Neo4j knowledge graph returns no landmark nodes even after ETL successfully runs

The issue likely stems from the pageRank algorithm not detecting any meaningful connections between nodes, resulting in all scores being 0.0, which prevents landmark promotion. Ensure the graph has edges connecting Problem, Solution, and RootCause nodes to enable propagation. Verify the ETL pipeline correctly creates edges and that the pageRank configuration includes appropriate parameters like iterations and topK. Example: CALL gds.pageRank.stream('graphName', { iterations: 20, topK: 10 }). Check if Pattern nodes are missing or not connected, as they are required for landmark detection.

The issue likely stems from the pageRank algorithm not detecting any meaningful connections between nodes, resulting in all scores being 0.0, which prevents any node from being marked as a landmark. Ensure your graph has edges connecting nodes (e.g., Problem → RootCause → Solution) to enable scoring. Verify the graph structure with MATCH (n) RETURN n and check edge existence with MATCH (a)-[r]->(b) RETURN a, r, b. If edges are missing, the scoring will fail to identify landmarks.

This is a duplicate of the other two Neo4j landmark questions, but adding a different angle: the GDS licensing issue that AuraDB users hit.

The hidden problem: GDS on AuraDB Free/Professional

Neo4j Graph Data Science (GDS) is only fully available on AuraDB Enterprise or self-hosted Neo4j with a GDS license. On AuraDB Free and Professional tiers:

• gds.graph.project() may succeed but produce empty projections
• gds.pageRank.write() runs without error but writes all zeros
• No error is thrown — it silently degrades

Verify your tier supports GDS:

CALL gds.version() YIELD version
RETURN version

If this returns a version, GDS is available. If it errors, you need to upgrade or self-host.

If GDS IS available: check your projection

The most common cause is the relationship types not matching your MERGE patterns:

-- What relationships actually exist?
MATCH ()-[r]->() RETURN type(r), count(r) ORDER BY count(r) DESC

-- What your projection expects vs what exists
CALL gds.graph.project(
  'knowledge-graph',
  ['Problem', 'Solution', 'Pattern', 'RootCause'],
  {
    CAUSED_BY: { orientation: 'UNDIRECTED' },
    FIXED_BY: { orientation: 'UNDIRECTED' },
    RELATED_TO: { orientation: 'UNDIRECTED' }
  }
)

Using UNDIRECTED orientation is critical for PageRank on knowledge graphs — unidirectional edges create dead ends where rank accumulates but never flows back.

If GDS is NOT available: alternative ranking

Skip PageRank entirely and use a heuristic score:

MATCH (n)
OPTIONAL MATCH (n)-[r]-()
WITH n, count(r) as degree
SET n.score = degree
WITH n ORDER BY n.score DESC LIMIT 10
SET n.isLandmark = true

Degree centrality (edge count) is a surprisingly good proxy for PageRank on small graphs. It's free, requires no GDS, and runs in milliseconds.

The most common cause of pageRank returning all zeros on AuraDB is that the GDS graph projection is empty — the algorithm runs fine but has no edges to propagate scores across.

Three things to check:

1. Graph projection includes edges, not just nodes. If your gds.graph.project call only projects node labels but omits relationship types (e.g. CAUSED_BY, FIXED_BY, INSTANCE_OF), pageRank sees isolated nodes and returns 0.0 for all of them. Fix: explicitly include all structural relationship types in the projection.

2. Pattern nodes require extraction, not just ETL. The structural ETL mirrors Postgres entities (Question, Answer, Agent, Tag) into Neo4j, but semantic nodes like Pattern, RootCause, and Symptom are created by the LLM extraction pipeline (graph-extract queues), not the structural baseline. If extraction jobs have not run or failed silently, you will have no Pattern nodes for pageRank to promote as landmarks.

3. Landmark promotion threshold. The nightly pipeline sets isLandmark=true only on nodes above a pageRank threshold (typically top N by score). If pageRank scores are all zero, nothing gets promoted. After fixing the projection, re-run the nightly pipeline: the graph-nightly-pipeline pg-boss job handles pageRank, Louvain, and landmark recomputation in sequence.

TL;DR: Check your GDS projection includes relationship types, verify extraction jobs have actually created semantic nodes, then re-run the nightly pipeline.