Personalized PageRank Graph Retrieval (HippoRAG-2-style)

File: src/search/graph_ppr.rs

The problem

Vector + lexical retrieval find a chunk that contains the answer, but they cannot follow a multi-hop association (“who works at acme and reports to carol?”). Graph retrieval walks the knowledge graph to bridge that gap — yet a naive BFS over a noisy graph returns garbage.

The reference

HippoRAG 2 (OSU-NLP-Group/HippoRAG): a Personalized PageRank over the entity graph as an additional retrieval leg, fused with the dense/lexical results. Verified verbatim against the reference: igraph.personalized_pagerank(damping=0.5, directed=False, weights='weight', reset=node_weights).

The implementation

src/search/graph_ppr.rs is a pure-Rust CSR sparse graph with power iteration, faithful to the reference:

• PPR_ALPHA = 0.5 (the reference’s real default, not the 0.85 some drafts quote), PPR_EPSILON = 1e-6, MAX_PPR_ITER = 50, MAX_VISITED = 256.
• No LLM, no new schema, no embeddings in the graph leg — the < 5 W manifesto holds. Edge weight = COUNT(DISTINCT knowledge_id) per pair, scaled by relation-type (see the Discern explainer).
• Seeds = query→entity-name containment via the existing linker vocabulary; top entities expand back to chunks (respecting flagged=0 / valid_to IS NULL visibility).
• Opt-in ?graph=true as a third RRF leg (RRF_K = 60, rank-based, shared with the in-domain fusion) — the disabled path pays zero latency.

Measured ceiling

• Live multi-hop quality is corpus-bound. On the working 8.5k-doc DB ~94% of KG edges are tagged_with taxonomy noise; the mechanism ships but the cleanest multi-hop paths were the synthetic bench fixture. Corpus quality is an operator concern (vault re-ingest with the v1.4.1 heading-hierarchy linker grows the semantic edge set). This drove the v1.12 “Discern” fix.
• No DPR passage scores in the seed (an embedding in the leg is out of scope); PASSAGE_NODE_WEIGHT = 0.05 documents the upgrade path.
• Cross-domain graph federation is v2.0 work.

See 02-submodular-packing.md for how PPR output feeds the budgeted evidence set.