Loading article...
Loading article...
Stream/Area:
-
Subject:
-
Paper ID:
IJSREM65877
Volume:
10 |
7 |
Jul 2026
Pages:
1-5
Publication Frequency:
Monthly (12 issues per year)
License:International Journal of Scientific Research in Engineering and Management follows the Creative Commons Attribution 4.0 (CC BY 4.0) License, allowing free use and distribution with proper attribution.
Department of CSE & CY
Lakshmi Narain College of Technology & Science, Bhopal
Department of AI-ML
Sagar Institute of Research & Technology, Bhopal
Retrieval-Augmented Generation (RAG) attempts to mitigate the knowledge limitations of Large Language Models (LLMs) by letting them tap into outside sources. But in practice, things rarely run that smoothly. In real-world deployments, the assumption of uniform consistency across retrieved documents is often invalid. Pull from a mix of databases, and pretty quickly you end up with overlapping ideas and sources flat out contradicting each other. Passing conflicting evidence directly into an LLM typically results in hallucinations or factually incorrect responses. To resolve these inconsistencies systematically, we propose a modular, multi-agent system with LangGraph that puts some order into the chaos. Our architecture breaks the process into distinct steps, using eight agents to reshape raw text into clear factual statements, check for logical clashes, and weigh how trustworthy each source is, all based on the metadata. We score credibility by looking at things like how current the source is and its overall authority, then filter out the contradictions before the language model gets a final shot at generating answers. We put this framework to the test on a hand-picked set of 150 document pairs, all loaded with contradictions across legal, financial, and technical topics. The agent-based filtering boosted factual consistency up to 26.5% over Naive RAG compared to baseline RAG setups. Conclusion: integrating a dedicated contradiction-checking layer significantly improves the factual reliability of LLMs and a solid foundation for building applications people can trust.
KEYWORDS
This paper lays out a straightforward multi-agent system that adds a contradiction detection and resolution layer right between retrieval and generation in RAG pipelines. Instead of piling everything onto one big model, it splits the work fact extraction, contradiction checks, credibility scoring, and resolution across distinct agents using LangGraph, each with a specific job. With this setup, the system consistently catches more contradictions and produces more accurate answers than Naive RAG, CoT RAG, or a simple Self-RAG baseline, and it doesn’t slow things down enough to be a problem for internal, accuracy-focused projects. There’s no heavy math or complicated models here just clear, understandable heuristics. That means an undergrad at the end of their degree could actually design, build, and test this system in a single semester. Still, it leaves plenty of room for building out better, more trustworthy, multi-source RAG systems in the future.
[1] P. Lewis et al., “Retrieval-augmented generation for knowledge-intensive NLP tasks,” NeurIPS, vol. 33, pp. 9459–9474, 2020. [2] Z. Ji et al., “Survey of hallucination in natural language generation,” ACM Computing Surveys, vol. 55, no. 12, 2023. [3] Y. Gao et al., “Retrieval-augmented generation for LLMs: A survey,” arXiv:2312.10997, 2023. [4] T. Wu et al., “AutoGen: Enabling next-gen LLM applications via multi-agent conversation,” arXiv:2308.08155, 2023. [5] K. Shuster et al., “Retrieval augmentation reduces hallucination in conversation,” Findings of ACL: EMNLP, 2021. [6] V. Karpukhin et al., “Dense Passage Retrieval for Open-Domain QA,” Proc. EMNLP, pp. 6769–6781, 2020. [7] J. Liu et al., “Lost in the Middle: How Language Models Use Long Contexts,” arXiv:2307.03172, 2023. [8] A. Asai et al., “Self-RAG: Learning to Retrieve, Generate, and Critique through SelfReflection,” arXiv:2310.11511, 2023. [9] Z. Jiang et al., “Active Retrieval Augmented Generation,” Proc. EMNLP, 2023. [10] A. Madaan et al., “Self-Refine: Iterative Refinement with Self-Feedback,” NeurIPS, 2023.
Publication Date
7/8/2026
Peer Review Type
Double-Blind Peer Review
Review Rounds: 2
Documents
Article Views
1
Total views since publication
PDF Downloads
14
Total downloads since publication
Geographic Reach
5
Digital Preservation: This article is permanently archived on Zenodo and complies with all data preservation standards across major indexing platforms.
Plagiarism Screening
This manuscript has been checked for plagiarism using TurnitinPlagiarismChecker software.
Conflict of Interest
The authors declare no conflict of interest related to this research work.
Funding Information
No funding information available for this research.
Related research in the same subject area
Loading similar articles...