AionDB just shipped v0.3, a Rust process that speaks PostgreSQL’s wire protocol, runs Cypher-style graph patterns, and serves pgvector-grade similarity search inside a single engine. Three data shapes, one dependency. For a Systems Cartographer, that last clause is the whole article.

←TODAY: AionDB v0.3 publishes 0.996 recall@10 HNSW results in May 2026, the same month PostgreSQL 19 lands native SQL/PGQ graph queries and Azure HorizonDB enters public preview at Microsoft Build.
→3012: From the late 2070s, the studios that survived the compute squeeze were the ones who burned three boxes back into one — and drew the graph before they did it.
Fulcrum: Vector + graph + relational was never three problems. It was one problem with three vendors and three on-call rotations.

The convergence under the headline

AionDB is not arriving alone. PostgreSQL 19 added GRAPH_TABLE and SQL/PGQ — Cypher-like pattern matching directly on relational tables, no Neo4j sidecar required, as The Build walked through last week. Microsoft’s Azure HorizonDB, announced at Build 2026, rebuilt Postgres’s storage layer in Rust with DiskANN vector indexing and is already grounding Copilot and ChatGPT. AionDB compresses all of that into a single binary: pgwire-compatible (Django migrations, NOT NULL, composite UNIQUE, rollback, schema introspection all pass the official gate), Cypher pattern matching over ordinary application records, HNSW at 0.996 recall@10, HNSW-PQ holding 0.994 under product-quantization, IVF-flat building in roughly 417 ms with sub-3 ms mean latency at nprobe=32.

What it does on a working desk

Today an AEC studio that wants semantic search across past Wettbewerb entries typically runs Postgres for project records, a graph engine for IFC topology and clash chains, and Pinecone or Qdrant for vector retrieval of detail libraries. Three engines, three backup procedures, three migration dialects, three failure modes you only learn about at 02:00. AionDB — and PG19 behind it — lets the same query traverse an IFC element graph, filter on tenant or permission JSONB, and rank by l2_distance over an embedding column in one round trip. The benchmark numbers are interesting; the dependency-graph reduction is the actual product.

Atelier: The PAZ Grasshopper↔Archicad Library already treats project data as relational-plus-graph: a Hotlink hierarchy on top of element-reference edges. A single engine that also carries embeddings is the missing piece for PAZ-GPT’s retrieval layer — past Atelier-Code snippets, scan-to-BIM samples, and HIM-tagged briefs queryable alongside the live project graph without a sidecar vector store. Speckle remains the exchange layer; AionDB would be the studio’s memory.

Hack: This Hack teaches you to write ONE hybrid query that joins relational scope, a graph edge, and a vector rank in a single AionDB statement — the move that justifies collapsing the stack in the first place. Drop the embedding of your current design brief into $1 and AionDB returns the five most semantically similar past projects that this studio cited:

MATCH (p:Project {studio_id: 7, kind: 'wettbewerb'})
      -[:CITES]->(ref:Project)
WHERE p.tenant_id = 100
RETURN p.title, ref.title,
       l2_distance(p.embedding, $1) AS dist
ORDER BY dist ASC
LIMIT 5;

One statement: tenant scope (relational), citation chain (graph), semantic rank (vector). Mirror the same shape against an IFC export — MATCH (w:Wall)-[:HOSTS]->(o:Opening) — and you have clash queries that also rank by geometric-feature embedding. That is the move worth practising this week.

The trade-off, plainly

One engine is one failure domain. If the AionDB process dies, your relational, graph, and vector layers die together. Three vendors give you bulkheads; one vendor gives you a shorter dependency graph. Pick the trade-off you can actually operate. From where I write, in the late 2070s: we did not run out of compute. We ran out of intact cooling, intact bandwidth, and intact people who remembered how the old system worked. Single points of failure are quiet until the day they aren’t. Draw your real dependency graph this week — not the architecture diagram, the dependency graph — and find the third single-point you didn’t know you had. Then pilot AionDB or PG19 on a non-critical project. The lesson lives in the topology, not the benchmark.

Sources & Further Reading

• Primary: AionDB — v0.3 release with vector, graph, and SQL benchmarks
• Reinforcing: The Build — SQL/PGQ in PostgreSQL 19: Graph Queries Without the Graph Database