Warehouse GraphRAG — ClickGraph & DeltaGraph

01 · BASELINE

Who are the most influential users?

Rank users by how many people follow them.

1 hop · aggregate vector store: roughly

read_neo4j_cypher→ warehouse

MATCH (u:User)<-[:FOLLOWS]-(f:User)
RETURN u.name AS influencer, count(f) AS followers
ORDER BY followers DESC, influencer LIMIT 5

5 rows

influencer	followers
Rachel	5
Tina	5
Xander	5
Alice	3
Jack	3

02 · RECOMMENDATION

Who should Alice follow?

Friends of friends she doesn't already follow — ranked by mutual connections.

2 hops · anti-join vector store: no

read_neo4j_cypher→ warehouse

MATCH (me:User {name:'Alice'})-[:FOLLOWS]->(f:User)-[:FOLLOWS]->(fof:User)
WHERE fof <> me AND NOT (me)-[:FOLLOWS]->(fof)
RETURN fof.name AS suggested, count(DISTINCT f) AS mutual_connections
ORDER BY mutual_connections DESC, suggested LIMIT 5

3 rows

suggested	mutual_connections
Ben	1
Henry	1
Victor	1

fof <> me is node-identity comparison — it resolves to the schema's user_id column, not a literal .id, so it runs unchanged on ClickHouse and Databricks (fix #1076).

03 · CONTENT GROUNDING

What is Alice's network talking about?

Pull the posts authored by the people Alice follows.

2 hops · scoped vector store: no

read_neo4j_cypher→ warehouse

MATCH (me:User {name:'Alice'})-[:FOLLOWS]->(f:User)-[:AUTHORED]->(p:Post)
RETURN f.name AS author, p.content AS post
ORDER BY p.created_at DESC LIMIT 5

5 rows

author	post
Xander	Graph neural networks
Tina	Subqueries and comprehensions
Tina	UNION queries in Cypher
David	ClickHouse integration patterns
David	Building social networks with graphs

This is GraphRAG, literally. The answer is grounded in warehouse rows reached by traversal, scoped to Alice's neighborhood — not a fuzzy top-k over the whole corpus.

04 · CONNECTION PATH

How is Alice connected to Rachel? To Ben?

Find the shortest chain of relationships between two specific people.

variable-length vector store: no

read_neo4j_cypher→ warehouse

// read the whole chain, by name
MATCH path = shortestPath((a:User {name:'Alice'})-[:FOLLOWS*1..5]->(b:User {name:'Rachel'}))
RETURN [n IN nodes(path) | n.name] AS connection_path, length(path) AS hops

1 row

Alice→David→Henry→Sam→Rachel

hops = 4 — the names come back in traversal order, materialized as a parallel array carried alongside the path.

Path readout by name (#1079) shipped from building this demo — the third real bug the MCP loop surfaced. [n IN nodes(path) | n.name] now returns the ordered chain, verified live on ClickHouse and Databricks.

05 · ENGAGEMENT

Which posts are resonating, and who wrote them?

Rank posts by distinct likers, joined back to their authors.

3-way join vector store: no

read_neo4j_cypher→ warehouse

MATCH (author:User)-[:AUTHORED]->(p:Post)<-[:LIKED]-(liker:User)
RETURN author.name AS author, p.content AS post, count(DISTINCT liker) AS likes
ORDER BY likes DESC, post LIMIT 5

5 rows

author	post	likes
Uma	Graph machine learning	5
Grace	Graph algorithms overview	4
Ben	Graph query languages comparison	4
Leo	Indexing strategies for graphs	4
Iris	Real-time graph analytics	4

06 · HYBRID — VECTOR + GRAPH

Semantic entry, then graph expansion

Use similarity to find the entry point; use the graph to traverse, explain, and re-rank. The question ≈ "deep learning spanning ML and NLP", encoded as a query vector.

vector + traversal vector store: step 1 only

1 VECTOR — what is the question about?

read_neo4j_cypher→ warehouse

MATCH (t:Topic)
RETURN t.name AS topic, round(1 - cosineDistance(t.embedding, $q), 4) AS similarity
ORDER BY similarity DESC LIMIT 3

3 rows

topic	similarity
Natural Language Processing	0.6844
Machine Learning	0.6844
Databases	−0.0001

The two AI topics surface; unrelated topics sit at ~0. A vector store stops here.

