Invoice information extraction assistant
Chile - Chamber of Deputies
Use case ID: 055
Author: Chamber of Deputies of Chile
Date: 20 August 2024
Objective:
Extract relevant information from expense accreditation documents such as receipts or invoices. This use case describes part of a suite of AI-based products. The CAMINAR platform supports legislative work with the help of AI. It emerged as an innovative response to the challenges faced by the National Congress of Chile during the COVID-19 pandemic and owing to the country’s political situation. The project seeks to leverage the digital strengths of the Chamber of Deputies by utilizing its regulatory repository and parliamentary databases.
Actors:
- Officials: to access an efficient and intuitive tool to support the regulation of public fund usage
Prerequisites:
- Database
- Vectorized regulations documentation
- Formatted files of expense accreditation documents
- Appropriate prompt
- System to integrate the tool into its production phase
Scenario:
- Officials involved in approving expenses related to parliamentary work receive a proposed endorsement result accompanied by the extracted information. They validate the proposal by indicating data correctness. The goal is to quickly extract the document, store it rapidly and make it available for consumption by other assistants.
Expected results:
- The time taken to extract data from documents is reduced.
- The extraction error rate is reduced owing to a second review.
- Accuracy rates are quantified for image data extraction technologies.
Potential challenges:
- Need for human validation at all times
- Possible error margins (independent of LLMs) in image data extraction technologies
Data requirements:
- Updated database for assistant consumption
- LLM’s own knowledge
Integrations with other systems:
- Parliamentary allowances management system (ASIGPAR)
Success Metrics:
- Accuracy of the response, when correctness can be determined
- Speed of the response, related to the requirements of repetitive tasks
- Impact on post-implementation operational performance variables