Generative AI Series

Ongoing (Q4 '24):

5-part Generative AI Series
Create RAG systems and AI agents with Sectors Financial API, LangChain and state-of-the-art LLM models -- capable of producing fact-based financial analysis and financial-specific reasoning. **Continually updated** to keep up with the latest major versions of the tools and libraries used in the series.

Generative AI Series: Table of Contents

1

Generative AI for Finance

An overview of designing Generative AI systems for the finance industry and the motivation for retrieval-augmented generation (RAG) systems.

Tool-Use Retrieval Augmented Generation (RAG)

3

Structured Output from AIs

From using Generative AI to extract from unstructured data or perform actions like database queries, API calls, JSON parsing and more, we need schema and structure in the AI's output.
4

Tool-use ReAct Agents w/ Streaming

Updated for LangChain v0.3.2, we explore streaming, LCEL expressions and ReAct agents following the most up-to-date practices for creating conversational AI agents.
5

Conversational Memory AI Agents

Updated for LangChain v0.2.3, we dive into Creating AI Agents with Conversational Memory

​
API-based RAG systems architecture

We’ve learned from the previous chapter that we can use APIs to retrieve data from the web. In this chapter, we’ll learn how to use APIs to retrieve data from the web and use it to build a simple financial model. We’ve also seen how such a system falls into a broader category of systems that are referred to as Retrieval-Augmented Generation (RAG) models.

Rather than generating text solely from scratch, RAG models such as the one we’ll build in this chapter use information retrieved from an external data source to form the basis of their text generation. Let’s take one more look at the architecture of our API-based RAG system:

Broadly speaking, the system consists of three main components:

  1. Orchestrator: We will be implementing this is Python using the LangChain framework. This script will be responsible for taking in the user’s input and passing it to the retriever and LLM components along with some pre-defined constraints.
  2. Retriever: We will create tools that retrieve data from Sector’s, a financial API platform. These retrievers are in the form of some Python functions that take in some parameters and retrieve data accordingly from our data source (“Sectors API”).
  3. LLM: The language model we will be using is llama3-groq-70b-8192-tool-use-preview, which is state-of-the-art and finetuned specifically for tool use.

Llama-3-Groq-70B-Tool-Use is the highest performing model on the Berkeley Function Calling Leaderboard (BFCL), outperforming all other open source and proprietary models. Our models have achieved remarkable results, setting new benchmarks for Large Language Models with tool use capabilities:

  • Llama-3-Groq-70B-Tool-Use: 90.76% overall accuracy (#1 on BFCL at the time of publishing)
  • Llama-3-Groq-8B-Tool-Use: 89.06% overall accuracy (#3 on BFCL at the time of publishing)

​
Practice: Building a RAG System

Available as a Colab notebook

This tutorial is also available as a Colab notebook. Click here to access it.

For the remainder of this section, you will need to set up your environment along with the necessary dependencies. You can do this by running the following commands:

pip install requests
pip install langchain
pip install langchain-groq

In a new Python file, you should be able to initiate a request with your key:

import requests

headers = {
    # your Sectors API key
    "Authorization": "7c39xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
}

def get_overview(stock: str, section: str) -> requests:
    url = f"https://api.sectors.app/v1/company/report/{stock}/?sections={section}"
    response = requests.get(url, headers=headers)
    return response

response = get_overview("BBRI", "financials")
print(response.json())

Since we’ll be using a model hosted on Groq, we’ll also need an API key obtained from Groq. You can sign up for a free account here. With both keys in hand, you can start setting up our script and load the keys into our environment.

I’m using dotenv to load my keys from a .env file. You can install it by running pip install python-dotenv and create a .env file in the same directory as your Python script with the following content:

GROQ_API_KEY=your_groq_api_key
SECTORS_API_KEY=your_sectors_api_key

And our Python script should look like this:

import os
import json
import requests
from dotenv import load_dotenv

load_dotenv() # load your .env file
GROQ_API_KEY = os.getenv("GROQ_API_KEY")
SECTORS_API_KEY = os.getenv("SECTORS_API_KEY")

def retrieve_from_endpoint(url: str) -> dict:
    headers = {"Authorization": SECTORS_API_KEY}

    try:
        response = requests.get(url, headers=headers)
        response.raise_for_status()
        data = response.json()
    except requests.exceptions.HTTPError as err:
        raise SystemExit(err)
    return json.dumps(data)

Notice that retrieve_from_endpoint is just a convenience function that we can use to retrieve data from the Sectors API. We are creating an abstraction layer so we can re-use this function throughout our script.

