RAG Pattern Picker: match your content feature to the right retriever-augments-generator architecture

Concept

A structured, one-day architecture diagnostic for teams about to build a retrieval-augmented content feature. The intake captures the output modality (text, image, audio, code, video), the task, the data store's size and refresh rate, the latency budget, and which pain dominates — stale knowledge, long-tail gaps, data-leakage risk, or training/inference cost. The tool maps those answers onto the survey's taxonomy of how a retriever can augment a generator (e.g., retrieval injected into the prompt at inference time, retrieval used to train or adapt the retriever or generator, retrieval consulted during decoding, or post-generation correction) and returns a ranked shortlist of augmentation patterns with trade-offs for each, plus an engineering checklist covering index construction, chunking, refresh cadence, and retrieval evaluation. The deliverable is a specific build recommendation — not 'use RAG' but 'use this augmentation pattern for this feature, here is the implementation path, and here are its known limitations.' Sold per-assessment to product teams, or as an internal standard for platform companies that ship many AI content features per year.

Why now

The new survey [0] is the enabling evidence, and this knowledge did not exist in consolidated form before: it "classif[ies] RAG foundations according to how the retriever augments the generator, distilling the fundamental abstractions of the augmentation methodologies for various retrievers and generators," and separately summarizes "additional enhancement methods for RAG, facilitating effective engineering and implementation of RAG systems," along with benchmarks and stated limitations of current systems. That taxonomy-plus-checklist is exactly the knowledge base a selection tool needs. The survey also names the four hurdles — updating knowledge, long-tail data, data leakage, and high training/inference cost — that any pattern choice must trade off, giving the diagnostic its scoring dimensions. The older GraphRAG survey [1] adds graph retrieval as a candidate branch when the corpus is entity-relationship heavy. Because the survey attributes RAG's value to delivering "higher accuracy and better robustness" in AIGC, choosing the wrong augmentation pattern means building a system that still fails on the very hurdle that motivated it.