app.py

# for setting/extracting environment variables such as API keys 
import os

### 1. For Web Scraping
# for querying Financial Modelling Prep API
from urllib.request import urlopen
import json

### 2. For Converting Scraped Text Into a Vector Store of Chunked Documents
# for tokenizing texts and splitting them into chunks of documents
from langchain.text_splitter import RecursiveCharacterTextSplitter
# for turning documents into embeddings before putting them in vector store
from langchain.embeddings import HuggingFaceEmbeddings
# for vector store for documents
from langchain.vectorstores import Chroma

### 3. For Querying LLM
# for loading HuggingFace LLM models from the hub
from langchain.llms import HuggingFaceHub
# for querying LLM conveniently using the context
from langchain.chains.question_answering import load_qa_chain

### 4. For Gradio App UI
import gradio as gr

fmp_api_key = os.environ['FMP_API_KEY']

# initialize the default model for embedding the tokenized texts, the articles are stored in this embedded form in the vector database
hf_embeddings = HuggingFaceEmbeddings()

if os.path.exists("chromadb_earnings_transcripts_extracted"):
    os.system("rm -r chromadb_earnings_transcripts_extracted")
if os.path.exists("earnings_transcripts_chromadb.zip"):
    os.system("rm earnings_transcripts_chromadb.zip")
os.system("wget https://github.com/damianboh/test_earnings_calls/raw/main/earnings_transcripts_chromadb.zip")
os.system("unzip earnings_transcripts_chromadb.zip -d chromadb_earnings_transcripts_extracted")

chroma_db = Chroma(persist_directory='chromadb_earnings_transcripts_extracted/chromadb_earnings_transcripts',embedding_function=hf_embeddings)

# Load the huggingface inference endpoint of an LLM model
# Name of the LLM model we are using, feel free to try others!
model = "mistralai/Mistral-7B-Instruct-v0.1"

# This is an inference endpoint API from huggingface, the model is not run locally, it is run on huggingface
hf_llm = HuggingFaceHub(repo_id=model,model_kwargs={'temperature':0.5,"max_new_tokens":300})


def source_question_answer(query:str,vectorstore:Chroma=chroma_db,llm:HuggingFaceHub=hf_llm):
  """
  Return answer to the query
  """
  input_docs = vectorstore.similarity_search(query,k=4)
  qa_chain = load_qa_chain(llm, chain_type="stuff")
  query = f"[INST]According to the earnings calls transcripts earlier, {query}[INST]"

  response = qa_chain.run(input_documents=input_docs, question=query)
  source_docs_1 =  input_docs[0].page_content
  source_docs_2 =  input_docs[1].page_content
  source_docs_3 =  input_docs[2].page_content
  source_docs_4 =  input_docs[3].page_content

  source_title_1 = input_docs[0].metadata['title']
  source_title_2 = input_docs[1].metadata['title']
  source_title_3 = input_docs[2].metadata['title']
  source_title_4 = input_docs[3].metadata['title']

  return response,source_docs_1 ,source_docs_2,source_docs_3,source_docs_4, source_title_1, source_title_2, source_title_3, source_title_4


with gr.Blocks() as app:

    with gr.Row():
        gr.HTML("<h1>Chat with Tesla 2023 Earnings Calls Transcripts</h1>")

    with gr.Row():
        query = gr.Textbox("How is Tesla planning to expand?", placeholder="Enter question here...", label="Enter question")
        btn = gr.Button("Ask Question")

    with gr.Row():
        gr.HTML("<h3>Answer</h3>")

    with gr.Row():
        answer = gr.Textbox(label="Answer")

    with gr.Row():
        gr.HTML("<h3>Sources Referenced from Tesla 2023 Earnings Calls Transcripts</h3>")

    with gr.Row():
        with gr.Column():
          source_title_1 = gr.Markdown()
          source1 = gr.Textbox(label="Source Text 1")
        with gr.Column():
          source_title_2 = gr.Markdown()
          source2 = gr.Textbox(label="Source Text 2")

    with gr.Row():
        with gr.Column():
          source_title_3 = gr.Markdown()
          source3 = gr.Textbox(label="Source Text 3")
        with gr.Column():
          source_title_4 = gr.Markdown()
          source4 = gr.Textbox(label="Source Text 4")
   
    query.submit(fn=source_question_answer, inputs=[query],
              outputs=[answer, source1, source2, source3, source4, source_title_1, source_title_2, source_title_3, source_title_4])
    
    btn.click(fn=source_question_answer, inputs=[query],
              outputs=[answer, source1, source2, source3, source4, source_title_1, source_title_2, source_title_3, source_title_4])

app.launch()