app.py

import gradio as gr
from dotenv import load_dotenv
from langchain import PromptTemplate, LLMChain, HuggingFaceHub
from langchain.llms import CTransformers
from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers import pipeline
from langchain.llms.huggingface_pipeline import HuggingFacePipeline
from sentence_transformers import SentenceTransformer, util
from sklearn.cluster import KMeans
import nltk
nltk.download('punkt')
from nltk.tokenize import word_tokenize
from nltk import tokenize
import numpy as np
import scipy.spatial
import csv


load_dotenv()

def generate_prompts(user_input):
    prompt_template = PromptTemplate(
        input_variables=["Question"],
        template=f"Just list 10 question prompts for {user_input} and don't put number before each of the prompts."
    )
    config = {'max_new_tokens': 64, 'temperature': 0.7, 'context_length': 64}
    llm = CTransformers(model="TheBloke/Mistral-7B-Instruct-v0.1-GGUF",
                        config=config)
    hub_chain = LLMChain(prompt = prompt_template, llm = llm)
    input_data = {"Question": user_input}

    generated_prompts = hub_chain.run(input_data)  
    questions_list = generated_prompts.split('\n') 
    

    formatted_questions = "\n".join(f"Question: {question}" for i, question in enumerate(questions_list) if question.strip())
    questions_list = formatted_questions.split("Question:")[1:]
    return questions_list

def answer_question(prompt):
    prompt_template = PromptTemplate(
        input_variables=["Question"],
        template=f"give one answer for {prompt} and do not consider the number behind it."
    )
    config = {'max_new_tokens': 64, 'temperature': 0.7, 'context_length': 64}
    llm = CTransformers(model="TheBloke/Llama-2-7B-Chat-GGML",
                        config=config)
    hub_chain = LLMChain(prompt = prompt_template, llm = llm)
    input_data = {"Question": prompt}
    generated_answer = hub_chain.run(input_data)  
    return generated_answer

def calculate_similarity(word, other_words, model, threshold=0.5):
    embeddings_word = model.encode([word])
    embeddings_other_words = model.encode(other_words)
    for i, embedding in enumerate(embeddings_other_words):
        similarity = 1 - scipy.spatial.distance.cosine(embeddings_word[0], embedding)
        if similarity > threshold and similarity < 0.85:
            return i, similarity
    return None, None


def highlight_similar_paragraphs_with_colors(paragraphs, similarity_threshold=0.75):
    model = SentenceTransformer('all-MiniLM-L6-v2')
    
    # Split each paragraph into sentences
    all_sentences = [tokenize.sent_tokenize(paragraph) for paragraph in paragraphs]
    
    # Initialize storage for highlighted sentences
    highlighted_sentences = [['' for sentence in para] for para in all_sentences]
    colors = ['yellow', 'lightgreen', 'lightblue', 'pink', 'lavender', 'salmon', 'peachpuff', 'powderblue', 'khaki', 'wheat']
    
    # Track which sentences belong to which paragraph
    sentence_to_paragraph_index = [idx for idx, para in enumerate(all_sentences) for sentence in para]
    
    # Encode all sentences into vectors
    flattened_sentences = [sentence for para in all_sentences for sentence in para]
    sentence_embeddings = model.encode(flattened_sentences)
    
    # Calculate cosine similarities between all pairs of sentences
    cosine_similarities = util.pytorch_cos_sim(sentence_embeddings, sentence_embeddings)
    
    # Iterate through each sentence pair and highlight if they are similar but from different paragraphs
    color_index = 0
    for i, embedding_i in enumerate(sentence_embeddings):
        for j, embedding_j in enumerate(sentence_embeddings):
            if i != j and cosine_similarities[i, j] > similarity_threshold and sentence_to_paragraph_index[i] != sentence_to_paragraph_index[j]:
                color = colors[color_index % len(colors)]
                if highlighted_sentences[sentence_to_paragraph_index[i]][i % len(all_sentences[sentence_to_paragraph_index[i]])] == '':
                    highlighted_sentences[sentence_to_paragraph_index[i]][i % len(all_sentences[sentence_to_paragraph_index[i]])] = ("<span style='color: "+  color  +"'>"+ flattened_sentences[i]+"</span>")
                if highlighted_sentences[sentence_to_paragraph_index[j]][j % len(all_sentences[sentence_to_paragraph_index[j]])] == '':
                    highlighted_sentences[sentence_to_paragraph_index[j]][j % len(all_sentences[sentence_to_paragraph_index[j]])] = ("<span style='color: "+  color  +"'>"+ flattened_sentences[j]+"</span>")
                color_index += 1  # Move to the next color

