Goodyear Engineer Builds TechGraphRAG to Fix AI's 80% Research Ceiling
A Goodyear engineer built a 13-step agentic RAG system over 2,100 technical papers, with self-correcting citation checks and a Neo4j knowledge graph — showing how domain experts can build research-grade AI tools for niche fields.
- What happened: A Goodyear engineer published TechGraphRAG, a 13-step agentic RAG system built over ~2,100 papers in intelligent tires and vehicle dynamics.
- Core innovation: A 100-point evidence sufficiency scoring framework that blocks generation when retrieved evidence is too weak, preventing hallucinated answers.
- Knowledge graph layer: A Neo4j graph built via LLM entity extraction and OpenAlex author validation enables relational, multi-hop reasoning across the corpus.
- Self-correcting loop: The system includes drift-guarded query reformulation, external academic database search (Crossref, OpenAlex, Semantic Scholar), and citation integrity verification before any output is produced.
- Limitation: No head-to-head benchmarks against other RAG systems; no public code release; fixed corpus with no described update mechanism.
- Best use case: A reference architecture for teams building domain-specific research assistants over large, curated technical corpora.
Most RAG tutorials still teach the 2023 playbook: embed a query, pull the top-k chunks, stuff them into a prompt, and generate. That pattern , Naive RAG , works fine for simple factual questions against a clean knowledge base, but plateaus at around 70–80% precision for anything more complex. For a domain expert trying to navigate thousands of highly technical papers, that ceiling is a wall.
That is the problem TechGraphRAG sets out to solve. The paper, authored by Kanwar Bharat Singh, a Principal Engineer at Goodyear, whose research spans intelligent tires, vehicle state estimation, and machine learning for automotive applications, describes a fully implemented agentic RAG framework built over a curated corpus of approximately 2,100 academic papers in intelligent tires, vehicle dynamics, and vehicle control. It is a rare case of a domain expert building a production-grade AI tool for their own field , and publishing the blueprint.
Why plain RAG breaks on technical literature
The core issue is that scientific knowledge is relational, not flat. Conventional RAG systems remain unsatisfactory when dealing with complex reasoning scenarios because flat retrieval directly returns fragmented chunks based on similarity matching, which limits their ability to model complex relationships between concepts and answer questions requiring multi-hop reasoning. Ask a system "which sensor fusion approaches have been validated on real tire data and cited by stability control papers?" and a vector search will return loosely related chunks , not a reasoned answer tracing the citation chain.
Standard RAG approaches treat the underlying corpus as a flat collection of documents, ignoring structured relationships such as scientific couplings or hierarchical parameter dependencies , a design choice that limits their effectiveness in technical domains where relational reasoning is critical. Add to this the problem of hallucinated citations , a model confidently citing a paper that does not exist , and you have a system that is actively dangerous for literature review.
A 13-step autonomous pipeline
TechGraphRAG's answer is an agentic pipeline that treats retrieval as a multi-stage reasoning process, not a single lookup. Rather than one pass from query to answer, the system runs through 13 autonomous steps before committing to a response. Here is how the key stages fit together:
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
