answer your questions only considering the documents provided'
            st.session_state.search_idx = 1

        def options_list(conversations: Dict[str, list]):
            if st.session_state.id_chat == 'New Conversation':
                return [st.session_state.id_chat] + list(conversations.keys())
            else:
                return ['New Conversation'] + list(conversations.keys())
        
        with st.sidebar:
            st.session_state.id_chat = st.selectbox(
                label='Choose a Conversation',
                options=options_list(conversations),
                index=0,
                key='chat_id'
            )

            st.session_state.messages = conversations[st.session_state.id_chat] if st.session_state.id_chat != 'New Conversation' else []
            
            def update_selectbox_remove(conversations_path, conversations):
                conversations.pop(st.session_state.chat_id)
                with open(conversations_path, 'wb') as fp:
                    packed_bytes = msgpack.packb(conversations, use_bin_type=True)
                    fp.write(packed_bytes)
                st.session_state.chat_id = 'New Conversation'
            
            st.button(
                'Delete Conversation', 
                use_container_width=True, 
                disabled=False if st.session_state.id_chat != 'New Conversation' else True,
                on_click=update_selectbox_remove,
                args=(conversations_path, conversations)
            )

            st.divider()

            def tooltip_change():
                if st.session_state.search_id == 'Exact Search':
                    st.session_state.tooltip = 'Search the exact definition'
                    st.session_state.search_idx = 0
                elif st.session_state.search_id == 'Documents Only Search':
                    st.session_state.tooltip = 'The AI answer your questions only considering the documents provided'
                    st.session_state.search_idx = 1
                elif st.session_state.search_id == 'Documents + LLM Search':
                    st.session_state.tooltip = """The AI answer your questions considering the documents provided, and if it doesn't found the answer in them, try to find in its own internal knowledge"""
                    st.session_state.search_idx = 2
            
            st.session_state.search_strategy = st.radio(
                label='Choose a Search Strategy',
                options=['Exact Search', 'Documents Only Search', 'Documents + LLM Search'],
                index=st.session_state.search_idx,
                on_change=tooltip_change,
                key='search_id',
                help=st.session_state.tooltip
            )

        def generate_conv_title(llm):
            if st.session_state.chat_id == 'New Conversation':
                output = llm.chat(
                    build_prompt_conv(),
                    SamplingParams(temperature=0.6,top_p=0.9, max_tokens=10, top_k=10)
                )
                print(f'OUTPUT : {output[0].outputs[0].text}')
                st.session_state.chat_id = output[0].outputs[0].text.replace('"', '')
                st.session_state.messages = []

                torch.cuda.empty_cache()
                
                conversations.update({st.session_state.chat_id: st.session_state.messages})
                with open(conversations_path, 'wb') as fp:
                    packed_bytes = msgpack.packb(conversations, use_bin_type=True)
                    fp.write(packed_bytes)
                    
        for message in st.session_state.messages:
            with st.chat_message(message["role"]):
                st.markdown(message["content"])     
                
        if prompt := st.chat_input(
            "Message Video Game Assistant", 
            on_submit=generate_conv_title, 
            key='user_input',
            args=(llm,)
        ):
            st.chat_message("user").markdown(prompt)
            st.session_state.messages.append({"role": "user", "content": prompt})
            
            ai_response = generate_answer(prompt, client, collection_name, model, dense_model, sparse_model, "\n".join([f'{msg["role"]}: {msg["content"]}' for msg in st.session_state.messages]), st.session_state.search_strategy)
            with st.chat_message("assistant"):
                message_placeholder = st.empty()
                full_response = ""
                for chunk in re.split(r'(\s+)', ai_response):
                    full_response += chunk + " "
                    time.sleep(0.05)
                    message_placeholder.write(full_response + '▌')
                message_placeholder.write(re.sub('▌', '', full_response))
                                          
            st.session_state.messages.append({"role": "assistant", "content": full_response})
                
            conversations.update({st.session_state.id_chat: st.session_state.messages})
            with open(conversations_path, 'wb') as fp:
                packed_bytes = msgpack.packb(conversations, use_bin_type=True)
                fp.write(packed_bytes)

        if "cached_files" not in st.session_state:
            st.session_state.cached_files = []

        with st.sidebar:
            st.divider()
        
            uploaded_files = st.file_uploader("Upload a file :", accept_multiple_files=True, type=['pptx', 'ppt'])
            
            for uploaded_file in uploaded_files:

                if uploaded_file not in st.session_state.cached_files:
                    st.session_state.cached_files.append(uploaded_file)
                    
                    file_name = os.path.basename(uploaded_file.name)
                    base_name, ext = os.path.splitext(file_name)
                
                    processing_time = datetime.now().strftime('%d %b %Y, %I:%M %p')

                    full_path = os.path.realpath(uploaded_file.name)
                    file_type = ext.lstrip('.')
                    
