// part of the AXON ecosystem

GLYPH

Builds a knowledge graph from documents (LLM extraction) and code (tree-sitter, deterministic), then serves graph-aware context: anchor entities, expand the neighborhood by hops, hand the subgraph to the agent instead of the top-k most similar chunks. Every claim below is benchmarked against a fair vector baseline over the same corpus, not a strawman.

github.com/sammyjdev/glyph-kg· Apache-2.0 · the graph-retrieval layer AXON consumes to decidewhat context to bring.

// measured, not vibes

Two results, including the one that didn't go as planned.

Code domain: the graph lostcommit b7f9886 · 2026-06-11

0.839 → 0.995faithfulness, graph vs. the fair vector baseline on code; graph ranked last

measured, two judgesN = 14 cases

// the founding thesis was "graph wins on code." It didn't hold on the robust metric; published anyway, not hidden.

Sense-making: half the costMETRICS-code-global.md · 2026-07-01

5,337tokens for a community summary vs 10,148–10,355 for vector/graph, equal-or-better quality

measured, two judgesN = 8 cases5,337 vs 10,148 tokens

// "how is this organized?" answered at roughly half the tokens, with overlapping confidence intervals on quality, so no arm wins on accuracy alone.