app.py

import streamlit as st

#from transformers import pipeline
from pinecone import Pinecone, ServerlessSpec
from sentence_transformers import SentenceTransformer, util
from openai import OpenAI
from datetime import datetime
import pandas as pd
import numpy as np
import os
import time
import json

from huggingface_hub import HfFileSystem
token = os.getenv('FILE_TOKEN')
fs = HfFileSystem(token=token)




api_key=os.environ["OPENAI_API_KEY"]


bi_encoder = SentenceTransformer('msmarco-distilbert-base-v4')
bi_encoder.max_seq_length = 256     # Truncate long documents to 256 tokens

# Store the index as a variable
INDEX_NAME = 'cl-search-idx'
INDEX_NAME = 'cl-kb'

pc_api_key= os.environ["clpine"]  #AWS
pc = Pinecone(api_key=pc_api_key)
index = pc.Index(name=INDEX_NAME)

system_instructions_text='''
          Your task is to extract the answer to a question from a body of text provided to you.
          The body of text will be enclosed within the delimiter tags  <text> and </text>
          For example,
          <text> General Preparation Tips for VARC Section:
          You need to develop an incessant habit of speed reading.
          Start with reading newspapers, editorials, fiction and nonfiction novels and simple passages.
          The more you read, the faster you read. Learn the basic grammar concepts like parts of speech, articles,verbs, adjectives, tenses, auxiliary verbs, modifiers, modals etc.
          Revise at least 50 new words every day
          </text>
          Question: What are some tips for preparing for VARC?
          Here are some tips for preparing for the VARC section:
          1. develop an incessant habit of speed reading
          2. Start reading newspapers, editorials, fiction and nonfiction novels
          3. Learn basic grammar concepts\n
          4. Revise at least 50 new words a day
          Question: How many new words are to be learnt in a day?
          It is advised that 50 new words are learn every day
          Your response should be based on the information contained in the provided text and should not included any other sources.
          If you are unable to answer the question from the text provided, please respond " Sorry. I do not have enough information to answer this"
           Do repeat the question. Do not make a pointed reference to the text provided. Directly answer the question
          '''
json_instructions='''
          Your task is to extract the answer to a question from a body of text provided to you in a json array.
          The json will contain two pieces of content in this format:
          [
              {"id":1 , "content": " first content"},
              {"id":2 , "content": " second content"}
          ]
          You need to check which content is most appropriate to answer the question and prepare
          an answer based on the content
          For example,
          [
          { "id":1 , "content" : "General Preparation Tips for Verbal Section:\n
          You need to develop an incessant habit of speed reading.
          Start with reading newspapers, editorials, fiction and nonfiction novels and simple passages.
          The more you read, the faster you read. Learn the basic grammar concepts like parts of speech, articles,verbs, adjectives, tenses, auxiliary verbs, modifiers, modals etc.
          Revise at least 50 new words every day"},
          { "id":2 , "content" : "General Preparation Tips for Quantitative Section:\n
          You need to develop an speed in solving math problems.
          Start with reading funda books,  math text books.
          Learn the basic  concepts like arithmetic, geometry, numbers, probability,  etc.
          Solve at least 50 new problems every day"}
          ]
          Question: What are some tips for preparing for Verbal exam?
          Here are some tips for preparing for the VARC section:
          1. develop an incessant habit of speed reading
          2. Start reading newspapers, editorials, fiction and nonfiction novels
          3. Learn basic grammar concepts\n
          4. Revise at least 50 new words a day
          Do repeat the question. Do not make a pointed reference to the content provided. Directly answer the question.
          Your response should be based on the information contained in the provided content in the json and should not included any other sources.
          If you are unable to answer the question from the content provided, please respond " Sorry. I do not have enough information to answer this"
           
