CodeGraph: Open-Taxonomy Knowledge Graph for Source Code with Wikidata Grounding
Summary
The paper introduces CodeGraph, a large-scale open-taxonomy knowledge graph for source code, built using a specialized LLM and grounded in Wikidata, covering millions of files and entities.
View Cached Full Text
Cached at: 09/28/26, 04:05 PM
Paper page - CodeGraph: Open-Taxonomy Knowledge Graph for Source Code with Wikidata Grounding
Source: https://huggingface.co/papers/2609.29474
Abstract
Publicsoftwarerepositories,likeGitHubandSoftwareHeritageArchive,storebillionsoffiles,yetextractingtheirimplicitengineeringknowledge---i.e.,thealgorithmstheyimplement,theparadigmstheyfollow,thepatternstheyinstantiate,andtheapplicationdomainstheyserve---remainschallenging,ascurrenttoolsareconstrainedtosyntacticandtoken-levelanalysis.Wepresentapipelineforbuildinganopen-taxonomysemanticannotationofsourcecodeusingacode-specialisedLargeLanguageModel.TheextractedentitiesaregroundedinWikidatathroughathree-stagelinkingprocedure:adeterministicSPARQLstagehandlesunambiguousentities,aDeepResearchAgentresolvestheresiduallongtail,andahierarchy-rollupstageimportstheparent-ofclosureofeachresolvedWikidataidentifier.Theresultingannotationsarematerialisedasasource-code-specificopen-taxonomyknowledgegraph.Wefurtherintroduceacalibratedquality-assuranceprotocolthatquantifiesannotationprecisionbycombiningasmallhumangoldsetwithanLLM-as-a-judgefilter.Weappliedourpipelinetothe167millionfilesoftheStack-Educorpus,creatingthefirstknownlarge-scaleopen-taxonomyknowledgegraphforsourcecode.Ourgraph,namedCodeGraph,containsapproximately158millionnodes,whichincludearound145millionfiles,about63,000extractedconceptentities(suchasalgorithms,paradigms,designpatterns,andapplicationdomains),androughly19,800groundedWikidataentities.Furthermore,CodeGraphfeaturesapproximately1billiontypededgesthatconnectfilestotheirrespectiveconcepts,linktheseconceptstotheirgroundedWikidataidentifiers,andrelatethemtotheirparentcategories,covering14programminglanguages.
View arXiv pageView PDFGitHub0Add to collection
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2609.29474 in a model README.md to link it from this page.
Datasets citing this paper1
#### halykoss/CodeGraph Viewer• Updated3 days ago • 1.18B • 4.88k
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2609.29474 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
colbymchenry/codegraph
CodeGraph is an open-source tool that creates a pre-indexed knowledge graph of a codebase, enabling Claude Code's exploration agents to query symbol relationships and call graphs instantly, reducing tool calls by up to 96% and exploration time by 77%.
@cathrynlavery: been starting to use codegraph in projects. It builds a local knowledge graph of every symbol, function, and connection…
Codegraph builds a local knowledge graph of every symbol, function, and connection in your code, allowing AI agents to look things up instantly instead of grep-searching, with reported cost and tool call reductions.
@BrooksWhaleX: Breaking: Someone just open sourced a knowledge graph engine for your codebase and it's terrifying how good it is. It's…
GitNexus is an open-source knowledge graph engine that indexes codebases, maps dependencies and execution flows, and integrates with AI coding assistants via MCP, providing full architectural clarity to prevent breaking changes.
vitali87/code-graph-rag
Code-Graph-RAG is an open-source tool that parses multi-language codebases with Tree-sitter, builds a knowledge graph in Memgraph, and lets you query, edit, and optimize code using plain English. It supports monorepos with mixed languages under a unified graph schema.
I built a code knowledge graph tool that's actually MIT licensed (fully local, no cloud)
An indie dev released repopedia, an MIT-licensed, fully local code knowledge-graph tool that builds a SQLite graph from Python/TypeScript repos using tree-sitter, plus an MCP server so coding agents like Claude Code can query call graphs and blast-radius impact directly instead of relying on grep.