app.py

import requests
os.system('pip install openpyxl')
os.system('pip install sentence-transformers')

def gpt3_question(prompt):
    api_key = "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"
    api_endpoint = "https://api.openai.com/v1/engines/text-davinci-003/completions"
    headers = {
        "Content-Type": "application/json",
        "Authorization": f"Bearer {api_key}"
    }
    data = {
        "prompt": prompt,
        "max_tokens": 400,
        "temperature": 0.5
    }
    print('sending request')
    response = requests.post(api_endpoint, headers=headers, json=data)
    print(response)
    generated_text = response.json()["choices"][0]["text"]

    return generated_text

def chatgpt3_question(context, question):
    api_key = "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"
    url = "https://api.openai.com/v1/chat/completions"

    prompt = f"""
    based on this context: {context}
    answer this use question: {question}
    """

    headers = {
        "Content-Type": "application/json",
        "Authorization": f"Bearer {api_key}"
    }

    data = {
        "model": "gpt-3.5-turbo",
        "messages": [{"role": "user", "content": prompt}]
    }

    response = requests.post(url, headers=headers, json=data)
    generated_text = response.json()['choices'][0]['message']['content']

    return generated_text

import os
import requests
import pandas as pd

def split_paragraph(text, keyword):
    list1 = [x.strip() for x in text.split('.')]
    list2 = []
    
    for sentence in list1:
        # Check if the sentence contains the phrase "chamber of commerce"
        if keyword in sentence.lower():
            list2.append(1)
        else:
            list2.append(0)

    #in case first sentence has no keyword, we add it
    if list2[0] == 0:
        list1[0] = f'the {keyword}: ' + list1[0]
        list2[0] = 1

    # print(list1)
    # print(list2)

    list3 = list()
    current_string = ''
    # Loop through each element of list1 and list2
    for i in range(len(list1)):
        # If the corresponding element in list2 is 1, add the current string to list3 and reset the current string

        if list2[i] == 1:
            list3.append(current_string)
            current_string = "" #reset
            current_string += list1[i]

        # Otherwise, concatenate the current string with the current element of list1
        if list2[i] == 0:
            current_string += '. '+list1[i]

    # Add the final concatenated string to list3
    list3.append(current_string)

    return [x.strip() for x in list3[1:]]

def context_dict2context_list(context_dict):
    list1 = list()
    for all_keys in context_dict:
        key = all_keys.split(';')[0]
        try:
            synonyms = all_keys.split(';')[1:]
        except:
            pass
        # print(key)
        str1 = context_dict[all_keys]
        
        split_list = [x.replace('\n', '').strip() for x in str1.split('\n\n')]
        split_list

        for sentence in split_list:
            for s in split_paragraph(sentence, key):
                #add synonyms
                for synonym in synonyms:
                    #manual replacement causes a wrong grammar
                    #gpt3 replacement
                    print(s, key, synonym)
                    prompt = f'in the following sentence: {s}. Replace {key} with {synonym} correcting the grammar'
                    answer = gpt3_question(prompt).replace('\n', '')
                    list1.append(answer)
                list1.append(s)
    return list1

def list2vec(list1):
    headers = {
        'Content-Type': 'application/json',
        'Authorization': 'Bearer ' + "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",
    }

    json_data = {
        'input': list1,
        'model': 'text-embedding-ada-002',
    }

    response = requests.post('https://api.openai.com/v1/embeddings', headers=headers, json=json_data)
    return [x['embedding'] for x in response.json()['data']]

    dict1 = dict()
    for index in range(len(json_data['input'])):
        dict1[json_data['input'][index]] = response.json()['data'][index]['embedding']
    return dict1

def text2vec(query):
    headers = {
        'Content-Type': 'application/json',
        'Authorization': 'Bearer ' + "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",
    }

    json_data = {
        'input': query,
        'model': 'text-embedding-ada-002',
    }

    response = requests.post('https://api.openai.com/v1/embeddings', headers=headers, json=json_data)
    query = response.json()['data'][0]['embedding'] #len=1536 #pricing=0.0004
    return query

