As a part of the credit adjudication process for a transaction, commercial bankers use an application to create summaries, memos and rating alerts as needed, which are instrumental for ongoing Capital at Risk (CaR) monitoring, Risk Profiling, Risk Adjusted Return on Capital (RAROC) computations, etc.

There is a significant amount of complexity involved in understanding this application due to the diversity in types of borrowers / loans, nature of collaterals, etc., e.g., How to create a transaction report for my deal? How to update an existing deal?

All of this information is spread across multiple user guides and FAQ documents that may run into hundreds of pages.

• Ringfenced a knowledge base comprised of the user guides of various functionalities (e.g., facility creation, borrower information, etc.)
• Built a custom-developed, React-based front-end for the conversational assistant to interact with the users
• Parsed, formatted and extracted text chunks from these documents using libraries such as PDF Miner, PyPDF2
• Created vector embeddings using sentence transformer embedding model (all-MiniLM-L6-v2) and stored as indices in the Facebook AI Similarity Search (FAISS) vector database
• Broke down the user query into vector embeddings, searched against the vector database and leveraged local LLM (Llama-2-7B-chat) to generate summarized responses based on the context passed to it by the similarity search

Our custom solution was a conversational agent built using Generative AI, which summarizes relevant information from multiple documents.

It significantly:

• Improved existing users’ ability to access relevant information on a timely basis
• Simplified the migration of bankers and integrations of lending applications resulting from merger or acquisition