          '''

def get_meta_score(url,question_embedding):
    qry = index.fetch(ids=[url], namespace="meta")

    emb=qry['vectors'][url]['values']
    vector1 = np.array(emb).reshape(1, -1)  # Reshape to ensure compatibility with sklearn
    vector2 = question_embedding.numpy().reshape(1, -1)

    # Calculate cosine similarity
    cosine_scores = util.cos_sim(question_embedding, emb)
    return  cosine_scores.item()


def query_from_pinecone(index,namespace, question_embedding, top_k=3):
    # get embedding from THE SAME embedder as the documents

    return index.query(
      vector=question_embedding,
      top_k=top_k,
      namespace=namespace,
      include_metadata=True   # gets the metadata (dates, text, etc)
    ).get('matches')

def response_generator(response):

    for word in response.split():
        yield word + " "
        time.sleep(0.05)

def write_log(query,url, score, ans, ts):
    # Construct new row
    score = str(score)
    ans = ans.replace('"', '""')
    new_row = f'\n"{query}","{url}",{score},"{ans}","{ts}"'

    with fs.open("datasets/sujitb/data/querylog.csv", "r") as f:
        buffer = f.read()

    # Append the new row to buffer
    buffer += new_row

    # Write the buffer to the file in "W" mode
    with fs.open("datasets/sujitb/data/querylog.csv", "w",encoding="utf-8") as f:
        f.write(buffer)

    return

logfile='querylog.csv'

qlist=[
'What are the best books for VARC in CAT?',
'What is the XAT exam pattern? How many sections ? How many questions are asked in each section?',
'I want to know about Personalized coaching for IGSE/IB',
'Which IIMs accept admissions under the IPM exam?',
'What topics are covered under CAT exam syllabus?',
'What is the pattern of the IPM exam?',
'Which Central Universities offer courses for CUET exam',
'For CAT preparation which is better - online classes or classroom program?',
'What programs are offered under CUET exam  by Central University of Jharkhand?',
'What is the pattern of the IPM exam?',
'When is the CAT 2024 exam going to be held?',
'What are program benefits of the MBA 2024 Online Classes?',
'What topics are covered in CUET General Test?',
'IIM A B C vs FMS - how to select the best bschool?'
]

try:
    df_log = pd.read_csv("hf://datasets/sujitb/data/querylog.csv", encoding="utf-8", index_col=0)
except:
    df_log=pd.DataFrame(columns=['query','url','score','ans', 'ts'])

#st.title('CLLM Answering Machine')
st.subheader('CLLM Answering Machine', divider='rainbow')

with st.sidebar:
    dispstr= 'Search History'
    st.markdown('*{}*'.format(dispstr))
    #st.write('Past Queries')
    #qlist = df_log.tail(30)['query'].tolist()
    for q in qlist[::-1]:
        st.write(q)

# 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"])

#with st.chat_message("user"):
#    st.write("Hello 👋  Ask any question related to careerlauncher.com in the text box below")

QUESTION = st.chat_input('Ask a question -e.g How to prepare for Verbal section for CAT?')


#QUESTION=st.text_area('Ask a question -e.g How to prepare for Verbal section for CAT?')  ##' How to prepare for Verbal section ?'
score=0
testing=True


if QUESTION:

    with st.chat_message("user"):
        st.markdown(QUESTION)
    # Add user message to chat history
    st.session_state.messages.append({"role": "user", "content": QUESTION})

    st.write('Searching knowledge base...')
    question_embedding = bi_encoder.encode(QUESTION, convert_to_tensor=True)

    THRESHOLD=.4
    ns='webpages'
    #ns='full'
    resp= query_from_pinecone(index,ns, question_embedding.tolist(), 10)
    resplist=[]
    id=0
    for r in resp:
        id+=1
        d={}
        d['id']=id
        d['content']=r['metadata']['data']
        d['title']=r['metadata']['title']
        d['url']=r['id']
        if ns=='webpages':
            d['url']= r['metadata']['url']
            
        meta_score= get_meta_score(d['url'],question_embedding)
        
        score=.5* r['score'] + .5*meta_score
        d['score']=score
        #st.write(d['url'], score, r['score'], meta_score)
        resplist.append(d)

