app.py

import streamlit as st
import requests
import os
from streamlit_chat import message
import random
from sentence_transformers import SentenceTransformer, util
import nltk
import numpy as np
import pandas as pd

nltk.download("punkt")


context = "To extract information from documents, use sentence similarity task. To classify sentiments, use text classification task. To do sentiment analysis, use text classification task. To detect masks from images, use object detection task. To extract name or address from documents use token classification task. To extract name or address from invoices, use token classification task. To build voice enabled applications, you can use automatic speech recognition task. You can retrieve information from documents using sentence similarity task. You can summarize papers using summarization task. You can convert text to speech using text-to-speech task. To detect language spoken in an audio, you can use audio classification task. To detect emotion in an audio, you can use audio classification task. To detect commands in an audio, you can use audio classification task. To decompose sounds in a recording, use audio-to-audio task. To answer questions from a document, you can use question answering task. To answer FAQs from your customers, you can use question answering task. To see if a text is grammatically correct, you can use text classification task. To augment your training data, you can use text classification task. To detect pedestrians, you can use object detection task."


link_dict = {
    "audio-to-audio": "https://huggingface.co/tasks/audio-to-audio",
    "audio classification": "https://huggingface.co/tasks/audio-classification",
    "automatic speech recognition": "https://huggingface.co/tasks/automatic-speech-recognition",
    "fill-mask":"https://huggingface.co/tasks/fill-mask",
    "image classification": "https://huggingface.co/tasks/image-classification",
    "image segmentation": "https://huggingface.co/tasks/image-segmentation",
    "question answering":"https://huggingface.co/tasks/question-answering",
    "text-to-speech":"https://huggingface.co/tasks/text-to-speech",
    "sentence similarity": "https://huggingface.co/tasks/sentence-similarity",
    "summarization":"https://huggingface.co/tasks/summarization",
    "text generation": "https://huggingface.co/tasks/text-generation",
    "translation": "https://huggingface.co/tasks/translation",
    "token classification": "https://huggingface.co/tasks/token-classification",
    "object detection": "https://huggingface.co/tasks/object-detection"}
    


model_name = 'sentence-transformers/msmarco-distilbert-base-v4'
max_sequence_length = 512

model = SentenceTransformer(model_name)
model.max_seq_length = max_sequence_length
corpus = []
sentence_count = []

for sent in context.split("."):

    sentences = nltk.tokenize.sent_tokenize(str(sent), language='english')
    sentence_count.append(len(sentences))
    for _,s in enumerate(sentences):
        corpus.append(s)
        
        
        
corpus_embeddings = np.load('task_embeddings_msmarco-distilbert-base-v4.npy')
corpus_embeddings.shape




def find_sentences(query):
    query_embedding = model.encode(query)
    hits = util.semantic_search(query_embedding, corpus_embeddings, top_k=5)
    hits = hits[0]

    for hit in hits:
        corpus_id = hit['corpus_id']

        # Find source document based on sentence index
        count = 0
        for idx, c in enumerate(sentence_count):
            count+=c
            if (corpus_id > count-1):
                continue
            else:
                doc = corpus[idx]

                break







message_history = [{"text":"Let's find out the best task for your use case! Tell me about your use case :)", "is_user":False}]

st.subheader("If you don't know how to build your machine learning product for your use case, Taskmaster is here to help you! 🪄✨")
st.write("Hint: Try to ask your use case in a question form.")
for msg in message_history:
    message(msg["text"], is_user = msg["is_user"])   # display all the previous message

placeholder = st.empty()  # placeholder for latest message


input = st.text_input("Ask me 🤗")
if input:

    message_history.append({"text":input, "is_user" : True})

    model_answer = find_sentences("How can I extract information from invoices?")


    if resp.status_code == 200:
    

        key_exists = False
        for key in link_dict:
            if key in model_answer:
                key_exists = True
                url = link_dict[key]
                response_templates = [f"I think that {model_answer} is the best task for this 🤩 Check out the page 👉🏼 {url}", f"I think you should use {model_answer} 🪄 Check it out here 👉🏼 {url}", f"I think {model_answer} should work for you 🤓 Check out the page 👉🏼 {url}"]
        
                bot_answer = random.choice(response_templates)
                message_history.append({"text":bot_answer, "is_user" : False})
        if key_exists == False:
            fallback_template = ["I didn't get the question 🧐 Could you please ask again? Try 'What should I use for detecting masks in an image?'",
                                     "Hmm, not sure I know the answer, maybe you could ask differently? 🤓",
                                     "Sorry, I didn't understand you, maybe you could ask differently? 🤓 Try asking 'What should I use to extract name in a document' 🤗"]
            bot_answer = random.choice(fallback_template)
            message_history.append({"text":bot_answer, "is_user" : False})
    
    with placeholder.container():
        last_message = message_history[-1]
        if last_message:
            message(last_message["text"], last_message["is_user"])