Engineering Capabilities for Enterprise Large Model Applications: Prompt Engineering, RAG, and Fine‑Tuning

In a previous article the author examined six basic factors for private model training using GitHub Copilot as a case study; this follow‑up continues the discussion by focusing on three key technologies—prompt engineering, Retrieval‑Augmented Generation (RAG), and model fine‑tuning—and how they relate to each other.

SE4AI Capability Building : Even if an enterprise does not meet all six private‑training requirements, it still needs a customized large‑model solution. The article presents a visual roadmap for engineering capabilities required to deliver such solutions.

Two Optimization Dimensions : The author identifies (1) behavior optimization, which stabilizes output format, tone, and preferences, and (2) context optimization, which enriches the model with private data such as internal code, documentation, and policies.

Three Optimization Methods :

Prompt Engineering – the most economical and quickest way to improve both behavior and context; it includes zero‑shot/few‑shot and in‑context learning techniques.

Retrieval‑Augmented Generation (RAG) – supplements the model with external knowledge at inference time, addressing the model’s static nature and lack of internal corporate knowledge.

Fine‑Tuning – adjusts model behavior using a small, task‑specific dataset, improving stability and output quality for high‑frequency, fixed tasks.

The article advises starting with prompt engineering, then introducing RAG when knowledge coverage becomes too large, and finally applying fine‑tuning when prompts become overly complex yet still unstable.

Model Application Testing : Traditional deterministic testing does not apply to generative AI. The author outlines the need to redesign test definitions, evaluation methods, and metrics (e.g., relevance, summarization, bias, toxicity, friendliness, safety, compliance, hallucination) for large‑model applications.

Conclusion : Prompt engineering, RAG, and fine‑tuning are complementary; successful enterprise AI solutions typically combine all three.

Event Invitation : The author invites readers to attend the “R&D Efficiency & Innovation Conference – IDCF 5‑Year Anniversary” on May 25 in Beijing, where he will present “AI‑Driven Software Engineering Practices to Improve Personalized Code Generation Accuracy”. The notice includes speaker line‑up, agenda, ticket information, and promotional details.