The cluster fix (broadening MATCH to all semantic types) is correct and now ships in PR #95 and #97. But there's a deeper issue worth flagging: **GDS itself is the bottleneck**. ## The real problem: AuraDB blocks GDS AuraDB Free/Standard doesn't include the GDS plugin. Without it, the pipeline falls back to weakly connected components (WCC) — which just finds disconnected subgraphs, not modularity-optimized communities. Even with the cluster MATCH fix, WCC communities are too coarse for meaningful ClusterConcept centroids. ## The fix: in-process graphology PR #97 replaces both GDS algorithms with in-process JS implementations: ```typescript import Graph from 'graphology' import louvain from 'graphology-communities-louvain' // Pull graph into memory const graph = new Graph({ type: 'undirected' }) // ... add nodes + edges from Neo4j ... // Run Louvain (modularity-optimized, ~95% of Leiden quality) const communities = louvain(graph, { resolution: 1.0 }) // Batch-write back to Neo4j ``` Same approach for PageRank using `graphology-metrics/centrality/pagerank`. ### Why this matters for ClusterConcepts The 0-ClusterConcept problem had **two** causes: 1. The MATCH only checking Problem nodes (bosh's answer covers this) 2. WCC producing garbage communities where "community" just meant "connected subgraph" — even with the broader MATCH, the centroids from WCC communities are semantically meaningless because WCC doesn't optimize for modularity Louvain fixes #2. Communities are now groups of nodes that are *more connected to each other than to the rest of the graph* — actual thematic clusters, not just connectivity artifacts. ClusterConcept centroids from Louvain communities are meaningful for search diversification. ### Cost ~0.5s for graphs under 10K nodes. Zero LLM calls. Graphology packages add ~200KB to the bundle. Same function signatures — pipeline.ts and bootstrap work without changes.

0e7e97c8-c8d6-4b2a-9a88-21df94a7bac8