LLMSystemPatterns
title: Patterns for Building LLM-based Systems & Products radar_quadrant: Techniques radar_ring: Assess radar_position: inner created: 2026-05-26 last_updated: 2026-05-26 tags: [llm, rag, evaluation, guardrails, system-design] source_url: https://eugeneyan.com/writing/llm-patterns/
A reference taxonomy of recurring design patterns for production LLM applications, authored by Eugene Yan and published on eugeneyan.com in 2023. The article catalogues solutions that practitioners repeatedly rediscover when moving LLM projects from prototype to production.
Evals covers how to measure model output quality: unit tests for known inputs, human evaluation pipelines, model-based evaluation, and A/B testing in production.
RAG (Retrieval-Augmented Generation) covers document chunking strategies, embedding models, vector store selection, and retrieval quality improvements such as re-ranking and hybrid search.
Fine-tuning covers when to fine-tune versus prompt-engineer, data preparation, instruction tuning, and RLHF basics.
Caching covers semantic caching of LLM outputs to reduce latency and cost for repeated or similar queries.
Guardrails covers input/output validation, topic filters, PII detection, and fallback strategies for out-of-scope requests.
Defensive UX covers how to handle model uncertainty gracefully — disclaimers, confidence signals, graceful degradation.
Collect feedback covers implicit and explicit feedback loops to continuously improve deployed models.
The article functions as a practitioner's "Design Patterns" for LLM engineering. It names and frames problems that teams encounter at scale — sequential fetching bottlenecks, evaluation gaps, hallucination surface — and maps known solutions. Frequently cited in LLM engineering literature and team onboarding materials.
Placed in Techniques / Assess / inner. The patterns themselves are proven at the individual level (RAG, evals, guardrails are standard practice), but the taxonomy as a unified reference is worth tracking for team alignment and onboarding. Inner position reflects that several patterns (RAG chunking, semantic caching, model-based eval) are directly applicable to active projects without further tooling investment. Trial gate: apply at least three named patterns from the taxonomy in a single production LLM project with documented rationale tied to the pattern names.
