Usually, problems in the AI project arise not from the quality of the model but from the poor design of the system surrounding it. The model is the engine; design patterns are the chassis, the steering, and the brakes. Without them, you end up with a good engine going nowhere. The development of large language models has advanced to such an extent that by 2025, we now have repeatable and battle-tested design patterns, just like we have in software engineering. This article provides an overview of those patterns, their purpose, the scenarios in which they are applicable, and how they can be incorporated into production systems.
1. The Prompt Chain Pattern
The simplest pattern. You break a complex task into a sequence of smaller prompts where the output of one becomes the input of the next. Instead of asking one giant prompt to do everything, you create a pipeline of focused steps.
Where an answer to a single prompt is inconsistent due to its complexity, writing a complete blog post using a single prompt is not possible. It needs to be divided into outline, draft, editing, and SEO.
Reference: Prompt chaining is now considered a foundational technique in production AI systems, sitting at the core of how developers build reliable multi-step reasoning pipelines.
2. RAG (Retrieval Augmented Generation)
RAG is the most prevalent AI architecture design currently being used in practice. Rather than depending on the knowledge that the model has acquired through its training process, the relevant documents are retrieved during the query and added to the prompt. The model makes reasoning based on factual, current, and concrete knowledge.
The core components
The RAG model has two main elements: a Retriever, which searches for documents related to the query from the knowledge base, like vector databases or search engines; and a Generator, which integrates the retrieved documents with the input query to produce responses.
When to use it
Any time your AI needs to answer questions about documents, internal knowledge bases, or information newer than the model's training cutoff. Customer support bots, document QA, internal search all RAG.
The basic RAG approach involves vectorizing the input question posed by the user, searching for similar documents in the vector database, and passing those documents along with the question itself to the language model. More complex approaches include self-RAG, where the model determines if retrieval should be done, and corrective RAG, where the model assesses the retrieved documents.