                    d = {
                        'icon': icon_to_types[file_type][0],
                        'document': base_name,
                        'type': icon_to_types[file_type][1],
                        'path': full_path,
                        'time': [datetime.strptime(processing_time, '%d %b %Y, %I:%M %p')],
                        'exact_search': False
                    }
                    
                    if (st.session_state.df.empty) or (base_name not in st.session_state.df['document'].tolist()):
                        st.session_state.df = pd.concat(
                            [
                                st.session_state.df, 
                                pd.DataFrame(data={
                                    'icon': icon_to_types[file_type][0],
                                    'document': base_name,
                                    'type': icon_to_types[file_type][1],
                                    'path': full_path,
                                    'time': [datetime.strptime(processing_time, '%d %b %Y, %I:%M %p')],
                                    'exact_search': False
                                })
                            ],
                            ignore_index=True
                        )
                    else:
                        idx = st.session_state.df.index[st.session_state.df['document']==base_name].tolist()[0]
                        st.session_state.df.loc[idx] = {
                            'icon': icon_to_types[file_type][0],
                            'document': base_name,
                            'type': icon_to_types[file_type][1],
                            'path': full_path,
                            'time': datetime.strptime(processing_time, '%d %b %Y, %I:%M %p'),
                            'exact_search': False
                        }

                    st.session_state.df.to_parquet(
                        os.path.join(
                            data_editor_path, 
                            'data_editor.parquet.lz4'
                        ), 
                        compression='lz4', 
                        engine='pyarrow'
                    )
                    
                    documents, ids = ppt_chunker(uploaded_file, llm)

                    dense, tokens_count = asyncio.run(embed_text(dense_model[0], documents))
                    sparse = [s for s in sparse_model.embed(documents, 32)]

                    embeddings_path = os.path.join(os.getenv('HF_HOME'), 'embeddings')

                    def generate_unique_id(existing_ids):
                        while True:
                            new_id = uuid.uuid4()
                            while new_id.hex[0] == '0':
                                new_id = uuid.uuid4()
                            new_id = new_id.int
                            if new_id not in existing_ids:
                                return new_id

                    for filename in os.listdir(embeddings_path):
                        if '_ids' in filename:
                            list_ids = np.load(os.path.join(embeddings_path, filename), allow_pickle=True).tolist()

                            for i, ids_ in enumerate(ids):
                                if ids_ in list_ids:
                                    ids[i] = generate_unique_id(list_ids)

                    metadatas_list = [{'url': full_path}] * len(documents)

                    payload_docs = [{ 'text': documents[i], 'metadata': metadata } for i, metadata in enumerate(metadatas_list)]

                    print(f'LEN : {len(ids)}, {len(payload_docs)}, {len(dense)}, {len([SparseVector(indices=s.indices.tolist(), values=s.values.tolist()) for s in sparse])}')
                    
                    client.upsert(
                        collection_name=collection_name,
                        points=Batch(
                            ids=ids,
                            payloads=payload_docs,
                            vectors={
                                'text-dense': dense,
                                'text-sparse': [SparseVector(indices=s.indices.tolist(), values=s.values.tolist()) for s in sparse]
                            }
                        )
                    )

                    payload_path = os.path.join(embeddings_path, base_name + '_payload.msgpack')
                    dense_path = os.path.join(embeddings_path, base_name + '_dense.npz')
                    sparse_path = os.path.join(embeddings_path, base_name + '_sparse.npz')
                    ids_path = os.path.join(embeddings_path, base_name + '_ids.npy')

                    with open(payload_path, "wb") as outfile_texts:
                        packed_payload = msgpack.packb(payload_docs, use_bin_type=True)
                        outfile_texts.write(packed_payload)
            
                    np.savez_compressed(dense_path, *dense)
                    max_index = 0
                    for embedding in sparse:
                        if len(embedding.indices) > 0:
                            max_index = max(max_index, max(embedding.indices))
        
                    sparse_matrices = []
                    
                    for embedding in sparse:
                        data = np.array(embedding.values)
                        indices = np.array(embedding.indices)
                        indptr = np.array([0, len(data)])
                        matrix = csr_matrix((data, indices, indptr), shape=(1, max_index + 1))
                        sparse_matrices.append(matrix)
            
                    combined_sparse_matrix = vstack(sparse_matrices)
                    save_npz(sparse_path, combined_sparse_matrix)

                    np.save(ids_path, np.array(ids), allow_pickle=True)

                    st.toast('Document(s) Ingested !', icon='🎉')