Graph-Oriented Generation – Beating RAG for Codebases by 89%
Replaces vector RAG with dependency graphs; 89% fewer tokens but benchmark setup is contrived.
Harmonic coherence as a universal relationship detection operator
Ancient geometry meets Fourier analysis—neat math, but application to LLMs and databases unproven.
ML researchers, database theorists, LLM architecture explorers
Positional encoding schemes (RoPE, ALiBi) · Fourier feature networks · Vector similarity search (FAISS, Pinecone)
Replaces vector RAG with dependency graphs; 89% fewer tokens but benchmark setup is contrived.
Type a name and you can literally watch characters turn into IDs, 16‑dim embeddings get added with positional encodings, and causal attention matrices animate per head — all matched numerically to Karpathy's 244‑line microGPT. The implementation is pure TypeScript (no ML libs) and includes a helpful scrollable sidebar with the reference math, which makes this an excellent, low‑friction learning tool — more pedagogical deep dive than research innovation.
Swap software PRNG for hardware entropy in vLLM sampling, but niche use case with steep setup cost.
Semantic grep with word embeddings when traditional grep only does syntax.
Novel BPE variant using tf-idf scoring produces shorter encodings than classic.
Cuts agent token costs by 98% compared to grep without needing GPU inference.