app.py

import streamlit as st
from openai import OpenAI
import glob
import pickle
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_core.runnables import RunnableParallel
from langchain import VectorDBQAWithSourcesChain
from langchain.chains import RetrievalQA
import json
from documents import read_documents_from_file, create_faq_documents

OPENAI_API_KEY = st.secrets["OPENAI_API_KEY"]

# 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

# Default
def_model = "gpt-3.5-turbo"
def_temperature = 0.0
def_k = 1
def_chunk_size = 500
def_chunk_overlap = 0
directory = "./chroma_db"
#vectorstore = Chroma(persist_directory=directory, embedding_function=OpenAIEmbeddings())
st.set_page_config(initial_sidebar_state="collapsed")

data_source = st.radio("Data source", options=['FAQ', 'Blog articles'])
if data_source == 'FAQ':
    docs=read_documents_from_file("./docs/faq_docs.json")
elif data_source == 'Blog articles':
    docs=read_documents_from_file()

with st.sidebar:
    if st.toggle("Experimental"):
        disabled = False
    else:
        disabled = True
    model = def_model
    temperature = st.number_input("Temperature", value=def_temperature, min_value=0.0, step=0.2, max_value=1.0, placeholder=def_temperature, disabled=disabled)
    k = st.number_input("Number of documents to include", value=def_k, min_value=1, step=1, placeholder=def_k, disabled=disabled)

    if st.toggle("Splitting", value=True, disabled=disabled):
        chunk_size = st.number_input("Chunk size", value=def_chunk_size, step=250, placeholder=def_chunk_size, disabled=disabled) # 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 = st.number_input("Chunk overlap", value=def_chunk_overlap, step=10, placeholder=def_chunk_overlap, disabled=disabled)
        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(), persist_directory=directory)
        if chunk_size != def_chunk_size | chunk_overlap != def_chunk_overlap:
            splits = text_splitter.split_documents(docs)
            vectorstore = Chroma.from_documents(documents=splits, embedding=OpenAIEmbeddings())
            print("Created new vectordb for this session.")
    else:
        vectorstore = Chroma.from_documents(documents=docs, embedding=OpenAIEmbeddings(), persist_directory=f"{directory}/complete")
        print("Used vectordb with all blog articles.")


prompt_template ="""
###
Wenn du gefragt wirst wer du bist: Du heißt "Volker". Du bist ein Assistent für die Beantwortung von Fragen zu Long-Covid (Post-Covid synonym verwenden). 

###
Du weißt nichts über Krankheiten wie 'tumorbedingte Fatigue', 'Multiple Sklerose', 'Hashimoto-Thyreoditis' oder 'Krebs'. 
Werden Fragen zu diesen Erkrankungen gestellt, beantworte sie mit "Dazu fehlen mir Informationen".

###
Du beantwortest keine Fragen zu 'Tod', 'Suizid', 'Selbstverletzung', oder anderen potenziell schädigenden Themen.
Werden Fragen zum 'Tod' gestellt, verweise auf den behandelnden Arzt.
Bei Fragen zu Suizid verweise auf die Telefonseelsorge: 0800 1110111

###
Du gibst keine Ratschläge zur Diagnose, Behandlung oder Therapie. 
Wenn du die Antwort nicht weißt, sag einfach, dass du es nicht weißt.

###
Wenn du allgemeine unspezifische Fragen gestellt bekommst, antworte oberflächlich und frage nach einer präziseren Fragestellung.
Antworte immer in ganzen Sätzen und verwende korrekte Grammatik und Rechtschreibung. Antworte nur auf Deutsch. 
Antworte kurz mit maximal fünf Sätzen außer es wird von dir eine ausführlichere Antwort verlangt.
Verwende zur Beantwortung der Frage nur den vorhandenen Kontext.
 
Frage: {question} 
Kontext: {context} 
Antwort:
""" # Source: hub.pull("rlm/rag-prompt")

# (1) Retriever
retriever = vectorstore.as_retriever(search_type="similarity_score_threshold", search_kwargs={"score_threshold": 0.3, "k": k})

# (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_from_docs = (
    RunnablePassthrough.assign(context=(lambda x: format_docs(x["context"])))
    | prompt
    | llm
    | StrOutputParser()
)

rag_chain = RunnableParallel(
    {"context": retriever, "question": RunnablePassthrough()}
).assign(answer=rag_chain_from_docs)


st.title("🐔 Volker-Chat")

def click_button(prompt):
    st.session_state.clicked = True
    st.session_state['prompt'] = prompt


c = st.container()
c.write("Beispielfragen")
col1, col2, col3 = c.columns(3)
if data_source == 'Blog articles':
    col1.button("Mehr zu 'Lernen'", on_click=click_button, args=["Was macht die Säule 'Lernen' aus?"])
    col1.button("Was macht die Fimo Health App?", on_click=click_button, args=["Was macht die Fimo Health App?"])
    col2.button("Mehr zu 'Tracken'", on_click=click_button, args=["Was macht die Säule 'Tracken' aus?"])
    col2.button("Was ist Pacing?", on_click=click_button, args=["Was ist Pacing?"])
    col3.button("Mehr zu 'Handeln'", on_click=click_button, args=["Was macht die Säule 'Handeln' aus?"])
elif data_source == 'FAQ':
    examples = ['Wie komme ich an meinen PDF-Report?', 'Wer steckt hinter den Kurs-Inhalten?', 'Wozu dient der Check-Out?']
    for i, col in enumerate(c.columns(3)):
        question = examples[i]
        col.button(question, on_click=click_button, args=[question])

if 'clicked' not in st.session_state:
    st.session_state.clicked = False

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 st.session_state.clicked:
    prompt = st.session_state['prompt']
    st.chat_message("user").write(prompt)
    with get_openai_callback() as cb:
        response = rag_chain.invoke(prompt)
        st.chat_message("assistant").write(response['answer'])
        with st.expander("Kontext ansehen"):
            for citation in response["context"]:
                st.write("[...] ", str(citation.page_content), " [...]")
                st.write(str(citation.metadata['source']))
                st.write(str("---")*20)
        with st.sidebar:
            sidebar_c = st.container()
            sidebar_c.success(cb)


if prompt := st.chat_input():
    st.chat_message("user").write(prompt)
    with get_openai_callback() as cb:
        response = rag_chain.invoke(prompt)
        st.chat_message("assistant").write(response['answer'])
        with st.expander("Kontext ansehen"):
            for citation in response["context"]:
                st.write("[...] ", str(citation.page_content), " [...]")
                st.write(str(citation.metadata['source']))
                st.write(str("---")*20)
        with st.sidebar:
            sidebar_c = st.container()
            sidebar_c.success(cb)


# cleanup
st.session_state.clicked = False
vectorstore.delete_collection()