app_gemma_7b_it.py

import streamlit as st

from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers.utils import is_flash_attn_2_available
from transformers import BitsAndBytesConfig
import pandas as pd
import os
import torch
import numpy as np
from scipy import sparse
from sklearn.metrics.pairwise import cosine_similarity
from scipy import sparse
from langchain_community.vectorstores import Chroma
from langchain_community.embeddings.sentence_transformer import (
    SentenceTransformerEmbeddings,
)

# SET TO WIDE LAYOUT
st.set_page_config(layout="wide")

#_______________________________________________SET VARIABLES_____________________________________________________

MODEL_ID = 'google/gemma-2b-it'
CHUNK_SIZE = 1000
OVERLAP_SIZE = 100
EMBEDDING = "all-MiniLM-L6-v2"
COLLECTION_NAME = f'vb_summarizer_{EMBEDDING}_test'
CHROMA_DATA_PATH = 'feedback_360'

#_______________________________________________LOAD MODELS_____________________________________________________
# LOAD MODEL
@st.cache_resource
def load_model(model_id) :

    HF_TOKEN = os.environ['HF_TOKEN']
    print(torch.backends.mps.is_available())
    #device = torch.device("mps") if torch.backends.mps.is_available() else "cpu"
    device = 'cpu'
    print(device)

    if device=='cpu' :
        print('Warning! No GPU available')

    # IMPORT MODEL
    
    print(model_id)

    quantization_config = BitsAndBytesConfig(load_in_4bit=True,
                                            bnb_4bit_compute_dtype=torch.float16)

    # if (is_flash_attn_2_available()) and (torch.cuda.get_device_capability(0)[0] >= 8):
    #   attn_implementation = "flash_attention_2"
    # else:
    #   attn_implementation = "sdpa"
    # print(f"[INFO] Using attention implementation: {attn_implementation}")

    tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path=model_id, token=HF_TOKEN)

    llm_model = AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path=model_id,
                                                    token=HF_TOKEN,
                                                    torch_dtype=torch.float16,
                                                    #quantization_config=quantization_config if quantization_config else None,
                                                    low_cpu_mem_usage=False,) # use full memory
                                                    #attn_implementation=attn_implementation) # which attention version to use
    llm_model.to(device)
    return llm_model, tokenizer, device

# LOAD VECTORSTORE
@st.cache_resource
def load_data(embedding) :
    # CREATE EMBEDDING
    embedding_function = SentenceTransformerEmbeddings(model_name=embedding)
    db3 = Chroma(collection_name = COLLECTION_NAME, persist_directory="./chroma", embedding_function = embedding_function)
    return db3


# Create a text element and let the reader know the data is loading.
model_load_state = st.text('Loading model...')
# Load 10,000 rows of data into the dataframe.
llm_model, tokenizer, device = load_model(MODEL_ID)
# Notify the reader that the data was successfully loaded.
model_load_state.text('Loading model...done!')

# Create a text element and let the reader know the data is loading.
data_load_state = st.text('Loading data...')
# Load 10,000 rows of data into the dataframe.
vectorstore = load_data(EMBEDDING)
# Notify the reader that the data was successfully loaded.
data_load_state.text('Loading data...done!')


#_______________________________________________SUMMARIZATION_____________________________________________________
# INFERENCE
# def prompt_formatter(reviews, type_of_doc):
#     return f"""You are a summarization bot.
#     You will receive {type_of_doc} and you will extract all relevant information from {type_of_doc} and return one paragraph in which you will summarize what was said.
#     {type_of_doc} are listed below under inputs.
#     Inputs: {reviews}
#     Answer :
#     """
# def prompt_formatter(reviews, type_of_doc):
#     return f"""You are a summarization bot.
#     You will receive {type_of_doc} and you will summarize what was said in the input.
#     {type_of_doc} are listed below under inputs.
#     Inputs: {reviews}
#     Answer :
#     """
def prompt_formatter(reviews):
    return f"""You are a summarization bot.
    You will receive reviews of Clockify from different users.
    You will summarize what these reviews said while keeping the information about each of the user.
    Reviews are listed below.
    Reviews: {reviews}
    Answer :
    """

def mirror_mirror(inputs, prompt_formatter, tokenizer):
    print('Mirror_mirror')
    prompt = prompt_formatter(inputs)
    input_ids = tokenizer(prompt, return_tensors="pt").to(device)
    outputs = llm_model.generate(**input_ids,
                                 temperature=0.3,
                                 do_sample=True,
                                 max_new_tokens=275)
    output_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return prompt, output_text.replace(prompt, '')


def summarization(example : str, results_df : pd.DataFrame = pd.DataFrame()) -> pd.DataFrame :
    
    # INFERENCE
    results = []
    for cnt in range(0,2) : 

        prompt, result = mirror_mirror(example, prompt_formatter, tokenizer)
        list_temp = [result, example]
        tokenized = tokenizer(list_temp, return_tensors="pt", padding = True)
        A = tokenized.input_ids.numpy()
        A = sparse.csr_matrix(A)
        score = cosine_similarity(A)[0,1]
        #print(cosine_similarity(A)[0,1])
        #print(cosine_similarity(A)[1,0])
        print(score)
        if score>0.1 :
            fin_result = result
            max_score = score
            break

        results.append(result)
        #print(result+'\n\n')

    # tokenize results and example together
    try  :
        fin_result 
    except :
    # if fin_result not already defined, use the best of available results
        # add example to results so tokenization is done together (due to padding limitations)
        results.append(example)
        tokenized = tokenizer(results, return_tensors="pt", padding = True)
        A = tokenized.input_ids.numpy()
        A = sparse.csr_matrix(A)
        # calculate cosine similarity of each pair 
        # keep only example X result column
        scores = cosine_similarity(A)[:,2]
        # final result is the one with greaters cos_score
        fin_result = results[np.argmax(scores)]
        max_score = max(scores)

