Title: TechGraphRAG: A Graph-Enhanced Agentic RAG Architecture for Reasoning in Technical Literature

Abstract:

This study introduces an agentic retrieval-augmented generation (RAG) framework designed to facilitate domain-specific technical reasoning. The system is built upon a specialized collection of roughly 2,100 scholarly articles focusing on vehicle control, vehicle dynamics, and intelligent tires. Moving beyond traditional single-pass RAG models, the proposed architecture utilizes a comprehensive 13-step autonomous workflow. This pipeline begins by categorizing user queries based on their intent and evaluates the adequacy of evidence using a multi-dimensional assessment rubric.

The system employs an agentic retry mechanism that reformulates queries while guarding against drift. It also conducts iterative "optimize–search–vet" cycles to query external academic repositories, including Crossref, OpenAlex, and Semantic Scholar. Furthermore, the framework navigates a Neo4j knowledge graph to extract relational context, verifies the integrity of citations, and executes post-generation quality audits that trigger automatic regeneration if necessary.

Key innovations of this work include: * A 100-point scoring system for evidence sufficiency across five distinct dimensions, featuring relevance damping and a hybrid review process combining rules and LLMs. * An external search architecture tailored to specific routes, supported by iterative agentic loops. * A knowledge graph developed through LLM-driven entity extraction and OpenAlex-based author validation, alongside intra-corpus citation resolution. * A self-correcting generation cycle that integrates citation verification with quality assessment.

Ultimately, this paper presents a practical, implemented case study demonstrating how agentic, evidence-based RAG systems can effectively aid in navigating and reasoning through large, domain-specific technical corpora.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC