InvestmentAdvisorGPT

Your friendly investment advisor has now turned into an LLM chatbot!

This project demonstrates the usage of custom agent with knowledge base and tools to generate insightful conversations with regards to investment-related topics and products.

The chatbot demonstrated in this repo exhibits certain persona defined aimed to conduct a sale outreach with users in attempt to promote and pitch certain financial products e.g., mutual funds embedded within our knowledge base. Conversation flows are also depended on previous conversation history. The attempt to generate context-aware sales conversation is adapted from SalesGPT, where the chatbot has the ability to shift conversation topics with minimal prompting from the user.

Tables of Contents

• Installation
  • Tools and LLM Setup
• Quick Start
  • UI
• Architecture
  • Conversation Context
  • Change In Conversation Flow
  • Knowledge Base and Tools
• Known Issues
• Future Works

Installation

The application consisting of the UI and server is container-ready and deployed via the docker-compose file.

To start the application, simply use the following command:

docker-compose up -build

To stop the application and clean up, use the following command:

docker-compose down --remove-orphans

Tools and LLM Setup

To ensure the full working experience of the chatbot, you will need to have the following API keys:

1. OpenAI via @ https://platform.openai.com/
   • This powers the underlying LLM model (ChatOpenAI).
2. Serper via @ https://serper.dev/
   • This powers the underlying web search capability (GoogleSerperAPIWrapper)

Once you have obtained the relevant API keys, you can stored them in a .env file, in the following format:

OPENAI_API_KEY="fill me in"
SERPER_API_KEY="fill me in"

Quick Start

Using your terminal:

# user_id can be anything, it just to use to create
# and identify your agent to retrieve relevant conversation history/context
export user_id="jensenlwt"

# initial message is empty as the bot will chat with you first
curl -X POST http://localhost:8000/chat/$user_id \
     -H "Content-Type: application/json" \
     -d '{"message": ""}'

>>> {
    "user_id":"jensenlwt",
    "raw_message":"Bobby Axelrod: Hi Jensen! How are you doing today? <END_OF_TURN>"
    }

# asking about new opportunities
curl -X POST http://localhost:8000/chat/$user_id \
     -H "Content-Type: application/json" \
     -d '{"message": "I am doing good Bobby! Do you have any new investment opportunities for me?"}'

>>> {"user_id":"jensenlwt",
    "raw_message":"Bobby Axelrod: Hi Jensen! I'm glad to hear that you're doing good. Yes, I do have some exciting new investment opportunities for you. We have recently launched various mutual funds from Abrdn that might be of interest to you. These funds are professionally managed by different fund managers and designed to meet your specific investment needs. Would you like me to provide more information about these funds? <END_OF_TURN>"
}

# seeking new information
curl -X POST http://localhost:8000/chat/$user_id \
     -H "Content-Type: application/json" \
     -d '{"message": "Sure! can you kindly provide me with more information?"}'

>>> {
  "user_id": "jensenlwt",
  "raw_message": "Bobby Axelrod: Absolutely, Jensen! Here are a few mutual funds from Abrdn that you might find interesting:\n\n1. Abrdn Global Technology Fund: This fund focuses on investing in global technology companies, aiming to capitalize on advancements in the tech sector.\n\n2. Abrdn Emerging Markets Equity Fund: This fund targets emerging market economies and invests in high-potential companies within those markets, aiming for long-term capital growth.\n\n3. Abrdn Sustainable Equity Fund: This fund invests in companies that actively contribute to sustainable development and environmental responsibility, aiming to generate both financial returns and positive impact.\n\n4. Abrdn Global Bond Fund: This fund focuses on fixed-income securities from around the world, aiming to provide stable income and potential capital appreciation.\n\nThese are just a few options available, Jensen. Let me know if any of these funds caught your interest, or if you'd like more information on any specific one. <END_OF_TURN>"
}

UI

In addition to using API calls, a UI is provided. The UI is developed using Chainlit that allows for fast iteration when building LLM app interfaces.

To visit the UI, you may access @ http://localhost:8080 after completing the relevant installation.

In any case where the UI takes some time to load on the browser, please docker-compose down --remove-orphans && docker-compose up --build again as the first call to the chat endpoint will be affected and will not display accordingly.

Architecture

The overall architecture of the application is as observed:

Conversation Context

The InvesmentAdvisorGPT agent requires certain configs to power basic conversation capability. Below contains an example set of configs found in src/model/configs/examples/agent_singaporean_male.json:

NOTE: You can change any configs here!

