RugBrain Architecture 🟡 Beta

 Corpus (md/code) ──▶ Chunker ──▶ Dual embed ──▶ Qdrant collections
                                    │  dense (4B deep) + sparse (BM25)
                                    ▼
 Query ──▶ Embed ──▶ Hybrid search ──▶ RRF fusion ──▶ Reranker ──▶ Top-k + sources
                        (dense + named-vector)

Dual-tier embeddings

Two model sizes serve different jobs:

Tier	Model	Latency	Use
Deep	qwen3-embedding-4b	~120 ms/doc chunk	Ingestion, high-stakes queries
Fast	qwen3-embedding-0.6b	~15 ms	Interactive search, agent loops

Both run locally on CPU via llama.cpp — no embedding API costs, no data leaves the box.

Why hybrid?

Pure vector search misses exact identifiers (getReserves, slot0). BM25 sparse vectors restore lexical recall; RRF fusion keeps both recall curves. Measured +18% hit rate on symbol-name queries in our eval set.