Internal Revenue Manual Research Aid (IRMA)
Description
The Internal Revenue Manual (IRM) Research Aide (IRMA) is a generative AI (GenAI) system that uses Retrieval Augmented Generation (RAG) search to answer questions on IRM policies and procedures that are asked in natural language. This tool has a custom user interface (UI) and utilizes a COTS platform via Application Programming Interface (API) to perform the RAG search and generate responses based on the full text of the IRM serving as a knowledge base. The generated responses also provide links back to the referenced sections of the IRM for further review.
Detailed example
This system uses an embedding model to produce a vector representation of texts for contextual searching and a Large Language Model (LLM) to generate a response to user inquiries. Outputs of the tool include the model generated response to a user's inquiry, as well as the text and links to relevant references identified in the retrieval stage and used by the model to generate the response.
AI / analytics pattern
Generative AI: AI that generates new or synthetic content (e.g., images, videos, audio, text, code).
Automation level / stage
b) Pilot – The use case has been deployed in a limited test or pilot capacity.
Expected benefit
Operational efficiency and reduced cost by providing the necessary information to taxpayer-facing customer service agents to quickly answer taxpayers' questions. This enables the agents to increase their capacity and help a larger number of taxpayers by reducing the time and effort needed to research responses. The public, in turn, benefits from a reduction in wait time to receive a response to their inquiries.
Audit / financial statement impact
The output is not presumed to be high-impact and is not used as the principal basis for significant decisions/actions
Controls / human review
ATO: Yes; PIA: Not published
Data needed
The model is trained and refined on the full text of the internal IRM, including the sections marked "Official Use Only (OUO)" that are redacted in the public version. The model is evaluated against a standard set of benchmarking questions whose responses have been provided and validated by subject matter experts (SMEs) with the relevant domain-knowledge.