Skip to content

System Structure

RPLH Systems

We hope to setup an systematic and modularzied way of doing multi-agent communication for less parametrized language models. We have implemented multiple instances of the RPLH system dpending on vanilla or efficient version. Here is the main breakdown of the code:

rplh/
├── env/
│   └── env.py
├── systems/
│   ├── d_efficient/
│   │   ├── memory.py
│   │   └── rplh_inference.py
│   ├── h_efficient/
│   │   ├── execution_checker.py
│   │   ├── memory/
│   │   │   ├── memory_standard.py
│   │   │   └── memory_agent.py
│   │   ├── rplh_inference.py
│   │   └── rplh_agent_inference.py
│   ├── h_vanilla/
│   │   ├── execution_checker.py
│   │   ├── memory.py
│   │   └── rplh_inference.py
├── llm/
│   ├── language_model.py
│   └── response_model.py
├── rendering/
│   ├── render_conversation.py
│   ├── render_states.py
│   └── animations.py
├── evaluation/
│   ├── evals.py
│   ├── energy.py
│   ├── embed.py
│   └── get_data.py
├── test.py
└── inference.py
Note the following:

  1. There are many shared features across all different type of communication system, which we ahve modularized into the llm folder, env folder, or the rendeiring folder for both using it in the vanilla and efficient models. In addition, because of implementation differences, files such as execution_checker.py, memory.py, and the main inference loop rplh_inference.py may be different, which is why each system has its own unique implementation.
  2. The d_efficient (efficient implementation of decentralized framework) is a particular instance of the h_efficient system (our hallucination system) where the inference loops and memory module is adjusted.
  3. Both h_efficient and h_vanilla are our hallucination system with slight differences in implementation, refer to these links for more details:
  4. We have created inference functions in each of the system to be called seperately, or an direct interface can be called by specifying args on the root folder level of rplh (inference.py).