RugBrain — Security Knowledge RAG 🟡 Beta
RugBrain is the retrieval-augmented knowledge system behind RMI's research features: a vector knowledge base over security research, exploit write-ups, protocol code, and curated vulnerability patterns — queried with hybrid dense + sparse retrieval and a reranker.
What it powers
- Semantic pattern lookup — "show me flash-loan oracle manipulation defenses" returns curated code snippets with relevance scores.
- Exploit corpus search — natural-language Q&A over historical exploit post-mortems.
- Protocol design reference — "how does the health factor gate liquidations" with real implementations as evidence.
- Agent grounding — RMI's AI agent cites RugBrain hits instead of hallucinating.
Stack
| Layer | Tech |
|---|---|
| Vector store | Qdrant |
| Embeddings | qwen3-embedding (local llama.cpp, dual-tier: 4B deep / 0.6B fast) |
| Retrieval | Hybrid: dense cosine + BM25 sparse, fused with RRF |
| Rerank | qwen3-reranker (local) |
Methodology boundary
RugBrain indexes public research and open-source code. The curated detection KB used in production scoring is internal and not part of this open repo.