ESG_analysis_tool / LocalT_ESG_RAG_1.25.py
Mjlehtim's picture
Rename LocalT_ESG_RAG_1.32.py to LocalT_ESG_RAG_1.25.py
7af9a42 verified
raw
history blame
29.9 kB
import os
import shutil
import streamlit as st
from fpdf import FPDF
from chromadb import Client
from chromadb.config import Settings
import chromadb
from langchain_community.utilities import SerpAPIWrapper
from llama_index.core import VectorStoreIndex
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough
from langchain_groq import ChatGroq
from langchain.chains import LLMChain
from langchain.agents import AgentType, Tool, initialize_agent, AgentExecutor
from llama_parse import LlamaParse
from langchain_community.document_loaders import UnstructuredMarkdownLoader
from langchain_huggingface import HuggingFaceEmbeddings
from llama_index.core import SimpleDirectoryReader
from dotenv import load_dotenv, find_dotenv
import pandas as pd
from streamlit_chat import message
from langchain_community.vectorstores import Chroma
from langchain_community.utilities import SerpAPIWrapper
from langchain.chains import RetrievalQA
from langchain_community.document_loaders import DirectoryLoader
from langchain_community.document_loaders import PyMuPDFLoader
from langchain_community.document_loaders import UnstructuredXMLLoader
from langchain_community.document_loaders import CSVLoader
from langchain.prompts import PromptTemplate
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.memory import ConversationBufferMemory
from langchain.prompts import PromptTemplate
import joblib
import nltk
import nest_asyncio # noqa: E402
nest_asyncio.apply()
load_dotenv()
load_dotenv(find_dotenv())
os.environ["TOKENIZERS_PARALLELISM"] = "false"
SERPAPI_API_KEY = os.environ["SERPAPI_API_KEY"]
GOOGLE_CSE_ID = os.environ["GOOGLE_CSE_ID"]
GOOGLE_API_KEY = os.environ["GOOGLE_API_KEY"]
LLAMA_PARSE_API_KEY = os.environ["LLAMA_PARSE_API_KEY"]
HUGGINGFACEHUB_API_TOKEN = os.environ["HUGGINGFACEHUB_API_TOKEN"]
groq_api_key=os.getenv('GROQ_API_KEY')
st.set_page_config(layout="wide")
css = """
<style>
[data-testid="stAppViewContainer"] {
background-color: #f8f9fa; /* Very light grey */
}
[data-testid="stSidebar"] {
background-color: white;
color: black;
}
[data-testid="stAppViewContainer"] * {
color: black; /* Ensure all text is black */
}
button {
background-color: #add8e6; /* Light blue for primary buttons */
color: black;
border: 2px solid green; /* Green border */
}
button:hover {
background-color: #87ceeb; /* Slightly darker blue on hover */
}
button:active {
outline: 2px solid green; /* Green outline when the button is pressed */
outline-offset: 2px; /* Space between button and outline */
}
.stButton>button:first-child {
background-color: #add8e6; /* Light blue for primary buttons */
color: black;
}
.stButton>button:first-child:hover {
background-color: #87ceeb; /* Slightly darker blue on hover */
}
.stButton>button:nth-child(2) {
background-color: #b0e0e6; /* Even lighter blue for secondary buttons */
color: black;
}
.stButton>button:nth-child(2):hover {
background-color: #add8e6; /* Slightly darker blue on hover */
}
[data-testid="stFileUploadDropzone"] {
background-color: white; /* White background for file upload */
}
[data-testid="stFileUploadDropzone"] .stDropzone, [data-testid="stFileUploadDropzone"] .stDropzone input {
color: black; /* Ensure file upload text is black */
}
.stButton>button:active {
outline: 2px solid green; /* Green outline when the button is pressed */
outline-offset: 2px;
}
</style>
"""
st.write(css, unsafe_allow_html=True)
#st.sidebar.image('lt.png', width=250)
#-------------
llm=ChatGroq(groq_api_key=groq_api_key,
model_name="Llama-3.1-70b-Versatile", temperature = 0.0, streaming=True)
#--------------
doc_retriever_ESG = None
doc_retriever_financials = None
#--------------
#@st.cache_data
def load_or_parse_data_ESG():
data_file = "./data/parsed_data_ESG.pkl"
parsingInstructionUber10k = """The provided document contain detailed information about the company's environmental, social and governance matters.
