How Amazon Finance Automation built a generative AI Q&A chat assistant using Amazon Bedrock

Amazon Finance Automation developed a generative AI Q&A chat assistant using Amazon Bedrock to help Accounts Payable and Accounts Receivable analysts quickly retrieve answers to customer queries.

• Initially, the solution had only 49% response accuracy
• Improved accuracy to 86% through several key strategies:
  • Semantic document chunking (49% to 64% accuracy)
  • Advanced prompt engineering (64% to 76% accuracy)
  • Using Amazon Titan Text Embeddings model (76% to 86% accuracy)
• Key components include:
  • Amazon OpenSearch Service as knowledge base
  • Amazon Titan Multimodal Embeddings model
  • Amazon Bedrock Guardrails for safety
  • Streamlit-based chat assistant UI

The solution dramatically reduces time spent researching customer queries by providing rapid, accurate responses using retrieval augmented generation (RAG).