​
Building tools for our tool-use model

Now that we have our data retrieval function, we can start building a set of tools, each tool specializing in handling a specific type of queries. For example, we can create a tool that retrieves financial data for a given stock, another tool that retrieves news articles for a given stock, and yet another tool might retrieve the financial reports for a given stock.

langchain provides a Tool class as well as a tool decorator that we can use to wrap any Python function and turn it into a tool. Here’s how we use it to create a grand total of three tools:

from langchain_core.tools import tool

@tool
def get_company_overview(stock: str) -> str:
    """
    Get company overview
    """
    url = f"https://api.sectors.app/v1/company/report/{stock}/?sections=overview"

    return retrieve_from_endpoint(url)

@tool
def get_top_companies_by_tx_volume(
    start_date: str, end_date: str, top_n: int = 5
) -> str:
    """
    Get top companies by transaction volume
    """
    url = f"https://api.sectors.app/v1/most-traded/?start={start_date}&end={end_date}&n_stock={top_n}"

    return retrieve_from_endpoint(url)

@tool
def get_daily_tx(stock: str, start_date: str, end_date: str) -> str:
    """
    Get daily transaction for a stock
    """
    url = f"https://api.sectors.app/v1/daily/{stock}/?start={start_date}&end={end_date}"

    return retrieve_from_endpoint(url)


tools = [
    get_company_overview,
    get_top_companies_by_tx_volume,
    get_daily_tx
]

​
Orchestrating our RAG system

We’ll be using the llama3-groq-70b-8192-tool-use-preview model from Groq. Not only is this model state-of-the-art and specializes in tool use, it is also open source and more easily accessible than other proprietary models.

In the following code, we instantiate a ChatGroq object with the model name and the Groq API key. This provides us with the llm object that we’ll pass, along with tools created earlier, to our orchestrator agent. What’s missing is the a prompt template that we’ll use to specify some system constraints and provide some overall guidance to the LLM.

from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain_groq import ChatGroq
from langchain.agents import create_tool_calling_agent, AgentExecutor

prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            """Answer the following queries, being as factual and analytical 
            as you can. If you need the start and end dates but they are not 
            explicitly provided, infer from the query. Whenever you return a 
            list of names, return also the corresponding values for each name. 
            If the volume was about a single day, the start and end 
            parameter should be the same."""
        ),
        ("human", "{input}"),
        # msg containing previous agent tool invocations 
        # and corresponding tool outputs
        MessagesPlaceholder("agent_scratchpad"),
    ]
)

llm = ChatGroq(
    temperature=0,
    model_name="llama3-groq-70b-8192-tool-use-preview",
    groq_api_key=GROQ_API_KEY,
)

agent = create_tool_calling_agent(llm, tools, prompt)
agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True)

From top to bottom, we’re performing a few key steps:

  1. We’re creating a ChatPromptTemplate object that contains a message for the system and a message for the user. The system message provides some guidance to the LLM on how to respond to the user’s input.
  2. We’re instantiating our LLM model. This LLM is responsible for interpreting the user’s input and choosing the right tool among the candidates to best respond to the user’s query.
  3. We combine llm, tools, and prompt into an agent object. This agent is responsible for orchestrating the interaction between the user, the LLM, and the tools.
  4. Finally, we create an AgentExecutor object that will execute the agent and handle the interaction between the user and the system. Setting verbose to True will print out the system’s responses to the user’s queries and additional messages that the system might generate.