It contains several tables, figures and statistical information about CO2 emissions and energy consumption.
Give only precide CO2 and energy consumotion levels inly from the context documents.
You must never provide false numeric or statistical data that is not included in the context document.
Include tables and numeric data always when possible. Only refer to other sources if the context document refers to them or if necessary to provide additional understanding to company's own data."""
parser = LlamaParse(api_key=LLAMA_PARSE_API_KEY,
result_type="markdown",
parsing_instruction=parsingInstructionUber10k,
max_timeout=5000,
gpt4o_mode=True,
)
file_extractor = {".pdf": parser}
reader = SimpleDirectoryReader("./ESG_Documents", file_extractor=file_extractor)
documents = reader.load_data()
print("Saving the parse results in .pkl format ..........")
joblib.dump(documents, data_file)
# Set the parsed data to the variable
parsed_data_ESG = documents
return parsed_data_ESG
#@st.cache_data
def load_or_parse_data_financials():
data_file = "./data/parsed_data_financials.pkl"
parsingInstructionUber10k = """The provided document is the company's annual reports and includes financial statement, balance sheet, cash flow sheet and description of the company's business and operations.
It contains several tabless, figures and statistical information. You must be precise while answering the questions and never provide false numeric or statistical data."""
parser = LlamaParse(api_key=LLAMA_PARSE_API_KEY,
result_type="markdown",
parsing_instruction=parsingInstructionUber10k,
max_timeout=5000,
gpt4o_mode=True,
)
file_extractor = {".pdf": parser}
reader = SimpleDirectoryReader("./Financial_Documents", file_extractor=file_extractor)
documents = reader.load_data()
print("Saving the parse results in .pkl format ..........")
joblib.dump(documents, data_file)
# Set the parsed data to the variable
parsed_data_financials = documents
return parsed_data_financials
#--------------
# Create vector database
@st.cache_resource
def create_vector_database_ESG():
# Call the function to either load or parse the data
llama_parse_documents = load_or_parse_data_ESG()
with open('data/output_ESG.md', 'a') as f: # Open the file in append mode ('a')
for doc in llama_parse_documents:
f.write(doc.text + '\n')
markdown_path = "data/output_ESG.md"
loader = UnstructuredMarkdownLoader(markdown_path)
documents = loader.load()
# Split loaded documents into chunks
text_splitter = RecursiveCharacterTextSplitter(chunk_size=400, chunk_overlap=30)
docs = text_splitter.split_documents(documents)
#len(docs)
print(f"length of documents loaded: {len(documents)}")
print(f"total number of document chunks generated :{len(docs)}")
embed_model = HuggingFaceEmbeddings()
vs = Chroma.from_documents(
documents=docs,
embedding=embed_model,
collection_name="rag",
)
doc_retriever_ESG = vs.as_retriever()
index = VectorStoreIndex.from_documents(llama_parse_documents)
query_engine = index.as_query_engine()
return doc_retriever_ESG, query_engine
@st.cache_resource
def create_vector_database_financials():
# Call the function to either load or parse the data
llama_parse_documents = load_or_parse_data_financials()
with open('data/output_financials.md', 'a') as f: # Open the file in append mode ('a')
for doc in llama_parse_documents:
f.write(doc.text + '\n')
markdown_path = "data/output_financials.md"
loader = UnstructuredMarkdownLoader(markdown_path)
documents = loader.load()
text_splitter = RecursiveCharacterTextSplitter(chunk_size=100, chunk_overlap=15)
docs = text_splitter.split_documents(documents)
embed_model = HuggingFaceEmbeddings()
vs = Chroma.from_documents(
documents=docs,
embedding=embed_model,
collection_name="rag"
)
doc_retriever_financials = vs.as_retriever()
index = VectorStoreIndex.from_documents(llama_parse_documents)
query_engine_financials = index.as_query_engine()
print('Vector DB created successfully !')
