app.py

import streamlit as st
from datetime import time as t
import time

from operator import itemgetter  
import os
import json
import getpass
import openai
  
from langchain.vectorstores import Pinecone
from langchain.embeddings import OpenAIEmbeddings  
import pinecone


from results import results_agent
from filter import filter_agent
from reranker import reranker
from utils import build_filter
from router import routing_agent

OPENAI_API = st.secrets["OPENAI_API"]
PINECONE_API = st.secrets["PINECONE_API"]
openai.api_key = OPENAI_API


pinecone.init(
    api_key= PINECONE_API,
    environment="gcp-starter" 
)
index_name = "use-class-db"

embeddings = OpenAIEmbeddings(openai_api_key = OPENAI_API)

index = pinecone.Index(index_name)

k = 5

st.title("USC GPT - Find the perfect class")

class_time = st.slider(
    "Filter Class Times:",
    value=(t(11, 30), t(12, 45)))

units = st.slider(
    "Number of units",
    1, 4,
    value = (1, 4)
)

days = st.checkbox("What days are you free?",
               options = ["Mon", "Tue", "Wed", "Thu", "Fri", "Sat"],
               default = None,
               placeholder = "Any day"
                )

assistant = st.chat_message("assistant")
initial_message = "How can I help you today?"

def get_rag_results(prompt):
    '''
1. Remove filters from the prompt to optimize success of the RAG-based step.
2. Query the Pinecone DB and return the top 25 results based on cosine similarity
3. Rerank the results from vector DB using a BERT-based cross encoder
    '''
    query = filter_agent(prompt, OPENAI_API)

  ##Get metadata filters  
    days_filter = list()
    query_filter = {
        "start": {"$gte": str(class_time[0])},
        "end": {"$lte": str(class_time[1])}
    }

    if units != "any":
        query_filter["units"] = str(int(units)) + ".0 units"

    if len(days) > 0:
        for i in range(len(days)):
            days_filter.append(days[i])
            for j in range(i+1, len(days)):
                two_day = days[i] + ", " +  days[j]
                days_filter.append(two_day)
        query_filter["days"] = {"$in": days_filter}

  ## Query the pinecone database
    response = index.query(
        vector = embeddings.embed_query(query),
        top_k = 25,
        filter = query_filter,
        include_metadata = True
    )
    response = reranker(query, response) # BERT cross encoder for ranking

    return response

    

if "messages" not in st.session_state:
    st.session_state.messages = []
    with st.chat_message("assistant"):
        st.markdown(initial_message)
    st.session_state.messages.append({"role": "assistant", "content": initial_message})
    

if prompt := st.chat_input("What kind of class are you looking for?"):
    st.session_state.messages.append({"role": "user", "content": prompt})
    with st.chat_message("user"):
            st.markdown(prompt)

    with st.chat_message("assistant"):
        message_placeholder = st.empty()
        full_response = ""

        messages = [{"role": m["role"], "content": m["content"]}
                for m in st.session_state.messages]
        message_history = " ".join([message["content"] for message in messages])

        route = routing_agent(prompt, OPENAI_API, message_history)

        if route == "1":
        ## Option for accessing Vector DB
            rag_response = get_rag_results(prompt)
            result_query = 'Original Query:' + prompt + 'Query Results:' + str(rag_response)
            assistant_response = results_agent(result_query, OPENAI_API)
        else:    
        ## Option if not accessing Database
            assistant_response = openai.ChatCompletion.create(
                model = "gpt-4",
                messages = [
                    {"role": m["role"], "content": m["content"]}
                    for m in st.session_state.messages
                ]
            )["choices"][0]["message"]["content"]

        ## Display response regardless of route
        for chunk in assistant_response.split():
            full_response += chunk + " "
            time.sleep(0.05)
            message_placeholder.markdown(full_response + "▌")
        message_placeholder.markdown(full_response)
        st.session_state.messages.append({"role": "assistant", "content": full_response})