As every company strives to find ways to implement ChatGPT in their respective contexts, we spent months in stealth mode developing and refining a model built on top of ChatGPT specifically designed to address the unique needs of financial services firms, all while maintaining the highest standards of accuracy, privacy, and security.
When we first started developing the model, we didn't think that this would become the biggest challenge. As you are currently reading this, we assume that you have played around with ChatGPT. Except when asking about current events, the model always tries to answer any question and can even seem opinionated at times. However, in our initial experiments with banks, this behaviour proved to be a significant obstacle. A model that suggests divorce as a tax-saving strategy is, obviously, unusable. Our AI Mid-Layer makes use of additional Classifiers and Rankers to make the model decline to provide inappropriate responses.
To train a large language model with company-specific knowledge, a significant amount of data pre-processing is required. Essentially, the cleaner and more structured the data is, the better the model's performance will be. For instance, as with any other large language model, ChatGPT has a limit on the amount of data that can be processed in a single feed. For example, gpt-3.5-turbo has a token limit of 4096 tokens, which includes both the question and answer. Our model was trained to retrieve and rank the most relevant pieces of content to make ChatGPT give a better answer. Additionally, to ensure that user queries were placed within the appropriate context of Swiss banking, the model needed to be initially trained with general knowledge about the industry.
After training the model to a specific context and teaching it to say "I don't know", we needed to make sure to run it efficiently. When exposing the model to end-users, a company is paying per request. We had to identify and eliminate out-of-context user queries before they get sent to ChatGPT and cosume tokens. This also helped us tackle challenge N°1 and continuously teach the model to say no. To scale the model for use by multiple users and parallel instances, we rely on Azure Cloud Services.
We soon understood that just sending an API call to a ChatGPT model is not going to be enough. Our AI Mid-Layer builds upon the strong capabilities of ChatGPT by adding the necessary components to make it work for financial services firms. We can train the model with a company specific context and make it work within pre-defined boundaries.
Our goal is to help clients experiment with large language models without having to start from scratch. We offer a modular easy-to-integrate AI Mid-Layer that can be used in many contexts. Imagination is the limit.
Traditional chatbots listen to trigger words and follow a pre-defined decision tree. Adding Natural Language Processing (NLP) to make sense of a user question and generate an adequate answer based on a trained context is a game changer. Our ChatGPT powered Financial Chatbot keeps track of chat history and combines NLP and decision tree-based conversations to achieve a stronger performance.
We use the summarising capabilities of ChatGPT to generate summaries of financial news adapted to the level of financial literacy of the target user. Our model can generate a personalised news summary based on your investment portfolio and tailored to your financial expertise level.
Imagine you can ask a question about a FINMA regulation, get back an answer with the source and be able to ask a follow up question to better understand a complex regulatory matter. Our model is trained with all FINMA publicly available documents and makes use of chat history to put your follow up questions into context.
Our conversational model builds upon the capabilities of ChatGPT to achieve a human-like onboarding experience. Our hybrid architecture keeps a conversation going without sending client data to ChatGPT. Delegate low-value tasks to a conversational bot and augment your user experience.