return doc_retriever_financials, query_engine_financials
#--------------
ESG_analysis_button_key = "ESG_strategy_button"
portfolio_analysis_button_key = "portfolio_strategy_button"
#---------------
def delete_files_and_folders(folder_path):
for root, dirs, files in os.walk(folder_path, topdown=False):
for file in files:
try:
os.unlink(os.path.join(root, file))
except Exception as e:
st.error(f"Error deleting {os.path.join(root, file)}: {e}")
for dir in dirs:
try:
os.rmdir(os.path.join(root, dir))
except Exception as e:
st.error(f"Error deleting directory {os.path.join(root, dir)}: {e}")
#---------------
uploaded_files_ESG = st.sidebar.file_uploader("Choose a Sustainability Report", accept_multiple_files=True, key="ESG_files")
for uploaded_file in uploaded_files_ESG:
st.write("filename:", uploaded_file.name)
def save_uploadedfile(uploadedfile):
with open(os.path.join("ESG_Documents",uploadedfile.name),"wb") as f:
f.write(uploadedfile.getbuffer())
return st.success("Saved File:{} to ESG_Documents".format(uploadedfile.name))
save_uploadedfile(uploaded_file)
uploaded_files_financials = st.sidebar.file_uploader("Choose an Annual Report", accept_multiple_files=True, key="financial_files")
for uploaded_file in uploaded_files_financials:
st.write("filename:", uploaded_file.name)
def save_uploadedfile(uploadedfile):
with open(os.path.join("Financial_Documents",uploadedfile.name),"wb") as f:
f.write(uploadedfile.getbuffer())
return st.success("Saved File:{} to Financial_Documents".format(uploadedfile.name))
save_uploadedfile(uploaded_file)
#---------------
def ESG_strategy():
doc_retriever_ESG, _ = create_vector_database_ESG()
prompt_template = """<|system|>
You are a seasoned specialist in environmental, social and governance matters. You write expert analyses for institutional investors. Always use figures, nemerical and statistical data when possible. Output must have sub-headings in bold font and be fluent.<|end|>
<|user|>
Answer the {question} based on the information you find in context: {context} <|end|>
<|assistant|>"""
prompt = PromptTemplate(template=prompt_template, input_variables=["question", "context"])
qa = (
{
"context": doc_retriever_ESG,
"question": RunnablePassthrough(),
}
| prompt
| llm
| StrOutputParser()
)
ESG_answer_1 = qa.invoke("Give a summary what specific ESG measures the company has taken recently and compare these to the best practices.")
ESG_answer_2 = qa.invoke("Does the company's main business fall under the European Union's taxonomy regulation? Is the company taxonomy compliant under European Union Taxonomy Regulation?")
ESG_answer_3 = qa.invoke("Explain what items of ESG information the company publishes. Describe what ESG transparency commitments the company has given. Does the company follow the Paris Treaty's obligation to limit globabl warming to 1.5 celcius degrees?")
ESG_answer_4 = qa.invoke("Does the company have carbon emissions reduction plan and has the company reached its carbod dioxide reduction objectives? Set out in a table the company's carbon footprint by location and its development from the context. Set out carbon dioxide emissions in relation to turnover.")
ESG_answer_5 = qa.invoke("Describe and set out in a table the following carbon emissions figures: (i) Scope 1 CO2 emissions, (ii) Scope 2 CO2, and (iii) Scope 3 CO2 emissions. Set out the material changes relating to these figures.")
ESG_answer_6 = qa.invoke("Set out in a table the company's energy and renewable energy usage for each material activity coverning the available years. Explain the energy efficiency measures taken by the company.")
ESG_answer_7 = qa.invoke("Does the company follow UN Guiding Principles on Business and Human Rights, ILO Declaration on Fundamental Principles and Rights at Work or OECD Guidelines for Multinational Enterprises that involve affected communities? Set out the measures taken to have the gender balance on the upper management of the company.")
ESG_answer_8 = qa.invoke("List the environmental permits and certifications held by the company. Set out and explain any environmental procedures and investigations and decisions taken against the company. Answer whether the company's locations or operations are connected to areas sensitive in relation to biodiversity.")
ESG_answer_9 = qa.invoke("Set out waste produces by the company and possible waste into the soil by real estate. Describe if the company's real estates have hazardous waste.")
