app.py

from llama_index import GPTVectorStoreIndex, SimpleDirectoryReader, ServiceContext, set_global_service_context, load_index_from_storage, StorageContext, PromptHelper
from llama_index.llms import OpenAI
from llama_index.evaluation import ResponseEvaluator
from langchain.chat_models import ChatOpenAI
from PyPDF2 import PdfReader
import gradio as gr
import sys
import os

try:
    from config import OPEN_AI_KEY
    os.environ["OPENAI_API_KEY"] = OPEN_AI_KEY
except:
    pass

# ===============================
# Settings
# ===============================

MAX_INPUT_SIZE = 4096
NUM_OUTPUT = 2048
CHUNK_OVERLAP_RATIO = 0.15
CHUNK_SIZE_LIMIT = 1000
TEMPERATURE = 0.5
DIRECTORY = 'dww_rev_cleaned'
DIRECTORY_PERSIST = 'dww_rev_cleaned_persist'

# Define LLM: gpt-3.5-turbo, temp:0.7
llm = OpenAI(model="gpt-3.5-turbo", temperature=TEMPERATURE, max_tokens=NUM_OUTPUT)

# Define prompt helper
prompt_helper = PromptHelper(context_window=MAX_INPUT_SIZE, num_output=NUM_OUTPUT, chunk_overlap_ratio=CHUNK_OVERLAP_RATIO, chunk_size_limit=CHUNK_SIZE_LIMIT)

# Set service context
service_context = ServiceContext.from_defaults(llm=llm, prompt_helper=prompt_helper)
set_global_service_context(service_context)


# ===============================
# Functions
# ===============================

def construct_index(directory_path, index_path):
    
    if os.listdir(index_path) != []:
        storage_context = StorageContext.from_defaults(persist_dir=index_path)
        index = load_index_from_storage(storage_context)
        return index

    else:
        # Load in documents
        documents = SimpleDirectoryReader(directory_path).load_data()

        # Index documents
        index = GPTVectorStoreIndex.from_documents(documents, service_context=service_context, show_progress=True)

        # Save index
        index.storage_context.persist(persist_dir=index_path)

        return index

INDEX = construct_index(DIRECTORY, DIRECTORY_PERSIST)
QE = INDEX.as_query_engine()
PDF_CONTENT = gr.State("")

def upload_file(file):
    try:
        read_pdf = PdfReader(file.name)
        pdf_text = "\n\n".join([w.extract_text() for w in read_pdf.pages])
        PDF_CONTENT.value = pdf_text
        return pdf_text
    except Exception as e:
        return f"Error: {str(e)}"

def chatfunc(input_text, chat_history, max_chat_length=4):
    prompt = """You are an insight bot that helps users (special educators and school psychologists) build individual education programs based on disability categories using DWW (a library of research-backed interviews and tools) as reference. 
                Refer to the DWW's context as much as you can to provide a detailed answer."""
    if PDF_CONTENT.value:
        prompt = prompt + "The following is the relevant document provided by the user" + PDF_CONTENT.value + "\n\n"
    for chat in chat_history[~max_chat_length:]:
        user_chat, bot_chat = chat
        prompt = f"{prompt}\nUser: {user_chat}\nAssistant: {bot_chat}"
    prompt = f"{prompt}\nUser: {input_text}\nAssistant:"
    response = QE.query(prompt)
    chat_history.append([input_text, response.response])
    return "", chat_history

with gr.Blocks() as iface:
    chatbot = gr.Chatbot(height=400)
    msg = gr.Textbox(label="Ask the DWW Bot anything about research-based practices in education")
    submit = gr.Button("💬 Submit")
    with gr.Row():
        clear = gr.ClearButton(value="🧹 Clear outputs", components=[msg, chatbot])
        upload_button = gr.UploadButton("📁 Upload a relevant document", file_types=[".pdf"], file_count="single")
    with gr.Accordion("📝 View your document"):
        syl = gr.Textbox(label="Your documents' content will show here")
    msg.submit(chatfunc, [msg, chatbot], [msg, chatbot])
    upload_button.upload(upload_file, upload_button, syl)  

iface.launch(share=False)