pending — blocked on Phase 2 close

Phase 3 — HippoRAG-lite Retrieval

Vector seed + networkx PageRank expand retrieval over the coupling graph. A query enters as a semantic vector, seeds a top-k set of cards, then PageRank expands through coupling edges to surface structurally adjacent machines. The retrieval subgraph is the context window for the reasoning model.

← Back to site map

Retrieval pipeline

Step 1

Query → semantic_emb (768d via gemma3)

Step 2

HNSW ANN search → top-k seed cards (k=5)

Step 3

networkx PageRank expand from seed set via coupling edges

Step 4

Re-rank by hybrid score: semantic sim + PageRank + state_emb proximity

Output

Top-N subgraph → context for Gemma 3 4B

Validation gate

Phase 3 closes when 20 hand-authored (query, expected-card-set) pairs achieve ≥0.80 recall@10 on the full corpus. Cross-era queries are weighted 2x in the evaluation set.