Day 4: Building Multi-documents RAG and packaging using Streamlit
This is part of the series?—?10 days of Retrieval Augmented Generation
Before we start our fourth day, let us have a look at what we have discussed and what lies ahead in this 10 days series:
In the last article we have seen how to build the RAG pipeline using LangChain, OpenAI and FAISS. In this article, we will package the application using StreamLit and build a UI based interface. Now, to add fun to this article instead of only one document indexing, we will index multiple documents and try to ask questions.
Creating Multiple Document RAG
In this article we will create a RAG which gives you answers after analyzing multiple documents. We will change the example of cricket from the last article, and we will create a Financial Investing expert. Let us first download the required documents -
We will save these documents first in a folder - fin_ed_docs. Now, let's first load these document using PyPDFLoader.
def load_and_process_pdfs(pdf_folder_path):
documents = []
for file in os.listdir(pdf_folder_path):
if file.endswith('.pdf'):
pdf_path = os.path.join(pdf_folder_path, file)
loader = PyPDFLoader(pdf_path)
documents.extend(loader.load())
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
splits = text_splitter.split_documents(documents)
return splits
splits = load_and_process_pdfs(pdf_folder_path)
The above code, once called, will load each document one by one, save it in a list inside a variable documents, and finally we will use RecursiveCharacterTextSplitter to split these documents so that indexing can be done later. The process is same as the previous article, only change is loading of multiple documents. Next, we will create an index of these documents
def initialize_vectorstore(splits):
return FAISS.from_documents(documents=splits, embedding=OpenAIEmbeddings(api_key="<YOUR API KEY>"))
vectorstore = initialize_vectorstore(splits)
The code above is exactly same as the previous article. Finally, the remaining steps and the code remains the same for creating Prompt Template and creation of the chain
prompt_template = """You are a finance expert. You need to answer the question related to finance.
Given below is the context and question of the user.
context = {context}
question = {question}
"""
prompt = ChatPromptTemplate.from_template(prompt_template)
llm = ChatOpenAI(model_name="gpt-3.5-turbo", temperature=0, api_key="API KEY")
def format_docs(docs):
return "\n\n".join(doc.page_content for doc in docs)
rag_chain = (
{"context": vectorstore.as_retriever() | format_docs, "question": RunnablePassthrough()}
| prompt
| llm
| StrOutputParser()
)
As the last step, we will pass the query and get the results
rag_chain.invoke("When should a person start investing?")
We will get the following answer -
'A person should start investing as early as possible. The context mentioned that the sooner one starts investing, the better. By investing early, one allows their investments more time to grow and accumulate principal, interest, or dividends year after year. The three golden rules for all investors are to invest early, invest regularly, and invest for the long term. The recommended age group for starting to invest is 18 to 35, as this is the perfect time horizon for an investor. However, it is never too late to start investing, and individuals should start as soon as they are able to save money.'
Now, let's move on to package the application using StreamLit.
领英推荐
Packaging our Application using StreamLit
As a Data Scientist or Machine Learning Engineer, we are responsible for building the backend which gives the response of Machine Learning models. But, displaying this response generally happens in the UI side. We convert our models into an API and then the frontend team connect with these APIs to show the results. Generally, React JS or Angular JS is used to build the Frontend. During the last few articles of this series we will look at how its done.
To present the backend to the management team, is generally hard task for engineers as they had to explain the code which upper management may not understand. Hence, to make the process of UI development fast a framework named StreamLit was made. Obviously, we cannot use this framework in production, but for presentation purpose it is used. Let's package our application and build a quick UI to present our responses.
First let's give a title to our application.
import streamlit as st
st.title("Finance Expert")
Next, a user needs to write the query. Let's create a text box for it.
user_input = st.text_input("Enter your question about finance:", "")
Lastly, a button is required to submit the query and once the response comes back a segment is required to show it.
if st.button("Submit"):
try:
response = rag_chain.invoke(user_input)
st.write(response)
except Exception as e:
st.write(f"An error occurred: {e}")
If we run this the code will execute and we will see the UI. But before we move on to run the code and see the UI, we need to cover one important point related to StreamLit.
The resources needs to be loaded in StreamLit cache, otherwise whenever we ask a query everything will get reloaded. In our case by resources I mean are the pdfs that I am loading (4 in our cases). Imagine asking a question and then these pdfs are again read and stored. This will be very time consuming. To solve this issue we will use st.cache().
@st.cache(allow_output_mutation=True)
def load_and_process_pdfs(pdf_folder_path):
...
return splits
@st.cache(allow_output_mutation=True)
def initialize_vectorstore(splits):
...
Let's understand the parameters that we passed in st.cache()
Now that we have written the complete code, let's see how the output will come
streamlit run test.py
So, we have successfuly created a multi documents RAG application with a UI built in StreamLit. In the next section we will discuss about creating a conversational bot with RAG built-in. Given below is the complete code for your reference.
import streamlit as st
from langchain import hub
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.document_loaders import PyPDFLoader
from langchain_community.vectorstores import FAISS
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough
from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain.prompts import ChatPromptTemplate
import os
# Cache the function to load and process PDF documents
@st.cache(allow_output_mutation=True)
def load_and_process_pdfs(pdf_folder_path):
documents = []
for file in os.listdir(pdf_folder_path):
if file.endswith('.pdf'):
pdf_path = os.path.join(pdf_folder_path, file)
loader = PyPDFLoader(pdf_path)
documents.extend(loader.load())
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
splits = text_splitter.split_documents(documents)
return splits
# Cache the function to initialize the vector store with documents
@st.cache(allow_output_mutation=True)
def initialize_vectorstore(splits):
return FAISS.from_documents(documents=splits, embedding=OpenAIEmbeddings(api_key="Your Key"))
pdf_folder_path = "./fin_ed_docs"
splits = load_and_process_pdfs(pdf_folder_path)
vectorstore = initialize_vectorstore(splits)
prompt_template = """You are a finance expert. You need to answer the question related to finance.
Given below is the context and question of the user. Don't answer question outside the context provided.
context = {context}
question = {question}
"""
prompt = ChatPromptTemplate.from_template(prompt_template)
llm = ChatOpenAI(model_name="gpt-3.5-turbo", temperature=0, api_key="Your Key")
def format_docs(docs):
return "\n\n".join(doc.page_content for doc in docs)
rag_chain = (
{"context": vectorstore.as_retriever() | format_docs, "question": RunnablePassthrough()}
| prompt
| llm
| StrOutputParser()
)
# Streamlit app
st.title("Finance Expert")
user_input = st.text_input("Enter your question about finance:", "")
if st.button("Submit"):
try:
response = rag_chain.invoke(user_input)
st.write(response)
except Exception as e:
st.write(f"An error occurred: {e}")