MDA in a RAG pipeline — how typed footnotes improved our retrieval precision

[landonkavanaugh55]

Sharing our experience integrating MDA into a production RAG system in case it's useful.

Setup: ~12K documents, mixed .md and .mda sources. We added rel-type annotations to about 30% of the corpus — the highest-value docs. The retrieval system uses the footnote graph to boost documents that are explicitly linked to the query's seed document.

Key finding: Precision at k=10 improved from 0.61 to 0.79 on our benchmark set after adding graph-based re-ranking on top of vector retrieval. The biggest wins were on queries that required following parent relationships across 2-3 hops.

Open question: Anyone handling documents where the relationship graph is incomplete? We fall back to pure vector search for unannotated docs, but that creates a two-tier retrieval quality problem.