LLM-powered Smart Contract Vulnerability Detection with RAG-based Knowledge Base and Corrective Quality Gating

Abstract

Once deployed, smart contract bytecode is final and cannot be patched. Any vulnerability that ships with the contract becomes a permanent attack surface. Existing tools have clear gaps. Static analyzers such as Slither flag aggressively and tend to produce many false positives, while large language models (LLMs) reason fluently but lack audit-specific knowledge. Retrieval-Augmented Generation (RAG) is meant to bridge the two. However, most current pipelines push every retrieved passage into the prompt without any quality check, and when retrieval misses, the injected noise actually makes the output worse instead of helping. We propose a detection pipeline with a retrieval quality gate. Abstract syntax tree (AST) parsing and Slither analysis run in parallel with a two-stage retriever over a 21,032-entry knowledge base built from 208 professional audit reports. A Corrective RAG (CRAG) gate inspects the reranker score and decides whether retrieved evidence is admitted into the prompt at all. On top of that, a Prompt Tiering layer adds audit-derived rules into the Gemini 2.5 Pro context only when CRAG says retrieval is trustworthy. We evaluate on SmartBugs-Curated paired with GPTScan Top200. The full pipeline reaches 80.3% F1 and brings the LLM-only baseline’s false positive rate (FPR) from 44.1% down to 21.2%, which is a 52% relative reduction, while keeping recall at 100%.