CrossCoder
A research framework for repository-level code generation that retrieves context from external libraries as well as the local repository, using a knowledge graph over both.
CrossCoder targets a gap in repository-level code generation: existing retrieval methods mostly pull context from the local repository, leaving correct use of external library APIs dependent on what the model already learned, which fails for unfamiliar or version-specific APIs. CrossCoder builds a unified knowledge graph over both the repository and its library dependencies, identifies important nodes through planning and semantic retrieval, and expands to neighboring nodes to gather multi-hop context before generating code.
The project also introduces VersionExec, an execution-based benchmark derived from BigCodeBench that checks whether generated code stays correct across different dependency versions, and reports improvements in functional correctness and version robustness on RepoExec, DevEval and VersionExec.
It is written in Python and accompanies an EMNLP 2026 paper.