    # Combine sentences back into paragraphs
    highlighted_paragraphs = [' '.join(para) for para in highlighted_sentences]

    # Combine all paragraphs into one HTML string
    html_output = '<div>' + '<br/><br/>'.join(highlighted_paragraphs) + '</div>'
    return highlighted_paragraphs

    
def calculate_similarity_score(sentences):
    # Encode all sentences to get their embeddings
    model = SentenceTransformer('all-MiniLM-L6-v2')
    embeddings = model.encode(sentences)

    # Calculate average cosine similarity
    total_similarity = 0
    comparisons = 0
    for i in range(len(embeddings)):
        for j in range(i+1, len(embeddings)):
            # Cosine similarity between embeddings
            similarity = 1 - cosine(embeddings[i], embeddings[j])
            total_similarity += similarity
            comparisons += 1

    # Average similarity
    average_similarity = total_similarity / comparisons if comparisons > 0 else 0

    # Scale from [-1, 1] to [0, 100]
    score_out_of_100 = (average_similarity + 1) / 2 * 100
    return score_out_of_100

def answer_question(prompt):
    prompt_template = PromptTemplate.from_template(
        input_variables=["Question"],
        template=f"give one answer for {prompt} and do not consider the number behind it."
    )
    config = {'max_new_tokens': 2048, 'temperature': 0.7, 'context_length': 4096}
    llm = CTransformers(model="TheBloke/Llama-2-7B-Chat-GGML",
                        config=config,
                        threads=os.cpu_count())
    hub_chain = LLMChain(prompt = prompt_template, llm = llm)
    input_data = {"Question": prompt}
    generated_answer = hub_chain.run(input_data)  
    return generated_answer


def process_inputs(llm, questions, relevance, diversity, email):
    # Save questions to a CSV file
    questions_list = questions.split('\n')
    df = pd.DataFrame(questions_list, columns=["Questions"])
    csv_file = "/mnt/data/questions.csv"
    df.to_csv(csv_file, index=False)

    # Email the CSV file
    sender_email = "[email protected]"
    sender_password = "opri fcxx crkh bvfj"
    receiver_email = email
    subject = "Your Submitted Questions"
    body = "Thank you for your submission. Please find attached the CSV file containing your questions."

    message = MIMEMultipart()
    message['From'] = sender_email
    message['To'] = receiver_email
    message['Subject'] = subject
    message.attach(MIMEText(body, 'plain'))

    attachment = open(csv_file, "rb")
    part = MIMEBase('application', 'octet-stream')
    part.set_payload((attachment).read())
    encoders.encode_base64(part)
    part.add_header('Content-Disposition', f"attachment; filename= questions.csv")
    message.attach(part)

    server = smtplib.SMTP('smtp.gmail.com', 587)
    server.starttls()
    server.login(sender_email, sender_password)
    text = message.as_string()
    server.sendmail(sender_email, receiver_email, text)
    server.quit()

    return f"Submitted questions:\n\n{questions}\n\nRelevance: {relevance}\nDiversity: {diversity}\nEmail: {email}"

    
text_list = []

def updateChoices(prompt):
    newChoices = generate_prompts(prompt)
    return gr.CheckboxGroup(choices=newChoices)

def setTextVisibility(cbg, model_name_input):

    sentences = [answer_question(text, model_name_input) for text in cbg]
    
    # Apply highlighting to all processed sentences, receiving one complete HTML string.
    highlighted_html = []
    highlighted_html = highlight_similar_paragraphs_with_colors(sentences, similarity_threshold=0.75)
    
    
    result = []
    # Iterate through each original 'cbg' sentence and pair it with the entire highlighted block.
    for idx, sentence in enumerate(highlighted_html):
        result.append("<p><strong>"+ cbg[idx] +"</strong></p><p>"+ sentence +"</p><br/>")

    score = round(calculate_similarity_score(highlighted_html))

    final_html = f"""<div>{result}<div style="text-align: center; font-size: 24px; font-weight: bold;">Similarity Score: {score}</div></div>"""


    return final_html
    

    # update_show = [gr.Textbox(visible=True, label=text, value=answer_question(text, model_name_input)) for text in cbg]
    # update_hide = [gr.Textbox(visible=False, label="") for _ in range(10-len(cbg))]
    # return update_show + update_hide

with gr.Blocks(theme=gr.themes.Soft()) as demo:

    gr.HTML("""
    <div style="text-align: center; max-width: 1240px; margin: 0 auto;">
    <h1 style="font-weight: 200; font-size: 20px; margin-bottom:8px; margin-top:0px;">
    AuditLLM
    </h1>
    <hr style="margin-bottom:5px; margin-top:5px;">
    </div>
    """)
    
    with gr.Tab("Live Mode"):
        gr.HTML("""
        <div>
        <h4> Live Mode Auditing LLMs <h4>
        <div>
        <div style = "font-size: 13px;">
        <p><In Live Auditing Mode, you gain the ability to probe the LLM directly./p>

