app.py

import gradio as gr
from dotenv import load_dotenv
from langchain.chains import LLMChain
from langchain_community.llms import CTransformers
from langchain_core.prompts import PromptTemplate
from sentence_transformers import SentenceTransformer
from sklearn.cluster import KMeans
from nltk.tokenize import word_tokenize
import numpy as np
import scipy.spatial
from scipy.spatial.distance import cosine


load_dotenv()
def generate_prompts(user_input):
    prompt_template = PromptTemplate(
        input_variables=["Question"],
        template= f"Your task is to formulate 5 unique queries for each given question. These queries must adhere to the criteria of relevance and diversity.write the questions in seperate lines.{user_input} "
    )
    config = {'max_new_tokens': 2048, 'temperature': 0.7, 'context_length': 4096}
    llm = CTransformers(model="TheBloke/Mistral-7B-Instruct-v0.1-GGUF",
                        config=config,
                        threads=os.cpu_count())
    hub_chain = LLMChain(prompt = prompt_template, llm = llm)

    input_data = {"Question": user_input}

    # Here you would integrate your prompt template with your model
    # For demonstration, this is just a placeholder
    generated_prompts = hub_chain.run(input_data)  # Modify this part based on how you run the model
    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, model_name):
    prompt_template = PromptTemplate(
        input_variables=["Question"],
        template=f"Give a short answer to this question '{prompt}' and do not consider the number behind it."
    )
    config = {'max_new_tokens': 512, 'temperature': 0.7, 'context_length': 512}
    llm = CTransformers(model=model_name, #"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 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_words_within_cluster(sentences, model, exclude_words):
    # Create a dictionary to map words to color codes
    word_to_color = {}
    color_codes = [
    "\033[41m",  # Background Red
    "\033[42m",  # Background Green
    "\033[43m",  # Background Yellow
    "\033[44m",  # Background Blue
    "\033[45m",  # Background Purple
    "\033[46m",  # Background Cyan
    "\033[100m", # Background Dark Gray
    "\033[101m", # Background Light Red
    "\033[102m", # Background Light Green
    "\033[103m", # Background Light Yellow
    "\033[104m", # Background Light Blue
    "\033[105m", # Background Light Purple
    "\033[106m", # Background Light Cyan
    "\033[47m"   # Background Gray
    ]
    html_color_codes = ["red", "green", "blue", "purple", "cyan", "fuchsia", "lime", "maroon", "olive", "navy", "teal", "gray"]
    color_index = 0

    highlighted_sentences = []
    for sentence in sentences:
        words = word_tokenize(sentence)
        other_sentences = [s for s in sentences if s != sentence]
        all_other_words = [word for s in other_sentences for word in word_tokenize(s) if word.lower() not in exclude_words and word.isalnum()]

        highlighted_words = []
        for word in words:
            if word.lower() not in exclude_words and word.isalnum():
                match_index, similarity = calculate_similarity(word, all_other_words, model)
                if match_index is not None:
                    # Assign color to the word if not already assigned
                    if word not in word_to_color:
                        word_to_color[word] = html_color_codes[color_index % len(html_color_codes)]
                        color_index += 1
                    # Highlight the word
                    #highlighted_word = f"{word_to_color[word]}{word}\033[0m"
                    highlighted_word = "<span style='color: "+ word_to_color[word] +"'>"+ word +"</span>"
                else:
                    highlighted_word = word
                highlighted_words.append(highlighted_word)
            else:
                highlighted_words.append(word)

        highlighted_sentences.append(' '.join(highlighted_words))
    return highlighted_sentences

# Rest of the code, including the cluster_sentences function, remains the same
def cluster_sentences(sentences, model, num_clusters=3):
    embeddings = model.encode(sentences)
    kmeans = KMeans(n_clusters=num_clusters)
    kmeans.fit(embeddings)
    return kmeans.labels_

model = SentenceTransformer('all-mpnet-base-v2')
exclude_words = {"a", "the", "for", "from", "of", "in", "over", "as", "on", "is", "am", "have", "an", "has", "had", "and", "by", "it", "its", "those", "these", "above", "to"}

text_list = []

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

def setTextVisibility(cbg, model_name_input):
    sentences = []
    result = []
    model = SentenceTransformer('all-mpnet-base-v2')
    exclude_words = {"a", "the", "for", "from", "of", "in", "over", "as", "on", "is", "am", "have", "an", "has", "had", "and", "by", "it", "its", "those", "these", "above", "to"}
    sentences_org = ["In a quaint little town nestled in the heart of the mountains, a small bakery famous for its artisanal breads and pastries had a line of customers stretching out the door, eagerly waiting to savor the freshly baked goods that were known far and wide for their delightful flavors.",
                "Within a picturesque mountain village, there stood a renowned bakery, celebrated for its handcrafted bread and sweet treats, attracting a long queue of patrons each morning, all keen to enjoy the baked delicacies that had gained widespread acclaim for their exceptional taste.",
                "A charming bakery, located in a small mountainous hamlet, renowned for producing exquisite handmade pastries and bread, was bustling with a crowd of eager customers lined up outside, each anticipating the chance to indulge in the famous baked items celebrated for their extraordinary deliciousness.",
                "In a cozy, mountain-encircled village, a beloved bakery was the center of attraction, known for its traditional baking methods and delightful pastries, drawing a consistent stream of people waiting outside, all desiring to experience the renowned flavors that made the bakery's products distinctively mouth-watering."]
    for text in cbg:
         sentences.append(answer_question(text, model_name_input))
    
    # Step 1: Cluster the sentences
    num_clusters = 1
    sentence_clusters = cluster_sentences(sentences, model, num_clusters)

    # Step 2: Highlight similar words within each cluster
    clustered_sentences = [[] for _ in range(num_clusters)]

    for sentence, cluster_id in zip(sentences, sentence_clusters):
        clustered_sentences[cluster_id].append(sentence)
    
    highlighted_clustered_sentences = []

    for cluster in clustered_sentences:
        highlighted_clustered_sentences.extend(highlight_words_within_cluster(cluster, model, exclude_words))
    
    for idx, sentence in enumerate(highlighted_clustered_sentences):
        result.append("<p><strong>"+ cbg[idx] +"</strong></p><p>"+ sentence +"</p><br/>")

    return result
    

    # 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;">
    Auditing LLMs
    </h1>
    <hr style="margin-bottom:5px; margin-top:5px;">
    
    
    
    </div>
    """)
    with gr.Tab("Live Mode"):
        with gr.Row():
            model_name_input = gr.Dropdown([("Llama", "TheBloke/Llama-2-7B-Chat-GGML"), ("Falcon", "TheBloke/Falcon-180B-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: 76%</div>
                """)

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

    with gr.Tab("Batch Mode"):
        with gr.Row():
            model_name_input = gr.Dropdown([("Llama", "TheBloke/Llama-2-7B-Chat-GGML"), ("Falcon", "TheBloke/Falcon-180B-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():
            prompt_input = gr.Textbox(label="RELAVENCY", placeholder="Relavancy")
            prompt_input = gr.Textbox(label="Diversity", placeholder="Diversity")

        with gr.Row():
            prompt_input = gr.Textbox(label="Enter your email address", placeholder="Enter Your Email Address")
        with gr.Row():
            generate_button = gr.Button("Submit", variant="primary")
            
# Launch the Gradio app
demo.launch()