app.py

import os
import string
from typing import Any, Dict, List, Tuple, Union

import chromadb
import numpy as np
import openai
import pandas as pd
import requests
import streamlit as st
from datasets import load_dataset
from langchain.document_loaders import TextLoader
from langchain.embeddings.sentence_transformer import SentenceTransformerEmbeddings
from langchain.text_splitter import CharacterTextSplitter
from langchain.vectorstores import Chroma
from scipy.spatial.distance import cosine

from utils.helper_functions import *

openai.api_key = os.environ["OPENAI_API_KEY"]


# Load the dataset from a provided source.
dataset = load_dataset(
    "eagle0504/youthless-homeless-shelter-web-scrape-dataset-qa-formatted"
)

# Initialize a new client for ChromeDB.
client = chromadb.Client()

# Generate a random number between 1 billion and 10 billion.
random_number: int = np.random.randint(low=1e9, high=1e10)

# Generate a random string consisting of 10 uppercase letters and digits.
random_string: str = "".join(
    np.random.choice(list(string.ascii_uppercase + string.digits), size=10)
)

# Combine the random number and random string into one identifier.
combined_string: str = f"{random_number}{random_string}"

# Create a new collection in ChromeDB with the combined string as its name.
collection = client.create_collection(combined_string)


# Embed and store the first N supports for this demo
L = len(dataset["train"]["questions"])
collection.add(
    ids=[str(i) for i in range(0, L)],  # IDs are just strings
    documents=dataset["train"]["questions"],  # Enter questions here
    metadatas=[{"type": "support"} for _ in range(0, L)],
)

# Front-end Design
st.title("Youth Homelessness Chatbot")

# Initialize chat history
if "messages" not in st.session_state:
    st.session_state.messages = []

# Display chat messages from history on app rerun
for message in st.session_state.messages:
    with st.chat_message(message["role"]):
        st.markdown(message["content"])

st.sidebar.markdown(
    """
    
    ### Instructions:
    
    This app guides you through YSA's website, utilizing a RAG-ready Q&A dataset [here](https://huggingface.co/datasets/eagle0504/youthless-homeless-shelter-web-scrape-dataset-qa-formatted) for chatbot assistance. 🤖 Enter a question, and it finds similar ones in the database, offering answers with a distance score to gauge relevance—the lower the score, the closer the match. 🎯 For better accuracy and to reduce errors, user feedback helps refine the database. ✨
    
    """
)
st.sidebar.success(
    "Please enter a distance threshold (we advise to set it around 0.2)."
)
special_threshold = st.sidebar.number_input(
    "Insert a number", value=0.2, placeholder="Type a number..."
)  # 0.3
clear_button = st.sidebar.button("Clear Conversation", key="clear")
st.sidebar.warning(
    "The 'distances' measures how close your question is to the questions in our database (lower the score the better). The 'ai_judge' measures independent similarity ranking of database answers and user's question (the higher the better)."
)

if clear_button:
    st.session_state.messages = []

# React to user input
if prompt := st.chat_input("Tell me about YSA"):
    # Display user message in chat message container
    st.chat_message("user").markdown(prompt)
    # Add user message to chat history
    st.session_state.messages.append({"role": "user", "content": prompt})

    question = prompt
    with st.spinner("Wait for it..."):
        results = collection.query(query_texts=question, n_results=5)
        idx = results["ids"][0]
        idx = [int(i) for i in idx]
        ref = pd.DataFrame(
            {
                "idx": idx,
                "questions": [dataset["train"]["questions"][i] for i in idx],
                "answers": [dataset["train"]["answers"][i] for i in idx],
                "distances": results["distances"][0],
            }
        )
        # special_threshold = st.sidebar.slider('How old are you?', 0, 0.6, 0.1) # 0.3
        filtered_ref = ref[ref["distances"] < special_threshold]
        if filtered_ref.shape[0] > 0:
            st.success("There are highly relevant information in our database.")
            ref_from_db_search = filtered_ref["answers"].str.cat(sep=" ")
            final_ref = filtered_ref
        else:
            st.warning(
                "The database may not have relevant information to help your question so please be aware of hallucinations."
            )
            ref_from_db_search = ref["answers"].str.cat(sep=" ")
            final_ref = ref

        try:
            llm_response = llama2_7b_ysa(question)
        except:
            llm_response = "Sorry, the inference endpoint is temporarily down. 😔"

        finetuned_llm_guess = ["from_llm", question, llm_response, 0]
        final_ref.loc[-1] = finetuned_llm_guess
        final_ref.index = final_ref.index + 1

        # add ai judge as additional rating
        independent_ai_judge_score = []
        for i in range(final_ref.shape[0]):
            this_content = final_ref["answers"][i]
            this_score = calculate_sts_openai_score(question, this_content)
            independent_ai_judge_score.append(this_score)

        final_ref["ai_judge"] = independent_ai_judge_score

        engineered_prompt = f"""
            Based on the context: {ref_from_db_search}

            answer the user question: {question}

            Answer the question directly (don't say "based on the context, ...")
        """

        answer = call_chatgpt(engineered_prompt)
        response = answer

    # Display assistant response in chat message container
    with st.chat_message("assistant"):
        with st.spinner("Wait for it..."):
            st.markdown(response)
            with st.expander("See reference:"):
                st.table(final_ref)
    # Add assistant response to chat history
    st.session_state.messages.append({"role": "assistant", "content": response})