​
Financial Queries with a custom RAG

Now is where the magic happens. We can start querying our system with financial queries and see how it responds. Each of the question is set up to require some form of tool use and actual data retrieval, failing which the system will not be able to generate a response.

query_1 = "What are the top 3 companies by transaction volume over the last 7 days?"
query_2 = "Based on the closing prices of BBCA between 1st and 30th of June 2024, are we seeing an uptrend or downtrend? Try to explain why."
query_3 = "What is the company with the largest market cap between BBCA and BREN? For said company, retrieve the email, phone number, listing date and website for further research."
query_4 = "What is the performance of GOTO (symbol: GOTO) since its IPO listing?"
query_5 = "If i had invested into GOTO vs BREN on their respective IPO listing date, which one would have given me a better return over a 90 day horizon?"

queries = [query_1, query_2, query_3, query_4, query_5]

for query in queries:
    print("Question:", query)
    result = agent_executor.invoke({"input": query})
    print("Answer:", "\n", result["output"], "\n\n======\n\n")

The first 3 queries should be relatively straightforward given the tools we’ve created in earlier steps.

For the remaining query_4 and query_5, you will need to implement additional tools that can retrieve historical performance of a stock since its IPO listing. This section: Company’s Performance since IPO of the API Documentation should provide you with the necessary information to extend your RAG system with this capability.

​
Financial Agent Responses

If implemented successfully, you should see the system generating an appropriate response for each of the query.

On query_1, we asked “What are the top 3 companies by transaction volume over the last 7 days?” and the system responded with:

> Entering new AgentExecutor chain...
Invoking: `get_top_companies_by_tx_volume` with `{'start_date': '2024-07-24', 'end_date': '2024-07-31', 'top_n': 3}`


{"2024-07-24": [{"symbol": "GOTO.JK", "company_name": "PT GoTo Gojek Tokopedia Tbk", "volume": 2660984600, 
"price": 54}, {"symbol": "BSBK.JK", "company_name": "PT Wulandari Bangun Laksana Tbk", "volume": 1833364900, 
"price": 83}, {"symbol": "DOOH.JK", "company_name": "PT Era Media Sejahtera Tbk", "volume": 511754100, 
"price": 53}], "2024-07-25": [{"symbol": "BSBK.JK", "company_name": "PT Wulandari Bangun Laksana Tbk", 
"volume": 1513766400, "price": 76}, {"symbol": "GOTO.JK", "company_name": "PT GoTo Gojek Tokopedia Tbk", 
"volume": 911091700, "price": 54}, {"symbol": "BBKP.JK", "company_name": "PT Bank KB Bukopin Tbk", 
"volume": 432596400, "price": 56}], "2024-07-26": [{"symbol": "BSBK.JK", "company_name": "PT Wulandari Bangun Laksana Tbk", 
"volume": 2408506000, "price": 72}, {"symbol": "GOTO.JK", "company_name": "PT GoTo Gojek Tokopedia Tbk", 
"volume": 913457100, "price": 53}, {"symbol": "MSJA.JK", "company_name": "PT Multi Spunindo Jaya Tbk.", 
"volume": 450977300, "price": 342}], "2024-07-29": [{"symbol": "GOTO.JK", "company_name": "PT GoTo Gojek Tokopedia Tbk", 
"volume": 1621366300, "price": 54}, {"symbol": "ATLA.JK", "company_name": "PT Atlantis Subsea Indonesia Tbk", 
"volume": 1255103900, "price": 50}, {"symbol": "BSBK.JK", "company_name": "PT Wulandari Bangun Laksana Tbk", 
"volume": 754254000, "price": 69}], "2024-07-30": [{"symbol": "GOTO.JK", "company_name": "PT GoTo Gojek Tokopedia Tbk", 
"volume": 1236811400, "price": 55}, {"symbol": "BSBK.JK", "company_name": "PT Wulandari Bangun Laksana Tbk", 
"volume": 797879400, "price": 69}, {"symbol": "BIPI.JK", "company_name": "PT Astrindo Nusantara Infrastruktur Tbk.", 
"volume": 629061500, "price": 59}]}

> Finished chain.

