Koragraph vs CodeGraph, GitNexus, Graphify

The question is whether a graph holds the right declarations and the right edges. Every number is produced by an independent oracle, a compiler front end rather than the tool under test, scored by one shared scorer at one shared library scope, over 12 languages and 36 pinned repositories, 3 per language. Koragraph wins 34 of 39 language and plane cells on recall and leads precision on all 39.

The systems are pinned at current builds: Koragraph shipping build, CodeGraph 1.6.0, GitNexus 1.6.10, Graphify 0.9.51. Each competitor’s index is written to a throwaway location and deleted before the oracle runs, so no system’s own cache is ever scored as source, and no path spends a paid model call.

One oracle per language, never a tree-sitter grammar where a compiler front end exists. C, C++ and Swift are scored on declarations only, since there is no build-free edge oracle for them that a fair referee could run.

Recall, koragraph against the best competitor in each cell. The handful of trails are precision and recall trade-offs or near-ties rather than planes another tool wins decisively.

Declarations are types, methods, fields and constants, pooled and matched at the name level within the oracle’s own file set. Koragraph leads recall on every language but two, and the two exceptions are a precision trade rather than a coverage gap.

Pooled declarations, recall / precision. The highlighted cell leads recall in its row, whichever system holds it. C# and JavaScript are the two rows where a competitor edges Koragraph, both by a fraction.

The intra-repo calls plane is the one that decides whether a graph can trace a call back to its callers. The callee must be a declared in-repo symbol, so no system is penalised for not modelling standard-library targets. Koragraph leads recall on every language with an independent edge oracle except Rust, and does it at 91 to 100% precision, while the tools that report higher raw recall elsewhere do so at far lower precision.

Intra-repo calls, recall / precision. The highlighted cell leads recall in its row. On JavaScript the gap is 87.6 against a best rival of 45.1. Rust is the one row where a competitor leads recall, highlighted for CodeGraph, and Koragraph keeps the higher precision there.