LLM-Based Systems
Understand key patterns of LLM based systems in enterprise AI, including classification, extraction, summarization, and question answering. Learn when to apply retrieval-augmented generation or fine-tuning to meet specific business needs. Discover best practices for evaluation, scoping, and handing off maintainable AI solutions.
We'll cover the following...
LLM-based systems form the backbone of enterprise AI deployments. Many of the highest-value workflows an FDE encounters are addressable without multi-step reasoning, tool orchestration, or autonomous action. A well-scoped LLM-based system delivers measurable value faster and with less operational overhead than a more complex alternative.
These systems follow a small set of recurring patterns. The FDE who recognizes those patterns during discovery can move from problem to architecture to scope document without rebuilding the approach from scratch on every project.
This lesson covers the core patterns, the two architectural extensions that most commonly appear alongside them, and the signals that tell the FDE this is the right system type for the problem.
Patterns of LLM-based systems
LLM-based systems fall into four patterns in enterprise deployments. Each has a distinct input type, a distinct output type, and a distinct evaluation approach the FDE defines before the build begins.
Classification: Takes an input and assigns it to a predefined category. A support ticket routing system that reads an incoming message and directs it to the relevant team follows this pattern. The FDE defines the category schema and labeled test set during scoping, which makes evaluation straightforward: precision and recall against the agreed test set.
Extraction: Pulls structured data from unstructured text. A contract analysis system that reads incoming agreements and returns named parties, effective dates, and key payment terms follows this pattern. Evaluation requires a ...