Answer: 
The top 3 companies by transaction volume over the last 7 days are PT GoTo Gojek Tokopedia Tbk, 
PT Wulandari Bangun Laksana Tbk, and PT Era Media Sejahtera Tbk.

To arrive at this answer, it would need to interpret “last 7 days” and infer the start and end dates for the query, given that the tool get_top_companies_by_tx_volume requires a start and end date to retrieve the data. It then correctly retrieves the data and have to perform a reduction operation to find the top 3 companies by transaction volume.

As a bonus, you may also want to provide a calculator tool that your RAG can use to perform the necessary summation and sorting operations to find the top 3 companies by transaction volume. In cases where it fail to interpret the intent correctly, your RAG system might instead return the top 3 entries for the last 7 days, which is probably a good approximation but not fully correct.

On query_2, we asked “Based on the closing prices of BBCA between 1st and 30th of June 2024, are we seeing an uptrend or downtrend? Try to explain why.” Our RAG model return with the following response:

> Entering new AgentExecutor chain...

To determine if we are seeing an uptrend or downtrend in BBCA\'s closing prices between June 1st and June 30th, 2024, we need 
to analyze the closing prices over this period. 

First, let\'s get the daily transaction data for BBCA for the specified period:

\```json
{
  "name": "get_daily_tx",
  "arguments": {
    "stock": "BBCA",
    "start_date": "2024-06-01",
    "end_date": "2024-06-30"
  }
}
\```

> Finished chain.
Answer: 

After analyzing the data, we can see that the closing prices have been increasing over the period, with the 
closing price on June 30th being higher than on June 1st. This indicates an uptrend in BBCA\'s closing prices 
during June 2024.

The uptrend could be due to various factors such as positive news or events related to the company, an increase 
in investor confidence, or a general upward trend in the market.

query_3 is considerably trickier as it requires the system to retrieve the market cap for both BBCA and BREN, compare them, and then retrieve additional information for the company with the largest market cap. Our original query is “What is the company with the largest market cap between BBCA and BREN? For said company, retrieve the email, phone number, listing date and website for further research.” and our RAG system responded with:

> Entering new AgentExecutor chain...

Invoking: get_company_overview with {'stock': 'BBCA'}
`{"symbol": "BBCA.JK", "company_name": "PT Bank Central Asia Tbk.", "overview": {"listing_board": "Main", "industry": "Banks", "sub_industry": "Banks", "sector": "Financials", "sub_sector": "Banks", "market_cap": 1263500995330048, "market_cap_rank": 1, "address": "Menara BCA, Grand Indonesia\r\nJalan MH Thamrin No. 1\r\nJakarta 10310", "employee_num": 27546, "listing_date": "2000-05-31", "website": "www.bca.co.id", "phone": "021-23588000", "email": "investor_relations@bca.co.id", "last_close_price": 10175, "latest_close_date": "2024-07-30", "daily_close_change": -0.00731707317073171}}`

Invoking: get_company_overview with {'stock': 'BREN'}
`{"symbol": "BREN.JK", "company_name": "PT Barito Renewables Energy Tbk.", "overview": {"listing_board": "Main", "industry": "Electric Utilities", "sub_industry": "Electric Utilities", "sector": "Infrastructures", "sub_sector": "Utilities", "market_cap": 1157248873136128, "market_cap_rank": 2, "address": "Wisma Barito Pacific II, Lantai 23,\nJl. Let. Jend. S. Parman Kav. 60, RT 010, RW 005, Slipi, Palmerah, Jakarta 11410,\nIndonesia", "employee_num": 634, "listing_date": "2023-10-09", "website": "www.baritorenewables.co.id", "phone": "(021) 530 6711", "email": "corpsec@baritorenewables.co.id", "last_close_price": 8650, "latest_close_date": "2024-07-30", "daily_close_change": 0.00581395348837209}}`

> Finished chain.

Answer: 
 The company with the largest market cap between BBCA and BREN is PT Bank Central Asia Tbk. (BBCA). Here are the details for further research:

- Email: investor_relations@bca.co.id
- Phone Number: 021-23588000
- Listing Date: 2000-05-31
- Website: www.bca.co.id

As for query_4 and query_5, you should expect a response similar to the one below:

Question: What is the performance of GOTO (symbol: GOTO) since its IPO listing?


