A bank wants customers to ask questions about their scanned mortgage statements and receive answers with amounts pulled from specific form fields. The statements arrive as image-based PDFs. Which pipeline is most appropriate?
Show answer and explanation
Answer: B
Scanned statements are images, so text, tables, and labeled fields must be extracted before a retrieval system can use them. Azure AI Document Intelligence performs that extraction and returns structured content. Indexing the extracted content with source references then lets the assistant answer from specific fields and cite them. Sending raw scans or relying on image similarity skips the step that makes the figures searchable.
Why the other options are wrong
- A. A chat model cannot reliably read every figure on a scanned page without an extraction step. Sending raw PDFs skips the structure needed for field-level answers.
- C. File names say nothing about the amounts inside a statement. Asking the model to guess would produce unsupported figures.
- D. Image similarity finds visually alike pages, not the specific numbers a customer asks about. It cannot return reliable field values.