---
name: rag-retrieval-example
description: MDA source illustrating how typed footnote relationships encode retrieval graph edges for a RAG pipeline, enabling author-asserted document relationships to augment vector similarity.
title: RAG Retrieval Pipeline
doc-id: b3be8a43-b79a-4ea8-b169-bfca74fcccbc
author: Landon Kavanaugh
tags: [example, rag, retrieval, knowledge-graph, llm]
created-date: "2026-05-08T00:00:00Z"
metadata:
  mda:
    version: "1.0.0"
  ai-script:
    version: "1"
    instructions: |
      Use the relationship graph in this document to determine retrieval order.
      Parent relationships indicate primary sources with higher retrieval weight.
      Child relationships indicate derived documents — retrieve if parent is insufficient.
      Peer relationships indicate complementary context — retrieve in parallel.
---

# RAG Retrieval Pipeline

This document models a retrieval-augmented generation pipeline using MDA
footnote relationships as typed edges in the retrieval graph.

## Retrieval Graph Structure

Documents declare their position in the retrieval graph via typed
footnote relationships[^retrieval-spec]. The agent runtime traverses
this graph to assemble context windows[^context-assembly].

## Ranking Heuristics

- **Parent** relationships → primary sources, highest retrieval weight
- **Peer** relationships → complementary context, retrieved in parallel
- **Child** relationships → derived documents, retrieved on fallback

This ranking reduces hallucination on multi-hop queries compared to
pure semantic similarity, because relationships are author-asserted
rather than inferred by the embedding model.

[^retrieval-spec]: {"rel-type": "implements", "doc-id": "retrieval-spec-v1", "rel-desc": "RAG retrieval graph specification"}
[^context-assembly]: {"rel-type": "references", "doc-id": "context-assembly-v1", "rel-desc": "Context window assembly strategy for multi-hop queries"}