Frontmatter the agent reads.

Nine fields. Each one earns its place by changing how the LLM picks, ranks, or refreshes the article.

The Sources section uses five wikilink prefixes so the relationship between article and source is explicit. cortex-lint flags untyped sources as warnings.

What a clean cortex article looks like end-to-end.

---
title: "Retrieval-augmented generation"
description: "Grounding LLM responses in your own documents. The standard architecture for memory-aware agents."
topic: "ai"
sources:
  - "40-raw/youtube/rag-explained.md"
  - "40-raw/papers/lewis-et-al-2020-rag.md"
created: 2026-04-12
last_compiled: 2026-05-08
verified_at: 2026-05-08
confidence: high
staleness_signal: "RAG architecture moves to graph-RAG by default, or vector DBs are replaced by long-context models"
---

# Retrieval-augmented generation

LLMs answer better when they retrieve relevant documents first.

## TL;DR

RAG pairs an LLM with a retriever that fetches relevant chunks from your own
corpus before generation. The model's response is grounded in real documents
instead of pure parametric memory, which reduces hallucination and lets you
update knowledge without retraining. Cost: retrieval quality is now your
bottleneck.

## Summary

[2-3 paragraphs.]

## Key Facts

- Lewis et al. (2020) introduced the term, pairing BART with a dense retriever.
- Modern stacks pick top-k chunks via embedding similarity, then concatenate.
- Retrieval quality dominates response quality once the LLM is good enough.

## Connections

- Related: [[10-cortex/ai/embeddings]], [[10-cortex/ai/long-context]]
- Used in: vault-side memory loop (cortex-compile + cortex-connect)
- Contrasts with: [[10-cortex/ai/finetuning]] -- finetuning bakes knowledge in;
  RAG keeps it swappable.

## Sources

- supports:: [[40-raw/papers/lewis-et-al-2020-rag.md]] -- original architecture description
- extends:: [[40-raw/youtube/rag-explained.md]] -- modern stack walkthrough (top-k, reranking, hybrid search)

Read-only diagnostic. Run /cortex-lint to surface decay and drift across 10-cortex/. Never mutates a file; the report is for you to read and act on.