GraphRAG over the Semantic Server

link → grounded plan → retrieve/expand → verbalise, with an honest similarity fallback. Facts carry provenance; similarity carries a caveat.
Methodology — what each strategy is
strategyselects records bywhat it proves
vectorF02 $semantic-search, plain cosine top-100, as served todaythe deck's "bundle vectors" baseline
vector_walkthe same seeds, then each seed's one-hop nodes and edges are fetched and RAG'dvector RAG done properly — the context walk, anchored by similarity
vector_enriched$semantic-search?_enrich=true (vector 0.6 + graph degree 0.2 + recency 0.2)F02's own hybrid re-rank
graphlink → grounded OQL/GraphQuery → structural retrieval + assembly; refuses when it cannotthe pure structured path
graph_fallbackgraph, then on 4 triggers a hybrid lookup: exact text faces first, then similarity — always labelled similaritywhat you would ship
All strategies assemble facts and verbalise the same way, so the bench (graphrag bench → out/bench.md) compares retrieval. retrieval_derivation_complete is true only for graph answers.