Graph-Oriented Generation – Beating RAG for Codebases by 89%

Name: Graph-Oriented Generation – Beating RAG for Codebases by 89%
Availability: InStock
Author: dchisholm125

by dchisholm125·Mar 6, 2026·12 points·2 comments

Visit Project View on HN

AI Analysis

●●SolidBig BrainWizardry

Replaces vector RAG with dependency graphs; 89% fewer tokens but benchmark setup is contrived.

Strengths

•Deterministic graph traversal O(1) instead of vector similarity O(n) reduces token bloat and latency for real codebase navigation.
•Reproducible benchmark with local Ollama test harness proves tiny 0.8B models work when context is surgical, not bloated.
•Concrete math behind the claim (SRM vs RAG comparison) with runnable code—not hand-waved.

Weaknesses

•Benchmark setup (inflated dummy repo with hidden 'needle') doesn't prove real-world performance on organic codebases with messy interdependencies.
•Still seeking arXiv endorsement suggests research-stage rigor; lacks production deployment or adoption proof to validate the approach scales.

Post Description

LLMs are better at being the "mouth" than the "brain" and I can prove it mathematically. I built a deterministic graph engine that offloads reasoning from the LLM. It reduces token usage by 89% and makes a tiny 0.8B model trace enterprise execution paths flawlessly. Here is the white paper and the reproducible benchmark.