app.py

### title: 010125-daysoff-assistant-api
### file: app.py

import asyncio
import os
import time
import json
import torch

from api_docs_mck import daysoff_api_docs

import chainlit as cl
#from chainlit import LLMSettings # hmm..
#from chainlit.config import config # hmm...

from langchain import hub
from langchain.chains import LLMChain, APIChain
from langchain_core.prompts import PromptTemplate
from langchain_community.llms import HuggingFaceHub
from langchain.memory.buffer import ConversationBufferMemory


HUGGINGFACEHUB_API_TOKEN = os.getenv("HUGGINGFACEHUB_API_TOKEN")
LANGCHAIN_API_KEY = os.environ.get("LANGCHAIN_API_KEY")
HF_TOKEN = os.environ.get("HF_TOKEN")

#os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:true"

dtype = torch.float16
device = torch.device("cuda")

daysoff_assistant_booking_template = """
You are a customer support assistant for Daysoff.no. Your expertise is
retrieving booking information for a given booking ID (’bestillingsnummer’)"
Chat History: {chat_history}
Question: {question}
Answer:
"""
daysoff_assistant_booking_prompt= PromptTemplate(
    input_variables=["chat_history", "question"],
    template=daysoff_assistant_booking_template
)

api_url_template = """
Given the following API Documentation for Daysoff's official
booking information API: {api_docs_mck}
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_mck', 'question'],
                                template=api_url_template)

api_response_template = """"
With the API Documentation for Daysoff's official API: {api_docs_mck}
and the specific user question: {question} in mind,
and given this API URL: {api_url} for querying, here is the
response from Daysoff's API: {api_response}.
Please provide a summary that directly addresses the user's question,
omitting technical details like response format, and
focusing on delivering the answer with clarity and conciseness,
as if a human customer service agent is providing this information.
Adapt to user's language. By default, you speak Norwegian.
Summary:
"""
api_response_prompt = PromptTemplate(input_variables=['api_docs_mck',
                                                      'question',
                                                      'api_url',
                                                      'api_response'],
                                     template=api_response_template)


# --model, memory object, and llm_chain
@cl.on_chat_start
def setup_multiple_chains():
    llm = HuggingFaceHub(repo_id="google/gemma-2-2b-it",
                         temperature=0.7,
                         huggingface_api_token=HUGGINGFACEHUB_API_TOKEN,
                         device=device)

    conversation_memory = ConversationBufferMemory(memory_key="chat_history",
                                                   max_len=200,
                                                   return_messages=True,
                                                   )
    llm_chain = LLMChain(llm=llm,
                         prompt=daysoff_assistant_booking_prompt,
                         memory=conversation_memory
                        )

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

    api_chain = APIChain.from_llm_and_api_docs_mck(
        llm=llm,
        api_docs_mck=daysoff_api_docs,
        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)


# --regex for alphanum. "LLLLLLxxxxxx", i.e. booking_id  |==> 308.9 trillion unique possibilities
BOOKING_ID = r'\b[A-Z]{6}\d{6}\b'

# --keywords based from email-data
BOOKING_KEYWORDS = [
    "booking",
    "bestillingsnummer",
    "bookingen",
    "ordrenummer",
    "reservation",
    "rezerwacji", 
    "bookingreferanse", 
    "rezerwacja", 
    "logg inn", 
    "booket", 
    "reservation number",
    "bestilling", 
    "order number",
    "booking ID",
    "identyfikacyjny płatności"  
]



# --wrapper function around the @cl.on_message decorator; chain trigger(s)
@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")

    is_booking_query = any(
        re.search(keyword, user_message, re.IGNORECASE) 
        for keyword in BOOKING_KEYWORDS + [BOOKING_ID]
    )

    if is_booking_query:
        response = await api_chain.acall(user_message,
                                         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