chat_with_docs.py

import os

import pandas as pd
import streamlit as st
from llama_index import GPTVectorStoreIndex, SimpleDirectoryReader
from llama_index import download_loader
from matplotlib import pyplot as plt
from pandasai.llm.openai import OpenAI

documents_folder = "documents"

# Load PandasAI loader, Which is a wrapper over PandasAI library
PandasAIReader = download_loader("PandasAIReader")

st.title("Welcome to `ChatwithDocs`")
st.header(
    "Interact with Documents such as `PDFs/CSV/Docs` using the power of LLMs\nPowered by `LlamaIndex🦙` \nCheckout the [GITHUB Repo Here](https://github.com/anoopshrma/Chat-with-Docs) and Leave a star⭐")


def get_csv_result(df, query):
    reader = PandasAIReader(llm=csv_llm)
    csv_response = reader.run_pandas_ai(
        df,
        query,
        is_conversational_answer=False
    )
    return csv_response


def save_file(doc):
    fn = os.path.basename(doc.name)
    # check if documents_folder exists in the directory
    if not os.path.exists(documents_folder):
        # if documents_folder does not exist then making the directory
        os.makedirs(documents_folder)
    # open read and write the file into the server
    open(documents_folder + '/' + fn, 'wb').write(doc.read())
    # Check for the current filename, If new filename
    # clear the previous cached vectors and update the filename 
    # with current name     
    if st.session_state.get('file_name'):
        if st.session_state.file_name != fn:
            st.cache_resource.clear()
            st.session_state['file_name'] = fn
    else:
        st.session_state['file_name'] = fn

    return fn


def remove_file(file_path):
    # Remove the file from the Document folder once 
    # vectors are created
    if os.path.isfile(documents_folder + '/' + file_path):
        os.remove(documents_folder + '/' + file_path)


@st.cache_resource
def create_index():
    # Create vectors for the file stored under Document folder. 
    # NOTE: You can create vectors for multiple files at once.
    try:
        documents = SimpleDirectoryReader(documents_folder).load_data()
        index = GPTVectorStoreIndex.from_documents(documents)
        return index
    except Exception as e:
        st.error("Failed to read documents")
    


def query_doc(vector_index, query):
    # Applies Similarity Algo, Finds the nearest match and 
    # take the match and user query to OpenAI for rich response
    query_engine = vector_index.as_query_engine()
    response = query_engine.query(query)
    return response


api_key = st.text_input("Enter your OpenAI API key here:", type="password")
if api_key:
    os.environ['OPENAI_API_KEY'] = api_key
    csv_llm = OpenAI(api_token=api_key)

tab1, tab2 = st.tabs(["CSV", "PDFs/Docs"])

with tab1:
    st.write("Chat with CSV files using PandasAI loader with LlamaIndex")
    input_csv = st.file_uploader("Upload your CSV file", type=['csv'])

    if input_csv is not None:
        st.info("CSV Uploaded Successfully")
        df = pd.read_csv(input_csv)
        st.dataframe(df, use_container_width=True)

    st.divider()

    input_text = st.text_area("Ask your query")

    if input_text is not None:
        if st.button("Send"):
            st.info("Your query: " + input_text)
            with st.spinner('Processing your query...'):
                response = get_csv_result(df, input_text)
            if plt.get_fignums():
                st.pyplot(plt.gcf())
            else:
                st.success(response)

with tab2:
    st.write("Chat with PDFs/Docs")
    input_doc = st.file_uploader("Upload your Docs")

    if input_doc is not None:
        st.info("Doc Uploaded Successfully")
        file_name = save_file(input_doc)
        index = create_index()
        remove_file(file_name)

    st.divider()
    input_text = st.text_area("Ask your question")

    if input_text is not None:
        if st.button("Ask"):
            st.info("Your query: \n" + input_text)
            with st.spinner("Processing your query.."):
                response = query_doc(index, input_text)
                print(response)

            st.success(response)

            st.divider()
            # Shows the source documents context which
            # has been used to prepare the response
            st.write("Source Documents")
            st.write(response.get_formatted_sources())