<div align="center">

RepoHyper: Search-Expand-Refine on Semantic Graphs for Repository-Level Code Completion

</div>

Introduction

We introduce RepoHyper, an novel framework transforming code completion into a seamless end-to-end process for use case on real world repositories. Traditional approaches depend on integrating contexts into Code Language Models (CodeLLMs), often presuming these contexts to be inherently accurate. However, we've identified a gap: the standard benchmarks don't always present relevant contexts.

To address this, RepoHyper proposes in three novel steps:

• Construction of a Code Property Graph, establishing a rich source of context.
• A novel Search Algorithm for pinpointing the exact context needed.
• The Expand Algorithm, designed to uncover implicit connections between code elements (akin to the Link Prediction problem on social network mining).

Our comprehensive evaluations reveal that RepoHyper sets a new standard, outperforming other strong baseline on the RepoBench benchmark.

Paper: https://arxiv.org/abs/2403.06095

Installation

pip install -r requirements.txt

Architecture

<img src="arch.png" width="750" height="350">

RepoHyper is a two-stage model. The first stage is a search-then-expand algorithm on Repo-level Semantic Graph (RSG) then use GNN link predictor that reranks the retrieved results from KNN search and graph expansion. The second stage is any code LLM model that takes the retrieved context and predicts the next line of code.

Checkpoints

We provide the checkpoints for the GNN model here. The GNN model is trained on the RepoBench-R dataset with gold labels. We also provide RepoBench-R RGSs to reproduce the results.

Usage

Data preparation

We need to clone Repobench dataset into data/repobench folder. Then download all the unique repositories used in this dataset

python3 -m scripts.data.download_repos --dataset data/repobench --output data/repobench/repos --num-processes 8

The next step is to generate call graph using PyCG. We use the following command to generate call graph for each repository. 60 processes are used to speed up the process (maximum RAM usage is around 350GB).

python3 -m scripts.data.generate_call_graph --repos data/repobench/repos --output data/repobench/repos_call_graphs --num-processes 60

Now we need to generate embeddings for each node for node embedding as well as create adjacency matrix by aligning Tree-sitter functions, classes, methods with call graph nodes.

python3 -m scripts.data.repo_to_embeddings --repos data/repobench/repos --call-graphs data/repobench/repos_call_graphs --output data/repobench/repos_graphs --num-processes 60

Final step is labeling which node is the most optimal for predicting next line using gold snippet from repobench dataset. In this step, we also generate the training data for GNN training by extracting the subgraph using KNN search and RSG expansion.

python3 -m scripts.data.matching_repobench_graphs -search_policy "knn-pattern" --rsg_path "YOUR RSG PATH" --output data/repobench/repos_graphs_labeled 

Training

We can train GNN linker seperately using following script

CUDA_VISIBLE_DEVICES=0 deepspeed train_gnn.py --deepspeed --deepspeed_config ds_config.json --arch GraphSage --layers 1 --data-path data/repobench/repos_graphs_labeled_cosine_radius_unix --output data/repobench/gnn_model --num-epochs 10 --batch-size 16

Evaluation for RepoBench-P

We can evaluate the model using the following script

python3 scripts/evaluate_llm.py --data data/repobench/repos_graphs_matched_retrieved --model "gpt3.5" --num-workers 8

Citing RepoHyper

More details can be found in our paper.

If you're using RepoHyper in your research or applications, please cite using this BibTeX:

@article{phan2024repohyper,
  title={Repohyper: Better context retrieval is all you need for repository-level code completion},
  author={Phan, Huy N and Phan, Hoang N and Nguyen, Tien N and Bui, Nghi DQ},
  journal={arXiv preprint arXiv:2403.06095},
  year={2024}
}

Contact us

If you have any questions, comments or suggestions, please do not hesitate to contact us.

• Website: fpt-aicenter
• Email: bdqnghi@gmail.com

License

MIT License