app.py

import streamlit as st
from openai import OpenAI
import glob

from langchain_community.vectorstores import Chroma
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain.callbacks import get_openai_callback
from langchain_core.prompts import ChatPromptTemplate

from langchain_community.document_loaders import UnstructuredFileLoader

import os
os.environ["OPENAI_API_KEY"] = "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"

# Get all the filenames from the docs folder
files = glob.glob("./docs/*.txt")
# Load files into readable documents
docs = []
for file in files:
    loader = UnstructuredFileLoader(file)
    docs.append(loader.load()[0])

# Config
chunk_size = 500 # Defines the chunks in amount of tokens in which the files are split. Also defines the amount of tokens that are feeded into the context. 
chunk_overlap = 100
temperature = 0.4
model = "gpt-3.5-turbo"


prompt_template ="""
You are called "Volker". You are an assistant for question-answering tasks. You only answer questions about Long-Covid (use Post-Covid synonymously) and the Volker-App. Only use the following pieces of retrieved context to answer the question. 
If you don't know the answer, just say that you don't know. Say why you don't know the answer. 
Never answer questions about other diseases (e.g. Cancer-related fatigue, Multiple Sklerose). 
Answer the questions in the language that they are posed in.
Keep the answer concise. Stay emphatic and positive.
Question: {question} 
Context: {context} 
Answer:
""" # Source: hub.pull("rlm/rag-prompt")

text_splitter = RecursiveCharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
splits = text_splitter.split_documents(docs)
vectorstore = Chroma.from_documents(documents=splits, embedding=OpenAIEmbeddings())

# (1) Retriever
retriever = vectorstore.as_retriever()

# (2) Prompt
prompt = ChatPromptTemplate.from_template(prompt_template)

# (3) LLM
# Define the LLM we want to use. Default is "gpt-3.5-turbo" with temperature 0. 
# Temperature is a number between 0 and 1. With 0.8 it generates more random answers, with 0.2 it is more focused on the retrieved content. With temperature = 0 it uses log-probabilities depending on the content.

llm = ChatOpenAI(model_name=model, temperature=temperature)

def format_docs(docs):
    return "\n\n".join(doc.page_content for doc in docs)


rag_chain = (
    {"context": retriever | format_docs, "question": RunnablePassthrough()}
    | prompt
    | llm
    | StrOutputParser()
)

st.title("💬 Volker-Chat")

if "messages" not in st.session_state:
    st.session_state["messages"] = [{"role": "assistant", "content": "Ahoi! Ich bin Volker. Wie kann ich dir helfen?"}]

for msg in st.session_state.messages:
    st.chat_message(msg["role"]).write(msg["content"])

if prompt := st.chat_input():
    st.chat_message("user").write(prompt)
    response = rag_chain.invoke(prompt)
    st.chat_message("assistant").write(response)