    #print(fin_result)
    # save final result and its attributes
    row = pd.DataFrame({'model' : MODEL_ID, 'prompt' : prompt, 'reviews' : example, 'summarization' : fin_result, 'score' : [max_score] })
    results_df = pd.concat([results_df,row], ignore_index = True)

    return results_df

def create_filter(group:str=None, platform:str=None, ReviewerPosition:str=None, Industry:str=None, CompanySize:str=None,
       UsagePeriod:str=None, LinkedinVerified:str=None, Date:str=None, Rating:str=None) :
    keys = ['group', 'Platform', 'ReviewerPosition', 'Industry', 'CompanySize',
            'UsagePeriod', 'LinkedinVerified', 'Date', 'Rating']
    input_keys = [group,platform, ReviewerPosition, Industry, CompanySize, UsagePeriod, LinkedinVerified, Date, Rating]
    
    # create filter dict 
    filter_dict = {}
    for key, in_key in zip(keys, input_keys) :
        if not in_key == None and not in_key == ' ':
            filter_dict[key] = {'$eq' : in_key}

    print(filter_dict)
    return filter_dict

#_______________________________________________UI_____________________________________________________

st.title("Mirror, mirror, on the cloud, what do Clockify users say aloud?")
st.subheader("--Clockify review summarizer--")

col1, col2, col3 = st.columns(3, gap = 'small')

with col1:
   platform = st.selectbox(label = 'Platform',
             options = [' ', 'Capterra', 'Chrome Extension', 'GetApp', 'AppStore', 'GooglePlay',
                        'Firefox Extension', 'JIRA Plugin', 'Trustpilot', 'G2',
                        'TrustRadius']
             )

with col2:
   company_size = st.selectbox(label = 'Company Size',
             options = [' ', '1-10 employees', 'Self-employed', 'self-employed',
                        'Small-Business(50 or fewer emp.)', '51-200 employees',
                        'Mid-Market(51-1000 emp.)', '11-50 employees',
                        '501-1,000 employees', '10,001+ employees', '201-500 employees',
                        '1,001-5,000 employees', '5,001-10,000 employees',
                        'Enterprise(> 1000 emp.)', 'Unknown', '1001-5000 employees']
             )

with col3:
   linkedin_verified = st.selectbox(label = 'Linkedin Verified',
             options = [' ', 'True', 'False'],
             placeholder = 'Choose an option'
             )

num_to_return = int(st.number_input(label = 'Number of documents to return', min_value = 2, max_value = 50, step = 1))

# group = st.selectbox(label = 'Review Platform Group',
#              options = ['Software Review Platforms', 'Browser Extension Stores', 'Mobile App Stores', 'Plugin Marketplace']
#              )



default_value = "Clockify"

query = st.text_area("Query", default_value, height = 50)
#type_of_doc = st.text_area("Type of text", 'text', height = 25)

# result = ''
# score = ''
# reviews = ''

if 'result' not in st.session_state:
    st.session_state['result'] = ''

if 'score' not in st.session_state:
    st.session_state['score'] = ''

if 'reviews' not in st.session_state:
    st.session_state['reviews'] = ''

col11, col21  = st.columns(2, gap = 'small')

with col11:
   button_query = st.button('Conquer and query!')
with col21:
    button_summarize = st.button('Summon the summarizer!')


if  button_query :
    print('Querying')
    # create filter from drop-downs
    filter_dict = create_filter(#group = group,
                                platform = platform,
                                CompanySize = company_size,
                                LinkedinVerified = linkedin_verified
                    )
    # FILTER BY META
    if filter_dict == {} :
        retriever = vectorstore.as_retriever(search_kwargs = {"k": num_to_return})

    elif len(filter_dict.keys()) == 1 :
        retriever = vectorstore.as_retriever(search_kwargs = {"k": num_to_return,
                                                              "filter":  filter_dict})
    else :
        retriever = vectorstore.as_retriever(search_kwargs = {"k": num_to_return,
                                                            "filter":{'$and': [{key : value} for key,value in filter_dict.items()]}
                                                        }
        )

    reviews = retriever.get_relevant_documents(query = query)
    # only get page content
    st.session_state['reviews'] = [review.page_content for review in reviews]
    print(st.session_state['reviews'])
    result = 'You may summarize now!'

if button_summarize :
    print('Summarization in progress')
    st.session_state['result'] = 'Summarization in progress'
    results_df = summarization("\n".join(st.session_state['reviews']))
    # only one input
    st.session_state['result'] = results_df.summarization[0]
    score = results_df.score[0]


col12, col22  = st.columns(2, gap = 'small')

with col12:
   chosen_reviews = st.text_area("Reviews to be summarized", "\n".join(st.session_state['reviews']), height = 275)
with col22:
    summarized_text = st.text_area("Summarized text", st.session_state['result'], height = 275)

score = st.text_area("Cosine similarity score", st.session_state['score'], height = 25)





# max_length = st.sidebar.slider("Max Length", min_value = 10, max_value=30)
# temperature = st.sidebar.slider("Temperature", value = 1.0, min_value = 0.0, max_value=1.0, step=0.05)
# top_k = st.sidebar.slider("Top-k", min_value = 0, max_value=5, value = 0)
# top_p = st.sidebar.slider("Top-p", min_value = 0.0, max_value=1.0, step = 0.05, value = 0.9)
# num_return_sequences = st.sidebar.number_input('Number of Return Sequences', min_value=1, max_value=5, value=1, step=1)s