context_dict = {
    "company; goliath; we" : 
    """
    Our company builds AI Recommendation Systems for Matching Platforms using the latest technology. Our company is estabilished and operates in Japan. Our company uses the AWS Cloud to manage Servers. Our company can use GPT3 as well. Our company also builds GPT3-based chatbots. Our company can use open-source models, if requested. Our company uses open source models. Our company operates in Japan. Our company has been operating for 1 year, and we are expanding in Hong Kong. Our company offers other services apart from recommendation systems, like GPT3 chatbots. Our company can also build recommendation systems for mobile apps.
    """
    ,
    "price" :
    """
    The price of a recommendation system depends on the amount of complexity that is required to build, as well as the volume of customers. Reach us to get a quotation. The price of a chatbot depends by its intended usage and complexity, contact us for a quotation.
    """
    ,
    "recommendation system" :
    """
    If your company wants to recommend products to customers, we can build a recommendation system for you. GPT3 can be used to build recommendation systems by using embeddings, mapping choices in a mathematical space. Once the recommendation system has been built, we will manage it in the future as well. Recommendation system could also be built for startups, though they will be in smaller size. We use AWS OpenSearch to host recommendation system.
    """
    ,
    "a matching platform" :
    """
    A matching platform is a business with thousands of users, who could be customers, individuals or companies, who are interacting with one another. For example dating apps, ecommerce platforms, or job recruiting platforms. 
    """
}

import pandas as pd
from sentence_transformers import SentenceTransformer, util

#prepare context
context_list = context_dict2context_list(context_dict)

#adding invidivual sentences
context_list += [
    'We can also use GPT3, if requested',
    'Our email is ma@goliath.jp',
    'You can contact us at ma@goliath.jp'
]

#create df
df = pd.DataFrame([context_list, list2vec(context_list)]).T
df.columns = ['description', 'text_vector_']
df['description'] = df['description'].apply(lambda x : x.strip())
df

qa_list = {
    'how long does it take to build a recommendation system?' : 'Usually, from a few weeks to one month',
    'how long does it take to build one' : 'Usually, from a few weeks to one month',
    'how many people are working for goliath?' : '5 people',
    'how many people are working for you?' : '5 people',
    'how much does it cost?' : 'The price depends by its intended usage and complexity, contact us for a quotation.',
    'do you use GPT3 API?' : 'yes, we can',
    'do you use GPT3?' : 'yes, we can',
    'do you use GPT4?' : 'yes, we can',
    'so you build chatbots' : 'yes, we built state-of-the art chatbots with GPT3 technology'
}
df_qa = pd.DataFrame([qa_list]).T.reset_index()
df_qa.columns = ['question', 'answer']
df_qa['text_vector_'] = list2vec(df_qa['question'].values.tolist())
df_qa

df_qa_ = df_qa.copy()
df_ = df.copy()

def qa(df_, df_qa_, min_qa_score, min_context_score, verbose, query):
    query_vec = text2vec(query)

    #first check if there is already a question in df_qa
    df_qa_['score'] = df_qa_['text_vector_'].apply(lambda x : float(util.cos_sim(x, query_vec)))
    df_qa_ = df_qa_.sort_values('score', ascending=False)
    df_qa_ = df_qa_[df_qa_['score']>=min_qa_score]
    #if we find at least one possible preset answer
    if len(df_qa_) > 0:
        if verbose : display(df_qa_)
        answer = df_qa_[0:1]['answer'].values.tolist()[0]
        return answer
    
    #then check if we can use the context to answer a question
    df_['score'] = df_['text_vector_'].apply(lambda x : float(util.cos_sim(x, query_vec)))
    df_ = df_.sort_values('score', ascending=False)
    df_ = df_[df_['score']>=min_context_score]
    #if we find at least one possible preset answer
    if len(df_) > 0:
        if verbose : display(df_)
        #in case we might decide to merge multiple context
        context = ' '.join(df_['description'][0:1].values.tolist())
        answer = chatgpt3_question(context, query)
        return answer
    else:
        return 'impossible to give an answer'

# print(
#     qa(
#         df_, 
#         df_qa_, 
#         min_qa_score=0.92, 
#         min_context_score=.75, 
#         verbose=False, 
#         query='What is a recommender system?'
#     )
# )

import subprocess
import random
import gradio as gr
import requests

history = None
history_prompt = None

def predict(input, history):
    #WE CAN PLAY WITH user_input AND bot_answer, as well as history
    user_input = input

    global history_prompt
    global block_predict

    bot_answer = qa(
        df_, 
        df_qa_, 
        min_qa_score=0.92, 
        min_context_score=.75, 
        verbose=False, 
        query=input
    )

    response = list()
    response = [(input, bot_answer)]
    
    history.append(response[0])
    response = history

    # print('#history', history)
    # print('#response', response)

    return response, history

demo = gr.Blocks()
with demo:
    gr.Markdown(
    """
        Chatbot
    """
    )
    state = gr.Variable(value=[]) #beginning
    chatbot = gr.Chatbot() #color_map=("#00ff7f", "#00d5ff")
    text = gr.Textbox(
        label="Question",
        value="What is a recommendation system?",
        placeholder="",
        max_lines=1,
    )
    text.submit(predict, [text, state], [chatbot, state])
    text.submit(lambda x: "", text, text)
    # btn = gr.Button(value="submit")
    # btn.click(chatbot_foo, None, [chatbot, state])

demo.launch(share=False)