Emerging Architectures for LLM Applications ↗

Editorial Notes

The most widely-referenced architectural map of the LLM application stack, based on a16z’s analysis of hundreds of AI startups and enterprise deployments. This article frames AI as fundamentally a data engineering problem, with in-context learning (RAG) as the dominant pattern for connecting models to proprietary data. For technical leaders evaluating architecture decisions, the key insight is the stack decomposition: data pipelines, embedding models, vector databases, orchestration, and operational tooling are distinct components with different vendor landscapes and build-vs-buy tradeoffs. The article has been updated multiple times since its 2023 publication, reflecting how rapidly the reference architecture evolves — pay attention to what has changed as much as what remains stable. Use this as a framework for evaluating your own stack against the emerging consensus.