ESG_answer_10 = qa.invoke("What percentage of women are represented in the (i) board, (ii) executive directors and (iii) upper management?")
ESG_answer_11 = qa.invoke("What policies has the company implemented to counter money laundering and corruption?")
ESG_output = f"**__Summary of ESG reporting and obligations:__** {ESG_answer_1} \n\n **__Compliance with taxonomy:__** \n\n {ESG_answer_2} \n\n **__Disclosure transparency:__** \n\n {ESG_answer_3} \n\n **__Carbon footprint:__** \n\n {ESG_answer_4} \n\n **__Carbon dioxide emissions:__** \n\n {ESG_answer_5} \n\n **__Renewable energy:__** \n\n {ESG_answer_6} \n\n **__Human rights compliance:__** \n\n {ESG_answer_7} \n\n **__Management and gender balance:__** \n\n {ESG_answer_8} \n\n **__Waste and other emissions:__** {ESG_answer_9} \n\n **__Gender equality:__** {ESG_answer_10} \n\n **__Anti-money laundering:__** {ESG_answer_11}"
financial_output = ESG_output
with open("ESG_analysis.txt", 'w') as file:
file.write(financial_output)
return financial_output
#-------------
@st.cache_data
def generate_ESG_strategy() -> str:
ESG_output = ESG_strategy()
st.session_state.results["ESG_analysis_button_key"] = ESG_output
return ESG_output
#---------------
#@st.cache_data
def create_pdf():
text_file = "ESG_analysis.txt"
pdf = FPDF('P', 'mm', 'A4')
pdf.add_page()
pdf.set_margins(10, 10, 10)
pdf.set_font("Arial", size=15)
#image = "lt.png"
#pdf.image(image, w = 40)
# Add introductory lines
#pdf.cell(0, 10, txt="Company name", ln=1, align='C')
pdf.cell(0, 10, txt="Structured ESG Analysis", ln=2, align='C')
pdf.ln(5)
pdf.set_font("Arial", size=11)
try:
with open(text_file, 'r', encoding='utf-8') as f:
for line in f:
# Replace '\u2019' with a different character or string
#line = line.replace('\u2019', "'") # For example, replace with apostrophe
#line = line.replace('\u2265', "'") # For example, replace with apostrophe
#pdf.multi_cell(0, 6, txt=line, align='L')
pdf.multi_cell(0, 6, txt=line.encode('latin-1', 'replace').decode('latin-1'), align='L')
pdf.ln(5)
except UnicodeEncodeError:
print("UnicodeEncodeError: Some characters could not be encoded in Latin-1. Skipping...")
pass # Skip the lines causing UnicodeEncodeError
output_pdf_path = "ESG_analysis.pdf"
pdf.output(output_pdf_path)
#----------------
#llm = build_llm()
if 'results' not in st.session_state:
st.session_state.results = {
"ESG_analysis_button_key": {}
}
loaders = {'.pdf': PyMuPDFLoader,
'.xml': UnstructuredXMLLoader,
'.csv': CSVLoader,
}
def create_directory_loader(file_type, directory_path):
return DirectoryLoader(
path=directory_path,
glob=f"**/*{file_type}",
loader_cls=loaders[file_type],
)
strategies_container = st.container()
with strategies_container:
mrow1_col1, mrow1_col2 = st.columns(2)
st.sidebar.info("To get started, please upload the documents from the company you would like to analyze.")
