DomAgent: Leveraging Knowledge Graphs and Case-Based Reasoning for Domain-Specific Code Generation

DomAgent is an autonomous coding agent designed to enhance the domain adaptability of large language models (LLMs). While general-purpose LLMs excel at generic code generation, they often struggle in real-world software development that requires domain-specific knowledge. To bridge this gap, DomAgent enables LLMs to generate domain-adapted code through structured reasoning and targeted knowledge retrieval. Its core component, DomRetriever, combines knowledge-graph reasoning (top-down understanding) with case-based reasoning (bottom-up examples) to dynamically retrieve and synthesize domain knowledge and representative cases. We train the large language models Qwen2.5-7B and LLaMA-3.1-8B using a Reinforcement Learning (RL) framework GRPO.

DomAgent can work as a standalone system or integrate with any LLM, helping small open-source models approach the performance of large proprietary ones in complex, domain-specific coding tasks.

Here, we only include the open-source benchmark dataset DS-1000. Due to the double-blind review policy, the truck CAN signal dataset used in our experiments will be released after the paper is accepted.

Package Directory Structure

.
├── OpenRLHF-RAG
├── README.md
├── requirements.txt
├── data
│   └── ds1000.jsonl
├── case-base-tool
│   ├── plain_text.json
│   └── semantic_search.py
├── kg-tool
│   ├── DS-KG.ttl
│   ├── kg_search_function.py
│   ├── prompt_list.py
│   └── utils.py
├── evaluation
│   ├── edit_ds1000.py
│   ├── execution.py
│   ├── extract_entity_from_query.py
│   ├── infer_domagent.py
│   ├── run_dist_inference.py
│   ├── run_inference.py
│   ├── run_openai.py
│   └── test_ds1000.py
├── reward-remote
│   └── reward_server.py
└── scripts
    ├── ray_start.sh
    └── reinforce_train.sh

---

Project Structure

• OpenRLHF-RAG/
  Contains tools for RLHF (Reinforcement Learning with Human Feedback) model training.
  This folder is adapted from OpenRLHF/OpenRLHF.

• data/
  Stores the training and testing datasets.

• evaluation/
  Used to load trained models and perform evaluation.

• kg-tool/
  Provides utilities for retrieving information from a knowledge graph.

• case-base-tool/
  Provides utilities for retrieving information from a case base.

• reward-remote/
  Implements the remote reward function used during RL training.

• scripts/
  Contains training scripts used in the RL training pipeline.

---

Requirements

• Python 3.x
• Install the required libraries:

  pip install -r requirements.txt

Usage

1. Enter the DomAgent folder:

   cd DomAgent

2. Training：

    ## Ray start
    bash scripts/ray_start.sh
   
    ## Start Reward Server
    python reward-remote/reward_server.py --port 1278
   
    ## Training
    bash scripts/reinforce_train.sh

3. Inference:

    python evaluation/infer_domagent.py --gpu_id 0 --temp 0.0 --port 5004 --prompt_type v0 --src_file data/eval_set/bamboogle_500.jsonl --model_path the_path_to_model

4. Execute the generated code to obtain the result：

    python evaluation/test_ds1000.py