The application support and global product assurance teams of a large global bank faced numerous challenges in delivering efficient and timely technical support as they had to manually identify solutions to recurring problems within the Known Error Database (KEDB), comprised of documents in various formats. With the high volume of support requests and limited availability of teams across multiple time zones, a large backlog of unresolved issues developed, leading to higher support costs.

Our team developed a conversational assistant using Gen AI by:

• Building an interactive customized React-based front-end
• Ringfencing a corpus of problems and solutions documented in the KEDB
• Parsing, formatting and extracting text chunks from source documents and creating vector embeddings using Sentence–Bidirectional Encoder Representations from Transformers (S-BERT)
• Storing these in a Facebook AI Similarity Search (FAISS) vector database
• Leveraging a local Large Language Model (Llama-2-7B-chat) to generate summarized responses

The responses generated using Llama-2-7B LLM were impressive and significantly reduced overall effort. Future enhancements to the assistant would involve:

• Creating support tickets based on information collected from users
• Categorizing tickets based on the nature of the problem
• Automating repetitive tasks such as access requests / data volume enquiries / dashboard updates
• Auto-triaging support requests by asking users a series of questions to determine the severity and urgency of the problem