Home / Current Issue / Paper 1715545
Tax Advisor AI: An Agentic RAG Framework for Indian Income Tax Guidance using Knowledge Graphs and Hybrid Retrieval
Subject area: Science,Engineering and Technology · Area of research: AI for financial and legal applications
Abstract
Starting off, understanding India’s tax law from 1961 isn’t simple - for regular folks or businesses. Because updates happen all the time, things get harder. Then there’s the split setup now: one old method, one new. That adds confusion. On top of that, sorting out deductions listed in Chapter VI-A involves subtle differences people miss. Most search tools just hunt for words, so they miss what someone actually means to ask. Meanwhile, big AI models talk well but make up rules that don’t exist - like mixing up how much you can claim under Section 80C compared to 80CCD Meet Tax Advisor AI - a smart tool built just for Indian tax questions. It gives clear answers backed by law, each one linked to real sources. At its core runs something new: Agentic RAG, not your usual search setup. Instead of just pulling documents, it thinks through problems step by step using ReAct logic, guided by the DeepSeek R1T engine. Queries get split into smaller pieces so nothing gets missed. For finding facts, it uses two methods at once - deep meaning scans plus keyword matching - blended smartly with RRF scoring. Beneath the surface lies a Neo4j graph, a web of connections that prevents conflicting rules from colliding. This system, when new information arrives, taps into live web sources, specifically the Tavily API, to ensure it reflects the latest official pronouncements. Tests demonstrate its ability to translate complex legal language into straightforward financial advice, outperforming simpler data retrieval techniques.
Keywords
Agentic RAG, Knowledge Graphs, DeepSeek, Indian Taxation, Hybrid Retrieval, Neo4j, Natural Language Processing, FinTech, Legal Tech.
References
[1] I. Chalkidis et al., ”LEGAL-BERT: The Muppets straight from Big Bird’s Mouth,” in Findings of EMNLP, 2020.
[2] P. Lewis et al., ”Retrieval-Augmented Generation for Knowledge- Intensive NLP Tasks,” in NeurIPS, 2020.
[3] S. Yao et al., ”ReAct: Synergizing Reasoning and Acting in Language Models,” in ICLR, 2023.
[4] DeepSeek-AI, ”DeepSeek LLM: Scaling Open-Source Language Mod- els,” arXiv preprint arXiv:2401.02954, 2024.
[5] H. Chase, ”LangChain: Building applications with LLMs through composability,” 2022. [Online]. Available: https://github.com/langchain- ai/langchain.
[6] Income Tax Department, ”Income Tax Act, 1961,” Government of India. [Online].
[7] J. Devlin et al., ”BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,” in NAACL-HLT, 2019.
[8] Y. Zhang et al., ”GraphRAG: Retrieving Knowledge Graphs for Open- Domain Question Answering,” arXiv preprint, 2023
How to cite this paper
@article{1715545,
author = {Rehan Rajesh, Nivin Roy, Harikrishnan S, Gautham P, Priya Jose},
title = {Tax Advisor AI: An Agentic RAG Framework for Indian Income Tax Guidance using Knowledge Graphs and Hybrid Retrieval},
journal = {Iconic Research And Engineering Journals},
year = {2026},
volume = {9},
number = {9},
pages = {2581-2588},
issn = {2456-8880},
url = {https://www.irejournals.com/formatedpaper/1715545.pdf},
abstract = {Starting off, understanding India’s tax law from 1961 isn’t simple - for regular folks or businesses. Because updates happen all the time, things get harder. Then there’s the split setup now: one old method, one new. That adds confusion. On top of that, sorting out deductions listed in Chapter VI-A involves subtle differences people miss. Most search tools just hunt for words, so they miss what someone actually means to ask. Meanwhile, big AI models talk well but make up rules that don’t exist - like mixing up how much you can claim under Section 80C compared to 80CCD Meet Tax Advisor AI - a smart tool built just for Indian tax questions. It gives clear answers backed by law, each one linked to real sources. At its core runs something new: Agentic RAG, not your usual search setup. Instead of just pulling documents, it thinks through problems step by step using ReAct logic, guided by the DeepSeek R1T engine. Queries get split into smaller pieces so nothing gets missed. For finding facts, it uses two methods at once - deep meaning scans plus keyword matching - blended smartly with RRF scoring. Beneath the surface lies a Neo4j graph, a web of connections that prevents conflicting rules from colliding. This system, when new information arrives, taps into live web sources, specifically the Tavily API, to ensure it reflects the latest official pronouncements. Tests demonstrate its ability to translate complex legal language into straightforward financial advice, outperforming simpler data retrieval techniques.},
keywords = {Agentic RAG, Knowledge Graphs, DeepSeek, Indian Taxation, Hybrid Retrieval, Neo4j, Natural Language Processing, FinTech, Legal Tech.},
month = {March},
doi = {https://doi.org/10.64388/IREV9I9-1715545}
}