{
    "advisor_name": "Bobby Axelrod",
    "advisor_role": "private wealth advisor",
    "nationality": "Singaporean",
    "formal_language": "english",
    "informal_language": "singlish",
    "company_name": "UOB",
    "company_business": "provide unit trusts professionally managed by various fund managers, designed to meet customers' specific investment needs",
    "conversation_purpose": "find out if the prospect is interested in the latest investment products, specifically various mutual funds from Abrdn",
    "conversation_type": "text",
    "source_of_contact": "investment seminar",
    "prospect_name": "Jensen Low",
    // these below can remain untouched
    "conversation_stage": "Introduction: Begin the cold call with a warm self-introduction. Include your name, company, and a credibility statement or reason for the prospect to stay engaged.",
    "conversation_history": [],
    "use_tools": true,
    "verbose": true
}

Explaination of Configs:

1. advisor_name: Name of your agent, so it doesn't address itself as a robot/llm.
2. advisor_role: Role for your agent, so it knows it should behave in the capacity of the role.
3. nationality: Attempt to enforce behaviour from a nationality (doesn't seem to work as expected, lacklustre)
4. formal_language and informal_language: Attempt to allow for a switch-up the way the conversational tone sounds depending on the context of the conversation (fluctuate between being very professional and too overly casual (e.g., "hey bro!", "... bro!"))
5. company_name and company_business: Allows the agent to identify himself appropriate or introduces himself by identifying his work association.
6. conversation_purpose: Drives the overall objective of the conversation and all conversation flow stems from the purpose.
7. source_of_contact and prospect_name: Refers to the user we are reaching out to/or reaching out to us. This attempts to ground the agent with regards to a certain level of familarity with the user and not hallucinating and making the premise of the conversation random.

Lastly, conversation_history, conversation_stage seeds the initial conversation. While use_tools controls whether to use a agent with tools (product search, web search) or just default prompts (greater potential to hallucinate), and verbose set the verbosity of the Agents.

NOTE: use_tools=False remains buggy at the moment, and the conversation does not flow as smoothly due to the conversation history being clustered with double/blank inputs (both ai and user-alike). Hence, highly recommend to set use_tools=True for the time-being.

Change In Conversation Flow

As demonstrated in SalesGPT, custom agents can have the ability to change conversational topics or guide the flow of conversation if the conversation history is analyzed and identified for what stage a sales call might be in. Adapting the ideas, a series of conversation stages were determined based on several research on cold-calls strategy and workflow to better capture how the conversation might panned out.

The stages are as followed:

1. Introduction: Begin the cold call with a warm self-introduction. Include your name, company, and a credibility statement or reason for the prospect to stay engaged.
2. Confirm: This is an important next stage right after [Introduction] to confirm if the prospect is the right person to discuss financial products/services. Check their age and authority for making financial decisions."
3. Understanding the Prospect (Repeatable): Ask open-ended questions multiple times to uncover the prospect's financial needs and situation. Repeat this stage until you have gathered sufficient background information. Attempt to figure out what life stage they are currently in, and if they have any major life events happening soon that may impact their finances. Listen attentively. You are to infer the prospect's financial ability in terms of income, expenditure and financial aspiration.",
4. Huge Claim: Present an attention-grabbing claim related to the product/service. Connect it to the prospect's background in [Understanding the Prospect] discussed earlier.
5. Product Introduction: Introduce some of the products you have that may best suit the prospect's background and needs (inferred in from [Understanding the Prospect]). If unsure of their needs, repeat [Understanding the Prospect] and ask more questions to generate a more informed understanding of the prospect.
6. Value Proposition: Explain how our financial products/services benefit the prospect. Focus on their needs and emphasize unique selling points.",
7. Addressing Doubts: Handle skepticism about previous claims or product presentation. Provide evidence or testimonials.
8. Closing: If the prospect is demonstrating keenness/enthuasisiam in your financial products/services, invite the prospect for a further discussion or meeting. Suggest potential dates and times.
9. End conversation: The prospect has to leave to call, the prospect is not interested, or next steps where already determined by the sales agent.

A chain (ConversationStageAnalyzerChain) is designed to analyzed the conversation history up till the latest point and retrieve the most relevant stage and feed back into the agent's prompt as a guideline on how to response.

Knowledge Base and Tools

To miminize hallucination from our agent, we also utilised a knowledge base based on embeddings generated from various investment-related documents. Referring to the architecture diagram, you can see that the agent leverages retrieval from the knowledge based to augment its thoughts and responses.

