Blog post · November 19, 2026 · aviznetworks.com
Aviz Networks details a SONiC Network Copilot using CPT, SFT and RAG
- About Aviz Networks
- Also names Argilla
- Docker
- It
- MinIO
- Network
- ONE
- Sonic
- Read the original
Summary
Aviz Networks outlines a SONiC-focused Network Copilot built with continued pre-training, supervised fine-tuning, and retrieval-augmented generation. The post argues general-purpose language models lack the known issues, workarounds, and support procedures network teams rely on in production.
Training Architecture
The Network Copilot training architecture includes a data collection pipeline, model adaptation, supervised dataset generation and review, model serving, and a retrieval knowledge base. Source material includes internal technical documentation, configuration and operational guides, troubleshooting runbooks, product documentation, and Zendesk support data. The collected information is stored in MinIO before processing.
The article notes that these sources often contain information absent from public training datasets. One example is a support case linking a switch platform, firmware release, ASIC behavior, and workaround, since troubleshooting depends on exact hardware, software version, and operating environment.
Three Complementary Techniques
Continued Pre-Training
Continued Pre-Training extends a foundation model's training using a domain-specific corpus. The model is trained on a GPU cluster using a CPT pipeline with a LoRA adapter, which adapts the model without full fine-tuning of all parameters. The stated aim is familiarity with networking terminology, network operating system behavior, vendor implementations, and protocols.
Supervised Fine-Tuning
Supervised Fine-Tuning teaches the model to respond to engineering and support questions. The SFT dataset is generated using foundation models and curated source material, producing question-and-answer pairs, troubleshooting scenarios, hardware-specific workarounds, and root-cause explanations. A BGP instability case, for instance, can be converted into a training example containing symptoms, environment details, investigation steps, and resolution.
Generated examples are reviewed by subject matter experts in Argilla before acceptance. Reviewers assess technical correctness, completeness of troubleshooting steps, vendor-specific accuracy, and consistency with documented support procedures. The reviewed dataset then serves as the input for the SFT stage.
Retrieval-Augmented Generation
The RAG layer uses a FAISS-based knowledge base to retrieve relevant documentation and support knowledge when an engineer asks a question. The retrieved information is provided to the model as context. The article states this approach keeps frequently changing information accessible without another training cycle.
Intended Use Cases
The article lists known issue identification, troubleshooting and root-cause analysis, production upgrade guidance, vendor-specific operational guidance, and reuse of support knowledge through natural-language interaction. The model is served through a vLLM-based Docker deployment with an OpenAI-compatible API on port 8000, using a LoRA adapter.
Complementary Roles
The article describes CPT as providing domain familiarity, SFT as providing support-oriented behavior, and RAG as providing access to evolving knowledge. It states that none of the three replaces the others, and that together they form the basis of the assistant architecture.
For enterprise network and support leaders, the post frames domain-specific knowledge and expert review as the factors that determine the reliability of a network AI assistant. This Blog Summary is a fact-based summary of the vendor blog.
Our summary. The original keeps its own byline.