        <p>First, select the LLM you wish to audit. Then, enter your question. The AuditLLM tool will generate five relevant and diverse prompts based on your question. You can now select these prompts for auditing the LLMs. Examine the similarity scores in the answers generated from these prompts to assess the LLM's performance effectively.</p>   
                  
        </div>
        """)
        with gr.Row():
            model_name_input = gr.Dropdown([("Llama", "TheBloke/Llama-2-7B-Chat-GGML"), ("Falcon", "TheBloke/Falcon-180B-Chat-GGUF"), ("Zephyr", "TheBloke/zephyr-quiklang-3b-4K-GGUF"),("Vicuna", "TheBloke/vicuna-33B-GGUF"),("Claude","TheBloke/claude2-alpaca-13B-GGUF"),("Alpaca","TheBloke/LeoScorpius-GreenNode-Alpaca-7B-v1-GGUF")], label="Large Language Model")
        with gr.Row():
            prompt_input = gr.Textbox(label="Enter your question", placeholder="Enter Your Question")
        with gr.Row():
            generate_button = gr.Button("Generate", variant="primary", min_width=300)
        with gr.Column():
            cbg = gr.CheckboxGroup(choices=[], label="List of the prompts", interactive=True)
        
        generate_button.click(updateChoices, inputs=[prompt_input], outputs=[cbg])

        with gr.Row() as exec: 
            btnExec = gr.Button("Execute", variant="primary", min_width=200)


        with gr.Column() as texts:
            for i in range(10):
                text = gr.Textbox(label="_", visible=False)
                text_list.append(text)

        with gr.Column():
            html_result = gr.HTML("""<div style="color: red"></div>""")

        #btnExec.click(setTextVisibility, inputs=[cbg, model_name_input], outputs=text_list)
        btnExec.click(setTextVisibility, inputs=[cbg, model_name_input], outputs=html_result)
        gr.HTML("""
        <div style="text-align: center; font-size: 24px; font-weight: bold;">Similarity Score: </div>
                """)

        clear = gr.ClearButton(link = "http://127.0.0.1:7865") 

    with gr.Tab("Batch Mode"):
        gr.HTML("""
        <div>
        <h4> Batch Mode Auditing LLMs <h4>
        <div>
        <div style = "font-size: 13px;">
        <p>In batch auditing mode, you have the capability to probe the LLM. To begin, you must first select the LLM you wish to audit and then input the questions you intend to explore. For each question submitted, the model will generate five prompts, each accompanied by its respective answers.</p>

        <p>To tailor the generation of these five prompts from your original question, you can adjust the relevance and diversity scores. The relevance score determines how closely the generated prompts should align with the original question, while the diversity score dictates the variance among the prompts themselves.</p>

        <p>Upon completion, please provide your email address. We will compile and send the answers to you promptly.</p>
        </div>
        """)

        gr.Markdown("## Batch Mode Auditing LLMs")
            gr.Markdown("In batch auditing mode, you have the capability to probe the LLM. To begin, you must first select the LLM you wish to audit and then input the questions you intend to explore. For each question submitted, the model will generate five prompts, each accompanied by its respective answers.")
            gr.Markdown("To tailor the generation of these five prompts from your original question, you can adjust the relevance and diversity scores. The relevance score determines how closely the generated prompts should align with the original question, while the diversity score dictates the variance among the prompts themselves.")
            gr.Markdown("Upon completion, please provide your email address. We will compile and send the answers to you promptly.")
            
            llm_dropdown = gr.Dropdown([("Llama", "TheBloke/Llama-2-7B-Chat-GGML"), ("Falcon", "TheBloke/Falcon-180B-Chat-GGUF"), ("Zephyr", "TheBloke/zephyr-quiklang-3b-4K-GGUF"), ("Vicuna", "TheBloke/vicuna-33B-GGUF"), ("Claude", "TheBloke/claude2-alpaca-13B-GGUF"), ("Alpaca", "TheBloke/LeoScorpius-GreenNode-Alpaca-7B-v1-GGUF")], label="Large Language Model")
            questions_textbox = gr.Textbox(label="Enter your question", placeholder="Enter your questions here...")
            file_upload = gr.File(label="Or You Can Click to Upload a File")
            relevance_slider = gr.Slider(0, 100, value=70, step=1, label="Relevance")
            diversity_slider = gr.Slider(0, 100, value=25, step=1, label="Diversity")
            email_input = gr.Textbox(label="Enter your email address", placeholder="[email protected]")
            
            submit_button = gr.Button("Submit")
            submit_button.click(fn=process_inputs, inputs=[llm_dropdown, questions_textbox, relevance_slider, diversity_slider, email_input], outputs="text")


    
# Launch the Gradio app
demo.launch()