Entering new AgentExecutor chain...
Invoking: get_performance_since_ipo with `{'stock': 'GOTO'}`

{"symbol": "GOTO.JK", "chg_7d": -0.0104712, "chg_30d": -0.418848, 
"chg_90d": -0.115183, "chg_365d": -0.740838}

Finished chain.
Answer: 
 The performance of GOTO (symbol: GOTO) since its IPO listing shows 
 a change of -0.418848 in the last 30 days, -0.740838 in the last 365 days, 
 -0.0104712 in the last 7 days, and -0.115183 in the last 90 days. 

======

Question: If i had invested into GOTO vs BREN on their respective IPO 
listing date, which one would have given me a better return over a 1 year horizon?


Entering new AgentExecutor chain...

Invoking: get_performance_since_ipo with {'stock': 'GOTO'}

{"symbol": "GOTO.JK", "chg_7d": -0.0104712, "chg_30d": -0.418848, 
"chg_90d": -0.115183, "chg_365d": -0.740838}

Invoking: get_performance_since_ipo with {'stock': 'BREN'}

{"symbol": "BREN.JK", "chg_7d": 1.82051, "chg_30d": 3.51282, 
"chg_90d": 5.92308, "chg_365d": null} 

Finished chain.
Answer: 
 Based on the performance data since their respective IPO listing dates, 
 BREN would have given you a better return over a 90-day horizon. 
 BREN had a 90-day change of 5.92308, while GOTO had a 90-day change 
 of -0.115183. 
======

​
Going Further: Improving the RAG System

Available as a Colab notebook

This tutorial is also available as a Colab notebook. Click here to access it.

There are numerous opportunities to improve our RAG system. Here are a few ideas to get you started:

  • Adding more tools: You can add more tools to your system to handle a wider range of queries. For example, you could add tools to retrieve news articles, analyst reports, or social media sentiment for a given stock.
  • Better prompts: You can experiment with different prompts to guide the LLM in generating more accurate and informative responses.
  • More descriptive tools: You can make your tools more descriptive by providing additional information about the data they retrieve. This helps your orchestrator agent make better decisions about which tool to use for a given query.
  • Different LLMs: You can try using different LLMs to see how they perform on your financial queries. You can experiment with different models from Groq or other providers. With how fast the field is evolving, it’s always a good idea to keep an eye on the latest models and see how they perform on your use case.
  • Error handling: You can add error handling to your system to handle cases where the LLM is unable to generate a response. For example, you could have the system prompt the user for more information or provide a default response.

In the spirit of motivating you to explore further, in the model that I have implemented for Generative AI for the Finance Industry workshop, my prompt template is more detailed and include the present date to help the LLM infer the start and end dates for the queries when user queries include time-sensitive information framed as “last 7 days”, or “since the start of the month” etc.

Here is my implementation:

def get_today_date() -> str:
    """
    Get today's date
    """
    from datetime import date

    today = date.today()
    return today.strftime("%Y-%m-%d")


prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            """
            Answer the following queries, being as factual and analytical as you can. 
            If you need the start and end dates but they are not explicitly provided, 
            infer from the query. Whenever you return a list of names, return also the 
            corresponding values for each name. If the volume was about a single day, 
            the start and end parameter should be the same. Note that the endpoint for 
            performance since IPO has only one required parameter, which is the stock. 
            Today's date is 
            """
            + get_today_date(),
        ),
        ("human", "{input}"),
        # msg containing previous agent tool invocations and corresponding tool outputs
        MessagesPlaceholder("agent_scratchpad"),
    ]
)

​
Challenge

Earn a Certificate

There is an associated challenge with this chapter. Successful completion of this challenge will earn you a certificate of completion and possibly extra rewards if you’re among the top performers.

If you’re up for a challenge, I have created two exercises that you can participate by making a copy of the Colab notebook and start working on them.

When you’re done, submit your work following this guide and I will be grading them.

1. Background: From GenAI to RAGs3. Structured Output AIs