button_container = st.sidebar.container()
if os.path.exists("ESG_analysis.txt"):
create_pdf()
with open("ESG_analysis.pdf", "rb") as pdf_file:
PDFbyte = pdf_file.read()
st.sidebar.download_button(label="Download Analyses",
data=PDFbyte,
file_name="strategy_sheet.pdf",
mime='application/octet-stream',
)
if button_container.button("Clear All"):
st.session_state.button_states = {
"ESG_analysis_button_key": False,
}
st.session_state.button_states = {
"portfolio_analysis_button_key": False,
}
st.session_state.results = {}
st.session_state['history'] = []
st.session_state['generated'] = ["Let's discuss the ESG issues of the company 🤗"]
st.session_state['past'] = ["Hey ! 👋"]
st.cache_data.clear()
st.cache_resource.clear()
# Check if the subfolder exists
if os.path.exists("ESG_Documents"):
for filename in os.listdir("ESG_Documents"):
file_path = os.path.join("ESG_Documents", filename)
try:
if os.path.isfile(file_path):
os.unlink(file_path)
except Exception as e:
st.error(f"Error deleting {file_path}: {e}")
else:
pass
if os.path.exists("Financial_Documents"):
# Iterate through files in the subfolder and delete them
for filename in os.listdir("Financial_Documents"):
file_path = os.path.join("Financial_Documents", filename)
try:
if os.path.isfile(file_path):
os.unlink(file_path)
except Exception as e:
st.error(f"Error deleting {file_path}: {e}")
else:
pass
# st.warning("No 'data' subfolder found.")
if os.path.exists("ESG_Documents_Portfolio"):
# Iterate through files in the subfolder and delete them
for filename in os.listdir("ESG_Documents_Portfolio"):
file_path = os.path.join("ESG_Documents_Portfolio", filename)
try:
if os.path.isfile(file_path):
os.unlink(file_path)
except Exception as e:
st.error(f"Error deleting {file_path}: {e}")
else:
pass
# st.warning("No 'data' subfolder found.")
folders_to_clean = ["data", "chroma_db_portfolio", "chroma_db_LT", "chroma_db_fin"]
for folder_path in folders_to_clean:
if os.path.exists(folder_path):
for item in os.listdir(folder_path):
item_path = os.path.join(folder_path, item)
try:
if os.path.isfile(item_path) or os.path.islink(item_path):
os.unlink(item_path) # Remove files or symbolic links
elif os.path.isdir(item_path):
shutil.rmtree(item_path) # Remove subfolders and all their contents
except Exception as e:
st.error(f"Error deleting {item_path}: {e}")
else:
pass
# st.warning(f"No '{folder_path}' folder found.")
with mrow1_col1:
st.subheader("Summary of the ESG Analysis")
st.info("This tool is designed to provide a comprehensive ESG risk analysis for institutional investors.")
button_container2 = st.container()
if "button_states" not in st.session_state:
st.session_state.button_states = {
"ESG_analysis_button_key": False,
}
if "results" not in st.session_state:
st.session_state.results = {}
if button_container2.button("ESG Analysis", key=ESG_analysis_button_key):
st.session_state.button_states[ESG_analysis_button_key] = True
result_generator = generate_ESG_strategy() # Call the generator function
st.session_state.results["ESG_analysis_output"] = result_generator
if "ESG_analysis_output" in st.session_state.results:
st.write(st.session_state.results["ESG_analysis_output"])
st.divider()
with mrow1_col2:
if "ESG_analysis_button_key" in st.session_state.results and st.session_state.results["ESG_analysis_button_key"]:
doc_retriever_ESG, query_engine = create_vector_database_ESG()
doc_retriever_financials, query_engine_financials = create_vector_database_financials()
memory = ConversationBufferMemory(memory_key="chat_history", k=3, return_messages=True)
search = SerpAPIWrapper()
# Updated prompt templates to include chat history
def format_chat_history(chat_history):
"""Format chat history as a single string for input to the chain."""
formatted_history = "\n".join([f"User: {entry['input']}\nAI: {entry['output']}" for entry in chat_history])
return formatted_history
prompt_financials = PromptTemplate.from_template(
template="""
You are a seasoned corporate finance specialist.
Use figures, numerical, and statistical data when possible. Never give false information, numbers or data.
Conversation history:
{chat_history}
Based on the context: {context}, answer the following question: {question}.
"""
)
prompt_ESG = PromptTemplate.from_template(
template="""
You are a seasoned finance specialist and a specialist in environmental, social, and governance matters.
Use figures, numerical, and statistical data when possible. Never give false information, numbers or data.
Conversation history:
{chat_history}
Based on the context: answer the following question: {question}.
