We develop an agentic system that leverages LLMs to automatically configure network switch devices based on natural language requests.
We use LLMs to translate natural language requests into configurations in SONiC. To enforce a structure in the LLM outputs, first we convert the natural language query into a YANG data file. This YANG output follows an explicit grammar that is defined by a set of provided YANG modules found in yang_modules/. We adopt these YANG modules from here. Then we translate the YANG configuration into SONiC via a deterministic translation of the configuration commands.
| Natural Language Request | YANG | SONiC ConfigDB |
|---|---|---|
| Configure connected interfaces between leaf and spine devices in separate class A IP network segments. Configure each leaf connected to the end-hosts appropriately so there is no conflict in IP addressing done for interfaces between leaf and spine devices. |  |
 |
| Please set up a policy on the Ethernet0 interface that drops all incoming traffic destined for the IP address 10.0.0.1. This rule should have the highest priority. Additionally, drop any traffic originating from the IP address 10.1.0.1 on source port 53. This second rule should have a lower priority than the first. |  |
 |
The user begins by describing a configuration objective in natural language (NL). There are three key stages:
Retrieval Stage: LLM Agent 1 performs dense retrieval on a vector database of the configuration files and identifies the parts of the network state that are relevant to this query. Asynchronously, LLM Agent 2 identifies the YANG modules that are relevant to this specific query.
Configuration Stage: Given the outputs from LLM Agents 1 and 2, LLM Agent 3 performs the configuration. It outputs a YANG configuration that satisfies the user query.
Verification Stage: We run the verifier pyang to check that the LLM output satisfies the syntax and constraints described in the YANG grammar. If there is an error, LLM Agent 4 attempts to correct the configuration based on the error log. The verification stage is repeated until the configuration passes the pyang tests, or until a specified number of iterations have failed.
After the YANG configuration has been verified, a compiler translates it into SONiC code via a deterministic mapping.
We use GPT-4 for all LLM agents. Instructions are provided to the agents via prompt templates found in prompt_templates/.
First, clone the repository and install the required dependencies:
git clone git@github.com:jzhou316/LLM-networking-control.git
pip install -r requirements.txt
Next, you will need to supply API keys for OpenAI and LangChain. Replace the empty values with your API keys in the fields OPENAI_API_KEY and LANGCHAIN_API_KEY found in handlers/.env.
Static SONiC configuration files have been provided in configs/sonic_configs. We also support real-time configurations by providing a script in handlers/network_handler.py to interact with the Cisco 8000 Emulator Sandbox for SONiC. You will need to set up a Cisco 8000 SONiC Notebook sandbox environment. Once you have set up your SONiC environment, upload the script server/server.py to the environment and run the file: python3 server.py. This will set up a RESTAPI to communicate between the sandbox environment and the local repository.
Run the following to open the Streamlit web interface. Natural language queries are provided to the LLM system via the interface. Moreover, the user can check the intermediate outputs from the LLM, as well as the final YANG and SONiC configurations.
streamlit run main.py
Here is an illustration of a user performing a VLAN configuration using our Streamlit interface.
@article{lin2025llm,
title={An LLM-based Agentic Framework for Accessible NetworkControl},
author={Lin, Samuel and Zhou, Jiawei and Yu, Minlan},
journal={ACM SIGMETRICS Performance Evaluation Review},
volume={53},
number={2},
pages={15--20},
year={2025},
publisher={ACM New York, NY, USA}
}
Work done by Samuel Lin, Jiawei Zhou, and Minlan Yu.
See the LICENSE.md file for details