LLM-Based AI Agent for Automated VNF/CNF Configuration

This repository contains an LLM-based AI agent for automating VNF/CNF deployment in NFV environments.
The system takes user-provided MOPs (Method of Procedure) as input and generates automated configuration workflows (Python / Ansible) using an LLM.
Before applying any generated configuration to the real network, the code is validated inside an NDT (Network Digital Twin) environment.
Only configurations that successfully pass NDT verification are returned to the user.

If the generated configuration fails during execution, the system monitors errors and allows the LLM to iteratively fix them using experimental techniques such as RAG, MAD, and RL.

Below is an example result showing improved configuration success rates across several LLMs relative to module execution time:

This work is currently being written as a research paper.
A preliminary version covering OpenStack-based experiments has been published.

---

📌 System Architecture

Requriements:

We use Python=3.11 TRL and Unsloth are updated weekly, so I have included the specific versions I use in the requirements.txt file. However, they may still fail to install in your environment.

pip install paramiko python-openstackclient python-docx ansible_runner astor pytz tqdm kubernetes trl datasets peft bitsandbytes nltk matplotlib chromadb transformers huggingface_hub sentence-transformers llama-cpp-python llm-blender mergekit tensorboard weave langchain-community
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"

The current prototype consists of the following major modules:

• LLM Module – Generates workflow code based on user-provided MOPs
• Test Module – Executes and validates generated workflows
• NDT (Network Digital Twin) – Supports Kubernetes and OpenStack environments
• RAG Module – Knowledge-enhanced LLM support (KB crawling, embedding fine-tuning, I/O filtering)
• MAD Module – Self-healing improvements (experimental)
• RL Module – Reinforcement learning integration (in development)

---

📁 Directory Structure and Key Files

The core AI Agent code is located in llm_based_config/.
Each component is described below:

Core Agent

• main.py
  Main AI agent logic and workflow orchestrator.

OpenStack Interaction

• make_new_floating_ip.py
• openstack_config.py
  Libraries for communicating with and manipulating OpenStack resources.

MAD Module

• mad.py
  Implements MAD-based self-correction logic (currently unused).

RAG Module

Directory: RAG/
Includes:

• Knowledge base crawling scripts
• Word embedding fine-tuning code
• Input/output filtering logic
  This is one of the core components under active research and paper preparation.

Prompting

• prompt.py
  Prompt templates used for requesting configuration generation from LLMs.

Workflow Preparation

• python_code_modify.py
  Pre-test processing for Python-generated workflows before applying them to the Test Module.

Fine-Tuning & RL (In Development)

• LLM_Finetuning.py
  LLM fine-tuning library.
• RL_LLM.py
  Reinforcement learning framework built on top of LLM_Finetuning.py.

---

📄 Additional Directories

• MOP/
  Contains the MOPs used in experiments.

• data_generating/
  Scripts to automatically generate MOPs using GPT-based generation pipelines.

---

📫 Contact

For inquiries related to this project, please contact:
📧 obiwan96@postech.ac.kr

---

📝 Citation

If you wish to cite this research, please reference the following paper (supported by Samsung Electronics Co., Ltd.):

S. Nam, N. Van Tu and J. W. -K. Hong,
"LLM-Based AI Agent for VNF Deployment in OpenStack Environment,"
NOMS 2025-2025 IEEE Network Operations and Management Symposium, Honolulu, HI, USA, 2025, pp. 1-7.
doi: 10.1109/NOMS57970.2025.11073607.

Paper Link: https://ieeexplore.ieee.org/abstract/document/11073607