Multimodal AI Assistant Boosts Network Config: 96.6% Accuracy, 26× Labor Cut

Background and Challenges

As enterprise networks grow in scale and complexity, traditional manual configuration becomes inefficient, requires deep protocol expertise, and often neglects performance guarantees, making it unsuitable for fast‑changing service‑quality demands.

Key Innovations of NLI2Conf

Multimodal network interface : NLI2Conf introduces a Text‑Attribute‑Graph (TAG) that directly integrates raw device configuration files, network topology, and performance metrics, eliminating the need for custom domain‑specific languages and using self‑attention to obtain semantic representations.

Intent‑configuration alignment : The model incorporates performance constraints into the update process, interpreting natural‑language intents while simultaneously satisfying forwarding policies and performance guarantees, thereby meeting strict service‑quality requirements.

Data‑driven training framework : A two‑stage training pipeline first performs masked pre‑training to enhance semantic understanding of configurations and structures, then applies LoRA fine‑tuning for efficient configuration‑update inference, reducing training cost while boosting performance.

Technical Architecture

NLI2Conf consists of three core modules:

CIEncoder : A multi‑layer self‑attention encoder that captures relationships between configuration files and natural‑language intents, producing contextual vectors for downstream inference.

NetEncoder : Built on Graph Neural Networks, it encodes topology and link‑state information, aggregating node features through message passing to achieve a global network view.

LoRA fine‑tuning : Low‑rank adaptation integrates multimodal network representations with a large language model without altering its core parameters, preserving language capabilities while lowering training and deployment costs.

Experimental Validation

Experiments across various network sizes demonstrate substantial advantages:

Text Accuracy (TA): 96.6% on a 100‑node network, far surpassing GraphLLM (31.2%) and GraphAdapter (42.4%).

Strategy Consistency (SC): 96.6% consistency, whereas competing models stay below 44.2%.

Full‑Satisfaction Rate (FSR): 96.6% on large‑scale networks, reducing manual review to 3.8% and improving operational efficiency by 26×.

The model also shows robust adaptability to different intent types (forwarding policies vs. performance constraints) and protocols (OSPF, BGP).

Future Outlook

NLI2Conf offers a promising solution for intent‑driven network automation, allowing operators to describe requirements in natural language instead of memorizing complex commands. As large language models and graph neural networks continue to evolve, such intelligent assistants are expected to play an increasingly important role in complex network scenarios, driving network management toward greater intelligence and efficiency.