app.py

# ===========================================
# ver01.01-5.workload-----app.py
# ===========================================

import asyncio
import os
import re
import time
import json

import chainlit as cl

#from tiktoken import encoding_for_model

from langchain import hub
from langchain_openai import OpenAI
from langchain.chains import LLMChain, APIChain
from langchain_core.prompts import PromptTemplate
from langchain.memory.buffer import ConversationBufferMemory
from langchain.memory import ConversationTokenBufferMemory
from langchain.memory import ConversationSummaryMemory


from api_docs_mck import api_docs_str 
#from faq_data import ansatte_faq_data, utleiere_faq_data
#from personvernspolicy import personvernspolicy_data

OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY")

daysoff_assistant_template = """
You are a customer support assistant (’kundeservice AI assistent’) for Daysoff. 
By default, you respond in Norwegian language, using a warm, direct, and professional tone. 
Your expertise is exclusively in retrieving booking information for a given booking ID assistance related to
to this.
You do not provide information outside of this scope. If a question is not about this topic, respond with 
"Ooops da, jeg driver faktisk kun med henvendelser omkring bestillingsinformasjon. Gjelder det andre henvendelser, 
kså må du nok kontakte kundeservice på [email protected]😊"
Chat History: {chat_history}
Question: {question}
Answer:
"""
daysoff_assistant_prompt = PromptTemplate(
    input_variables=['chat_history', 'question'],
    template=daysoff_assistant_template
)

api_url_template = """
Given the following API Documentation for Daysoff's official
booking information API: {api_docs}
Your task is to construct the most efficient API URL to answer
the user's question, ensuring the
call is optimized to include only the necessary information.
Question: {question}
API URL:
"""
api_url_prompt = PromptTemplate(input_variables=['api_docs', 'question'],
                                template=api_url_template)

# If the response includes booking information, provide the information verbatim (do not summarize it.)
    
api_response_template = """
With the API Documentation for Daysoff's official API: {api_docs} in mind,
and the specific user question: {question},
and given this API URL: {api_url} for querying,
and response from Daysoff's API: {api_response},
never refer the user to the API URL as your answer!
You should always provide a clear and concise summary (in Norwegian) of the booking information retrieved.
This way you directly address the user's question in a manner that reflects the professionalism and warmth 
of a human customer service agent.
Summary:
"""
api_response_prompt = PromptTemplate(
    input_variables=['api_docs', 'question', 'api_url', 'api_response'],
    template=api_response_template
)

# ---------------------------------------------------------------------------------------------------------
# 100 tokens ≃ 75 words
# system prompt(s), total = 330 tokens
# average api response = 250-300 tokens (current)
# user input "reserved" = 400 tokens (300 words max. /English; Polish, Norwegian {..}?@tiktokenizer), could be reduc3d to 140 tokens ≃ 105 words
# model output (max_tokens) = 2048

# ConversationBufferMemory = maintains raw chat history; crucial for "nuanced" follow-ups (e.g. "nuanced" ~ for non-English inputs)
# ConversationTokenBufferMemory (max_token_limit) = 1318 (gives space in chat_history for approximately 10-15 exchanges, assuming ~100 tokens/exchange) 
# ConversationSummaryMemory = scalable approach, especially useful for extended or complex interactions, caveat: loss of granular context
# ---------------------------------------------------------------------------------------------------------


@cl.on_chat_start
def setup_multiple_chains():

    llm = OpenAI(
        model='gpt-3.5-turbo-instruct',
        temperature=0.7, 
        openai_api_key=OPENAI_API_KEY,
        max_tokens=2048, 
        top_p=0.9,  
        frequency_penalty=0.1,
        presence_penalty=0.1   
    )

    # --ConversationBufferMemory
    conversation_memory = ConversationBufferMemory(memory_key="chat_history", 
                                                   max_len=30,  # --retains only the last 30 exchanges
                                                   return_messages=True,
    )
    
    # --ConversationTokenBufferMemory
    #conversation_memory = ConversationTokenBufferMemory(memory_key="chat_history",
                                                        #max_token_limit=1318,
                                                        #return_messages=True,
    #)

    # --ConversationSummaryMemory
    #conversation_memory = ConversationSummaryMemory(memory_key="chat_history",
                                                    #return_messages=True,
    #)
                                                        
    llm_chain = LLMChain(llm=llm,
                         prompt=daysoff_assistant_prompt,
                         memory=conversation_memory
                        )

    cl.user_session.set("llm_chain", llm_chain)

    api_chain = APIChain.from_llm_and_api_docs(
        llm=llm,
        api_docs=api_docs_str,
        api_url_prompt=api_url_prompt,
        api_response_prompt=api_response_prompt,
        verbose=True,
        limit_to_domains=None 
    )

    cl.user_session.set("api_chain", api_chain)

@cl.on_message
async def handle_message(message: cl.Message):
    user_message = message.content #.lower()
    llm_chain = cl.user_session.get("llm_chain")
    api_chain = cl.user_session.get("api_chain")
    
    booking_pattern = r'\b[A-Z]{6}\d{6}\b' 
    endpoint_url = "https://670dccd0073307b4ee447f2f.mockapi.io/daysoff/api/V1/booking"

    if re.search(booking_pattern, user_message):  
        bestillingskode = re.search(booking_pattern, user_message).group(0)  
        question = f"Retrieve information for booking ID {endpoint_url}?search={bestillingskode}"
   
        response = await api_chain.acall(
            {
                "bestillingskode": bestillingskode,
                "question": question
              
            },
            callbacks=[cl.AsyncLangchainCallbackHandler()])

    else:    
        response = await llm_chain.acall(user_message, callbacks=[cl.AsyncLangchainCallbackHandler()])

    response_key = "output" if "output" in response else "text"
    await cl.Message(response.get(response_key, "")).send()
    return message.content