Internal Revenue Manual Research Aid (IRMA)

Overview

Development stage

PilotSource: b) Pilot – The use case has been deployed in a limited test or pilot capacity.

High-impact designation

Not high-impactSource: c) Not high-impact

Impact justification

The output is not presumed to be high-impact and is not used as the principal basis for significant decisions/actions

Start date

2025-04-01 (day precision)Source: 4/1/2025

Withheld from public reporting

NoSource: a) No

Mission

Topic area

Information Technology

Operational functions

Chatbot or virtual assistant, Search and knowledge retrieval Derived by keyword rules; see methodology

Problem the AI is intended to solve

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.

Expected benefits

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.

System outputs

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.

Technology

AI classification

Generative AISource: Generative AI: AI that generates new or synthetic content (e.g., images, videos, audio, text, code).

System name(s)

Not reported

Custom-developed code

Yes

Public source code

Not reported

Sourcing

How it was built

Contract and in-houseSource: c) Developed with both contracting and in-house resources

Vendor (as reported)

Not reported

Vendors (standardized)

None on the standard list

A vendor is the supplier named by the agency. It does not identify the underlying model or AI technology.

Data and privacy

Involves PII

No

Privacy Impact Assessment

Not reported

Authorization to Operate

Yes

Demographic features

k) None of the above

Training and evaluation data

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.

Federal Data Catalog

Not reported

Governance

0 of 8 minimum-practice questions answered. A blank answer means the agency reported nothing; it does not mean the practice is absent.

Potential impacts and how they were identified

Not reported