app.py

import os
import time
import csv
import shutil
from datetime import datetime
import openai
import gradio as gr

# Embeddings
from langchain.embeddings.openai import OpenAIEmbeddings
from langchain.vectorstores import Chroma

# Chat Q&A
from langchain.chat_models import ChatOpenAI
from langchain.schema import AIMessage, HumanMessage, SystemMessage

# This sets up OpenAI embeddings model
embeddings = OpenAIEmbeddings()

# Loads database from persisted directory
db_directory = "2023_12_04_chroma_db"
db = Chroma(persist_directory=db_directory, embedding_function=embeddings)

# This is code that retrieves relevant documents based on a similarity search (in this case, it grabs the top 2 relevant documents or chunks)
retriever = db.as_retriever(search_type='similarity', search_kwargs={"k":2})

with open('system_prompt.txt', 'r') as file:
    ORIG_SYSTEM_MESSAGE_PROMPT = file.read()

openai.api_key = os.getenv("OPENAI_API_KEY")

#chat = ChatOpenAI(model_name="gpt-3.5-turbo",temperature=0) # Faster for experiments
chat = ChatOpenAI(model_name="gpt-3.5-turbo",temperature=0)

# Make sure we don't exceed estimation of token limit:
TOKEN_LIMIT = 4096  # GPT-3.5 Turbo token limit
BUFFER = 100  # Extra tokens to consider for incoming messages
PERSISTENT_LOG_PATH = "/data/downvoted_responses.csv" # File in which to log downvoted responses
LOCAL_LOG_PATH = "./data/downvoted_responses.csv"

def estimate_tokens(texts):
    return sum([len(t.split()) for t in texts])

def truncate_history(history):
    tokens = estimate_tokens([msg.content for msg in history])
    while tokens + BUFFER > TOKEN_LIMIT and len(history) > 3:
        history = history[0:1] + history[3:]
        tokens = estimate_tokens([msg.content for msg in history])
    return history

# Here is the langchain
def predict(history, input):
    context = retriever.get_relevant_documents(input)
    print(context) #For debugging
    history_langchain_format = []
    history_langchain_format.append(SystemMessage(content=f"{ORIG_SYSTEM_MESSAGE_PROMPT}"))
    for human, ai in history:
        history_langchain_format.append(HumanMessage(content=human))
        history_langchain_format.append(AIMessage(content=ai))
    history_langchain_format.append(HumanMessage(content=input))
    history_langchain_format.append(SystemMessage(content=f"If you need to answer a question based on the previous message, here is some info: {context}"))
    
    # Truncate if history is too long
    history_langchain_format = truncate_history(history_langchain_format)
    
    gpt_response = chat(history_langchain_format)
    
    # Extract pairs of HumanMessage and AIMessage
    pairs = []
    for i in range(len(history_langchain_format)):
        if isinstance(history_langchain_format[i], HumanMessage) and (i+1 < len(history_langchain_format)) and isinstance(history_langchain_format[i+1], AIMessage):
            pairs.append((history_langchain_format[i].content, history_langchain_format[i+1].content))
    
    # Add the new AI response to the pairs for subsequent interactions
    pairs.append((input, gpt_response.content))
    
    return pairs

# Function to handle user message (this clears the interface)
def user(user_message, chatbot_history):
    return "", chatbot_history + [[user_message, ""]]

# Function to handle AI's response
def bot(chatbot_history):
    user_message = chatbot_history[-1][0] #This line is because we cleared the user_message previously in the user function above
    # Call the predict function to get the AI's response
    pairs = predict(chatbot_history, user_message)
    _, ai_response = pairs[-1]  # Get the latest response

    response_in_progress = ""
    for character in ai_response:
        response_in_progress += character
        chatbot_history[-1][1] = response_in_progress
        time.sleep(0.05)
        yield chatbot_history

def log_to_csv(question, answer):
    """Append a line to a CSV. Create a new file if needed."""
    now = datetime.today().strftime("%Y%m%d_%H:%M:%S")
    if not os.path.isfile(PERSISTENT_LOG_PATH):
        # Add the column names to the CSV
        with open(PERSISTENT_LOG_PATH, "w+") as csv_file:
            writer = csv.writer(csv_file)
            writer.writerow(["datetime", "user_question", "bot_response"])

    # Write the disliked message to the CSV
    with open(PERSISTENT_LOG_PATH, "a") as csv_file:
        writer = csv.writer(csv_file)
        writer.writerow([now, question, answer])

    # Copy file from persistent storage to local repo
    shutil.copyfile(PERSISTENT_LOG_PATH, LOCAL_LOG_PATH)

def get_voted_qa_pair(history, voted_answer):
    """Return the question-answer pair from the chat history, given a 
    particular bot answer. Note: This is required because the 'vote'
    event handler only has access to the answer that was liked/disliked.
    """
    for question, answer in history:
        if answer == voted_answer:
            return question, answer

def vote(data: gr.LikeData, history):
    """This is a function to do something with the voted information"""
    print(history)
    if data.liked:
        print("You upvoted this response: " + data.value)
    else:
        print("You downvoted this response: " + data.value)
        # Find Q/A pair that was disliked
        question, answer = get_voted_qa_pair(history, data.value)
        log_to_csv(question, answer)

# The Gradio App interface
with gr.Blocks() as demo:
    gr.Markdown("""<h1><center>DylanAI by CIONIC</center></h1>""")
    gr.Markdown("""<p><center>For best results, please ask DylanAI one question at a time. Unlike Human Dylan, DylanAI cannot multitask.</center></p>""")
    chatbot = gr.Chatbot(label="DylanAI")
    textbox = gr.Textbox(label="Type your question here")
    clear = gr.Button("Clear")
    
    # Chain user and bot functions with `.then()`
    textbox.submit(user, [textbox, chatbot], [textbox, chatbot], queue=False).then(
        bot, chatbot, chatbot,
    )
    clear.click(lambda: None, None, chatbot, queue=False)
    chatbot.like(vote, chatbot, None)

# Enable queuing
demo.queue()
demo.launch(debug=True, share=True)