Beyond Repository Boundaries: Cross-Repository Graph Retrieval for Code Generation

Abstract

Repository-level code generation requires generated code to be compatible not only with the target repository but also with its dependency environment. Existing retrieval-based methods mainly retrieve context from the local repository, leaving external API usage dependent on the model’s pretrained knowledge, which can be insufficient for unseen or version-specific APIs. Moreover, current retrieval strategies largely focus on one-hop evidence and overlook the structural relationships among code components. We propose CrossCoder, a cross-repository code generation framework that explicitly incorporates external libraries into the retrieval context through a unified knowledge graph over repository and library entities. CrossCoder identifies important nodes via planning and semantic retrieval, then selectively expands neighboring nodes to retrieve richer multi-hop contextual evidence for generation. To further evaluate dependency-version compatibility, we introduce VersionExec, an execution-based benchmark derived from BigCodeBench that evaluates generation under different dependency versions. Experimental results on RepoExec, DevEval, and VersionExec demonstrate that CrossCoder consistently improves both functional correctness (up to 6.3% on pass@1) and robustness to dependency-version changes.

Setup

Create and activate the environment:

conda create -n crosscoder python=3.13.7
conda activate crosscoder
pip install -r requirements.txt

Extract the shared data archive:

unzip data.zip -d .

Set Fireworks credentials:

export FIREWORK_API_KEY=<your_fireworks_key>
export FIREWORK_BASE_URL=https://api.fireworks.ai/inference/v1

Use the model name after the Fireworks prefix in commands. For example, gpt-oss-120b becomes accounts/fireworks/models/gpt-oss-120b inside the pipeline.

Process Benchmark Data

Prepare RepoExec:

cd benchmark/RepoExec
unzip test-apps.zip
cd ../..

python src/utils/process_lib.py --dataset RepoExec
python src/utils/download_lib.py --benchmark RepoExec --wrap --download

Prepare DevEval:

cd benchmark/DevEval
tar -xzf data.tar.gz
wget https://huggingface.co/datasets/LJ0815/DevEval/resolve/main/Source_Code.tar.gz
tar -xvzf Source_Code.tar.gz
cd ../..

python src/utils/process_lib.py --dataset DevEval
python src/utils/download_lib.py --benchmark DevEval --wrap --download

Prepare VersionExec:

python src/utils/download_lib.py --benchmark versionexec_old --download
python src/utils/download_lib.py --benchmark versionexec_new --download

Build Graph

bash src/graph_builder/parser/run.sh --dataset <RepoExec|DevEval|versionexec_old|versionexec_new>

Example:

bash src/graph_builder/parser/run.sh --dataset versionexec_new

Load Benchmark Samples

python src/load_benchmark.py \
  --benchmark <RepoExec|DevEval|versionexec_old|versionexec_new> \
  --extend True \
  --model <model>

Example:

python src/load_benchmark.py \
  --benchmark versionexec_new \
  --extend True \
  --model gpt-oss-20b

Generate Plan and Graph Expansion

python src/retriever/gen_plan.py \
  --benchmark <RepoExec|DevEval|versionexec_old|versionexec_new> \
  --model <model> \
  --retriever unixcoder

Then, run the following script

python src/retriever/rebuild_predicted_components.py \
  --benchmark <RepoExec|DevEval|versionexec_old|versionexec_new> \
  --model <model> \
  --top_k 5

Build Prompts and generate code

python src/retriever/create_prompt_rerank.py \
  --benchmark <RepoExec|DevEval|versionexec_old|versionexec_new> \
  --model <model> \
  --threshold 0.25 \
  --include_import true

This writes prompts to:

data/prompt/

Then, run the following command:

python src/generator/generate.py \
  --benchmark <RepoExec|DevEval|versionexec_old|versionexec_new> \
  --model <model> \
  --num_sample 3

Evaluation

Evaluate RepoExec:

Read benchmark/RepoExec/README.md.

Evaluate DevEval:

Read benchmark/DevEval/README.md.

Evaluate VersionExec:

Start Docker Desktop before running the evaluation commands.

Setup VersionExec:

conda create -n versionexec python=3.10.0
conda activate versionexec

pip install -r benchmark/versionexec_old/Requirements/requirements.txt

Evaluate VersionExec Old:

conda activate versionexec
cd benchmark/versionexec_old

bash run_docker_full.sh \
  --clean-docker \
  --folder <output_folder_contains_jsonl_result_file>

Evaluate VersionExec New:

conda activate versionexec
cd benchmark/versionexec_new

bash run_docker_full.sh \
  --clean-docker \
  --folder <output_folder_contains_jsonl_result_file>