    #check youtube
    respmeta= query_from_pinecone(index, 'meta', question_embedding.tolist(), 5)
    for r in respmeta:
        if 'youtube' in r['id'] and r['score']>=.8:
            d['id']=r['id']
            d['content']=r['metadata']['data']
            d['title']=r['metadata']['title']
            d['url']= r['metadata']['url']   
            d['score']=r['score']
            resplist.append(d)

    if len(resplist)>0:
        sorted_indices = sorted(range(len(resplist)), key=lambda i: resplist[i]['score'], reverse=True)

        # Get the elements with the top 2 highest values
        top_2 =  [resplist[i] for i in sorted_indices[:2]]

        # covert to array

        json_data = json.dumps(top_2)


        goodmatch=False
        if resplist[sorted_indices[0]]['score']>=THRESHOLD:
            st.write('Preparing answers...')
            goodmatch=True
            mode = "two"  # two passages

            client = OpenAI()

            if mode=="one":
                instr=system_instructions_text

                out= resplist[sorted_indices[0]]['content']
                content="""
                <text>
                {}
                </text>
                """.format(out)

            if mode=="two":
                instr=json_instructions
                content=json_data

            response = client.chat.completions.create(
              model="gpt-3.5-turbo",
              messages=[
                {"role": "system", "content":instr },
                {"role": "user", "content": content},
                {"role": "user", "content": "Question:"+QUESTION}
              ]
            )

            ans= response.choices[0].message.content
        else:
            ans='Weak match to your query. Please try reframing your question'

            #st.write("Matched URL:{}  Score:{}".format(url,score))
            testing = False
            if testing:
                if len(resp)>=1:
                    st.write("2nd Matched URL:{}  Score:{}".format(resp[1]['id'],resp[1]['score']))
                if len(resp)>=2:
                    st.write("3rd Matched URL:{}  Score:{}".format(resp[2]['id'],resp[2]['score']))
        ##  Send RESPONSE
        with st.chat_message("assistant"):
            response = st.write_stream(response_generator(ans))
            if goodmatch:
                st.write('Resources:')
                for k in range(2):
                    disp_title=top_2[k]['title']
                    disp_url= top_2[k]['url']
                    if 'youtube' in disp_url:
                        disp_title='Youtube: '+disp_title

                    if k>0:
                        if top_2[k]['url']==top_2[k-1]['url']:
                            break
                    st.write("["+disp_title+"]("+disp_url+")")

            else:
                st.write('Resources:',top_2[0]['score'], ' confidence')
                st.write(top_2[0]['url'])                
            # Add assistant response to chat history
            st.session_state.messages.append({"role": "assistant", "content": response})
        #st.write(ans)

        #st.write(' ----------------------')
        #st.write(out)

        now= str(datetime.utcnow())
        url = top_2[0]['url'] + ' ; '+top_2[1]['url']

        df_log.loc[len(df_log)]=[QUESTION,url,score,ans,now]
        #write_log(QUESTION,url, score, ans, now)
        #df.to_csv("hf://datasets/sujitb/data/test.csv")
        storage_options={"token":token}
        df_log.to_csv("hf://datasets/sujitb/data/"+logfile,storage_options= storage_options)

    else:  ## Zero response from pinecone query
        #st.write("No matches for query")
        ans= "No matches for query"
        response = st.write_stream(response_generator(ans))
        # Add assistant response to chat history
        st.session_state.messages.append({"role": "assistant", "content": response})

        now= str(datetime.utcnow())
        df_log.loc[len(df_log)]=[QUESTION,'No match',0,'-',now]
        storage_options={"token":token}
        df_log.to_csv("hf://datasets/sujitb/data/"+logfile,storage_options= storage_options)
        #write_log(QUESTION,'No match', 0, '-', now)