These documents used include prospectus sheet, product highlight sheets from Abrdn (just randomly chose to use this company's product for demonstration purposes). The data is ingested into the knowledge base when placed in the project root directory (./data) before the application starts up, and when the (./db) directory is absent. The ./db directory is automatically created when you first ingest documents into the vector store (ChromaDB). Therefore, if the ./db directory is available, no data ingestion will occur (preventing constant calls for embeddings and associated costs). In this repository, the knowledge base is already provided. Hence, no embedding costs will be incurred if you choose to try out the application.

How the knowledge based is being leveraged is based on usage of custom tools. There two main tools available to our agent, ProductSearch and WebSearch. As the name implies, product search allows for search into our knowledge base to make more informed responses. Meanwhile, web search allow the agent to conduct latest news retrieval if one were to ask about any particular development in any country of interest that might help to drive a conversation on further investments.

Example of using both ProductSearch and WebSearch in a conversation:

Known Issues

1. ConversationBufferwindowMemory remains buggy is currently duplicating the agent's response twice in the row as observed in our conversation history:

   **Current Conversation Stage Context:**
   web_1  |     Product Introduction: Introduce some of the products you have that may best suit the prospect's background and needs (inferred in from [Understanding the Prospect]). If unsure of their needs, repeat [Understanding the Prospect] and ask more questions to generate a more informed understanding of the prospect.
   web_1  |     **Conversation History:**
   web_1  |     Jensen Low:
   web_1  |     Bobby Axelrod: Hey Jensen, it's Bobby Axelrod from UOB. Remember we met at the investment seminar? How's everything going?
   web_1  |     Bobby Axelrod: Hey Jensen, it's Bobby Axelrod from UOB. Remember we met at the investment seminar? How's everything going? <END_OF_TURN>
   web_1  |     Jensen Low: I am good. Thanks for asking. <END_OF_TURN>
   web_1  |     Jensen Low:
   web_1  |     Bobby Axelrod: Hey Jensen, glad to hear that you're doing well. I hope you found the investment seminar informative. Speaking of investments, have you considered exploring the latest mutual funds from Abrdn? They offer a range of professionally managed funds that can help meet your specific investment needs. Would you be interested in learning more about them? <END_OF_TURN>
   web_1  |     Bobby Axelrod: Hey Jensen, glad to hear that you're doing well. I hope you found the investment seminar informative. Speaking of investments, have you considered exploring the latest mutual funds from Abrdn? They offer a range of professionally managed funds that can help meet your specific investment needs. Would you be interested in learning more about them? <END_OF_TURN>
   web_1  |     Jensen Low: Yea sure, why not? <END_OF_TURN>
   

   In addition, blank user's message is also added into the conversation history. More work is required to debug the usage for this particular memory as it is crucial, considering the existing length of our prompt template, requiring the need for ease in pruning conversation history to ensure we do not exceed the token limit.
2. LLM often faced issues with its output parser when using a tool, as it does not follow the reAct prompting. In this repo, we setup the reAct prompt template for tools usage as:

   ### Tools ###
   
   To fulfill these certain objectives, {advisor_name} has access to the following tools:
   
   {tools}
   
   To use a tool, use the following format:
   
   # we use triple backtick in actual prompt
   # used two here to prevent codeblock formatting issue
   ``
   Question: the input question you must answer
   Thought: you should always think about what to do
   Action: the action to take, should be one of [{tools}]
   Action Input: the input to the action
   Observation: the result of the action
   ... (this Thought/Action/Action Input/Observation can repeat N times)
   Thought: I now know the final answer
   Final Answer: the final answer to the original input question
   ``
   
   If the result of the action is "I don't know." or "Sorry I don't know", then you have to say that to the user.

   However, the LLM still sometimes return back response such as "Thought: Search the web for xxx...", without completing the whole general reasoning framework. This results in the LLM to output weird response to a valid query. Currently, the only way to deal with it is through repeating the query/question in a slightly different way or moving on entirely.

Future Works

• Refine existing prompt to better exhibit persona and reduce LLM output parsing error when using tools.
• Enable easier configuration of conversation context e.g., prospect_name, advisor_name etc.
• Change agent architecture to better allow the usage of ConversationBufferWindowMemory instead of using stand-in method of removing conversation from a list (or using a list to store conversation history).
• Leverage on a better storage, insted of a using a dict object to store agent(s) in our server. Considering using a cache (redis) to cache agent and demonstrate ability to handle multiple concurrent conversations.