A leading North American bank had large commercial lending and credit risk units. Analysts in those units typically refer to numerous sections in a financial statement, including balance sheets, cash flows, and income statements, supplemented by footnotes and leadership commentaries, to extract decision-making insights. Switching between multiple pages of different documents took a lot of work, making the analysis extra difficult.

Many tasks were automated using Gen AI tools. Our steps:

• Ingest multiple URLs of financial statements
• Convert these to text using the PDF Reader library
• Build vector indices using LlamaIndex
• Create text segments and corresponding vector embeddings using OpenAI’s API for storage in a multimodal vector database e.g., Deep Lake
• Compose graphs of keyword indices for vector stores to combine data across documents
• Break down complex queries into multiple searchable parts using LlamaIndex’s DecomposeQueryTransform library

The solution delivered impressive results in financial analysis, notably reducing manual efforts when multiple documents were involved. Since the approach is still largely linguistic in nature, considerable Prompt engineering may be required to generate accurate responses. Response limitations due to the lack of semantic awareness in Large Language Models (LLMs) may stir considerations about the usage of qualifying information in queries.