2 GRAPH — how does that topic connect, by name?

MATCH path = (start:Topic {name:'Machine Learning'})-[:RELATED_TO*1..2]->(t2:Topic)
RETURN [n IN nodes(path) | n.name] AS topic_chain, length(path) AS hops

3 rows · path readout #1079

Machine Learning→Natural Language Processing→Computer Vision

[n IN nodes(path) | n.name] reads the neighborhood back in traversal order — similarity search cannot represent this.

3 HYBRID — re-rank the grounded neighborhood

MATCH (t:Topic)-[:DISCUSSES]->(d:Document)
WHERE t.name IN ['Machine Learning','Natural Language Processing','Computer Vision']
RETURN d.title AS document, round(1 - cosineDistance(d.embedding, $q), 4) AS similarity
ORDER BY similarity DESC LIMIT 5

top 3 of 5

document	via_topic	similarity
Introduction to Neural Networks	Machine Learning	0.6844
Transformer Architecture Explained	Natural Language Processing	0.6844
Object Detection with YOLO	Computer Vision	−0.0016

The graph scopes the candidates; the vector re-ranks them. The two foundational ML/NLP papers rise; the off-topic CV paper sinks. Relevant and explainable.

4 ALL AT ONCE — the whole loop in one statement

MATCH (ml:Topic {name:'Machine Learning'})-[:RELATED_TO*1..2]->(t2:Topic)-[:DISCUSSES]->(d:Document)
RETURN DISTINCT d.title AS document, t2.name AS via_topic,
       round(1 - cosineDistance(d.embedding, $q), 4) AS similarity
ORDER BY similarity DESC

3 rows · unblocked by #1083

document	via_topic	similarity
Transformer Architecture Explained	Natural Language Processing	0.6844
Deep Learning for NLP	Natural Language Processing	0.0943
Object Detection with YOLO	Computer Vision	−0.0016

Variable-length graph expansion feeds a vector re-rank in one Cypher statement — no client round-trip between traversal and ranking. The NLP-adjacent paper rises; the CV paper sinks.

ClickHouse-native vectors. cosineDistance / L2Distance / gds.similarity.* and CALL db.index.vector.queryNodes(…) run on the ClickGraph/ClickHouse endpoint. Variable-length expansion feeds the chained DISCUSSES hop directly — no WITH barrier — which composes once the anonymous-alias collision (#1083, fixing #1081) is in place. One naming caveat: the path start node is ml, not t — a variable-length-path endpoint named exactly t collides with ClickGraph's reserved internal VLP-CTE alias and is rejected with an actionable error (#1085); any other name resolves correctly.

Scenario	Traversal	Could a vector store do it?
01 · Influencers	1 hop, aggregate	roughly
02 · Recommendation	2 hops + anti-join	no — answer is in graph shape
03 · Content grounding	2 hops, scoped	no — needs edge-scoping
04 · Connection path	variable-length	no — not a vector
05 · Engagement	3-way join	no — relationship aggregation
06 · Hybrid vector + graph	vector seed → traverse → re-rank, one statement	step 1 only — traversal & re-rank compose in one query

An AI agent queries your warehouse as a graph.

Two tools are all it takes

Discover

Translate

Traverse

Ground

What traversal does that similarity can't

Who are the most influential users?

Who should Alice follow?

What is Alice's network talking about?

How is Alice connected to Rachel? To Ben?

Which posts are resonating, and who wrote them?

Semantic entry, then graph expansion

Same warehouse tables. Zero copies.

Swap the port. Same Cypher runs.