CodeGraph: Open-Taxonomy Knowledge Graph for Source Code with Wikidata Grounding

CodeGraph. Open-Taxonomy Knowledge Graph for Source Code with Wikidata Grounding. By Federico Pennino and four others, from the University of Bologna, Télécom Paris, and one other institution. Published on the twenty-fourth of September. Its code repository is public and real: about thirty source files.

CodeGraph is a knowledge graph about what source code does, rather than how it is written. It records which algorithms a file implements, which paradigms it follows, which design patterns it instantiates, and which application domains it serves. The authors describe it as the first known large-scale open-taxonomy knowledge graph for source code. Its concepts are extracted from the code itself, and then tied to Wikidata.

The graph has approximately one hundred and fifty-eight million nodes. Around one hundred and forty-five million of those are files. About sixty-three thousand are extracted concept entities, things like algorithms, paradigms, design patterns, and application domains. Roughly nineteen thousand eight hundred are grounded Wikidata entities.

Connecting them are approximately one billion typed edges. They link files to their concepts, concepts to their grounded Wikidata identifiers, and those to their parent categories. The files cover fourteen programming languages.

It was built from the one hundred and sixty-seven million files of the Stack-Edu corpus. A code-specialised large language model produces the semantic annotations. Those extracted entities are then grounded in Wikidata through three stages. A deterministic SPARQL stage handles the unambiguous entities. A Deep Research Agent resolves the long tail that remains. A hierarchy-rollup stage then imports, for each resolved Wikidata identifier, the full closure of its parents.

To say how accurate the annotations are, the authors add a calibrated quality-assurance protocol. It combines a small human gold set with an LLM-as-a-judge filter to quantify annotation precision. The division of labour mirrors the linking stages: a small amount of careful human judgment, and an automated judge to carry that judgment across a corpus far too large for people to check by hand.

The motivation is the scale of public software repositories such as GitHub and the Software Heritage Archive, which store billions of files. The engineering knowledge in those files is implicit, and current tools are limited to syntactic and token-level analysis. Those tools can tell you what a file says, token by token, but not which algorithm it implements or which domain it serves. CodeGraph is for anyone who wants to ask questions at the level of concepts instead, and to connect the answers to an identifier that means the same thing outside this one graph.

What separates this from a pile of labels is the grounding. Each extracted concept is tied to a Wikidata identifier and to the parent categories above it, so the graph links a single file, through its concepts, to a shared public vocabulary.