"""
)
# LCEL Chains with memory integration
financials_chain = (
{
"context": doc_retriever_financials,
# Lambda function now accepts one argument (even if unused)
"chat_history": lambda _: format_chat_history(memory.load_memory_variables({})["chat_history"]),
"question": RunnablePassthrough(),
}
| prompt_financials
| llm
| StrOutputParser()
)
ESG_chain = (
{
"context": doc_retriever_ESG,
"chat_history": lambda _: format_chat_history(memory.load_memory_variables({})["chat_history"]),
"question": RunnablePassthrough(),
}
| prompt_ESG
| llm
| StrOutputParser()
)
# Define the tools with LCEL expressions
# Define the vector query engine tool
vector_query_tool_ESG = Tool(
name="Vector Query Engine ESG",
func=lambda query: query_engine.query(query), # Use query_engine to query the vector database
description="Useful for answering questions that require ESG figures, data and statistics.",
)
vector_query_tool_financials = Tool(
name="Vector Query Engine Financials",
func=lambda query: query_engine_financials.query(query), # Use query_engine to query the vector database
description="Useful for answering questions that require financial figures, data and statistics.",
)
# Create a function to validate responses
def validate_esg_response(query):
esg_response = vector_query_tool_ESG.func(query)
esg_validation = ESG_chain.invoke({
"context": doc_retriever_ESG,
"chat_history": format_chat_history(memory.load_memory_variables({})["chat_history"]),
"question": esg_response
})
return esg_validation
def validate_financials_response(query):
financials_response = vector_query_tool_financials.func(query)
financials_validation = financials_chain.invoke({
"context": doc_retriever_financials,
"chat_history": format_chat_history(memory.load_memory_variables({})["chat_history"]),
"question": financials_response
})
return financials_validation
# Update the tools list to include validation
tools = [
Tool(
name="Search Tool",
func=search.run,
description="Useful when other tools do not provide the answer.",
),
Tool(
name="Validate ESG Response",
func=validate_esg_response,
description="Validates the response of the Vector Query Engine ESG tool.",
),
Tool(
name="Validate Financials Response",
func=validate_financials_response,
description="Validates the response of the Vector Query Engine Financials tool.",
),
vector_query_tool_ESG,
vector_query_tool_financials,
]
# Initialize the agent with LCEL tools and memory
agent = initialize_agent(
tools, llm, agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION, verbose=True, memory=memory, handle_parsing_errors=True)
def conversational_chat(query):
# Get the result from the agent
result = agent.invoke({"input": query, "chat_history": st.session_state['history']})
# Handle different response types
if isinstance(result, dict):
# Extract the main content if the result is a dictionary
result = result.get("output", "") # Adjust the key as needed based on your agent's output
elif isinstance(result, list):
# If the result is a list, join it into a single string
result = "\n".join(result)
elif not isinstance(result, str):
# Convert the result to a string if it is not already one
result = str(result)
# Add the query and the result to the session state
st.session_state['history'].append((query, result))
# Update memory with the conversation
memory.save_context({"input": query}, {"output": result})
# Return the result
return result
# Ensure session states are initialized
if 'history' not in st.session_state:
st.session_state['history'] = []
if 'generated' not in st.session_state:
st.session_state['generated'] = ["Let's discuss the ESG matters and financial matters 🤗"]
if 'past' not in st.session_state:
st.session_state['past'] = ["Hey ! 👋"]
if 'input' not in st.session_state:
st.session_state['input'] = ""
# Streamlit layout
st.subheader("Discuss the ESG and financial matters")
st.info("This tool is designed to enable discussion about the ESG and financial matters concerning the company.")
response_container = st.container()
container = st.container()
with container:
with st.form(key='my_form'):
user_input = st.text_input("Query:", placeholder="What would you like to know about ESG and financial matters", key='input')
submit_button = st.form_submit_button(label='Send')
if submit_button and user_input:
output = conversational_chat(user_input)
st.session_state['past'].append(user_input)
st.session_state['generated'].append(output)
user_input = "Query:"
#st.session_state['input'] = ""
# Display generated responses
if st.session_state['generated']:
with response_container:
for i in range(len(st.session_state['generated'])):
message(st.session_state["past"][i], is_user=True, key=str(i) + '_user', avatar_style="shapes")
message(st.session_state["generated"][i], key=str(i), avatar_style="icons")