What Microsoft announced
Microsoft says HorizonDB is generally available, AI Model Management and Knowledge Graphs are in Public Preview, and Database Agent for PostgreSQL is in Limited Public Preview, with access by request. The HorizonDB service is production-ready, but the graph feature is not. Don't read the preview label on one feature as applying to the whole service, or the other way around.
The Community Hub post uses a supplier-delay scenario. A key supplier announces a delay, and planners need to know which parts and open orders are affected and whether another supplier can cover them. In that post, Knowledge Graphs builds the graph "straight from supplier notices, manuals, and other documents". The map grows as new documents arrive, so nobody has to write the Cypher by hand.
How it works, based on Microsoft's example
The workflow in the example has four parts:
- Start with an ontology. This is the set of entity types, properties and relationships your application expects.
- Register it with
ai.create_graph(). This call also creates the underlying Apache AGE graph. - Anchor node identity on IDs you already trust. In the sample,
SupplierandPlantare node labels andSHIPS_FROMis an edge between them. Suppliers are identified by a supplier ID, and the example carries adocument_idon the edge. - Run pipelines. The example uses one pipeline to extract facts from notices into a table. A second pipeline maps those facts to AGE nodes and edges with
ai.graph_sink(). Each node MERGEs on its identity, so a repeated fact lands on the same node. Initial population is a manualai.run()call.
The example also normalizes messy extracted names. Variants such as "Plant 2", "plant 2" and " PLANT 2 " all need to resolve to the same node. Microsoft notes that for harder cases, such as abbreviations, you should anchor nodes on IDs you already trust.
The payoff in the demo is a query that joins two separate notices. One says Supplier B's valves ship from Plant 2. The other says Plant 2 is paused for maintenance. A keyword or vector search would struggle to surface that dependency. A graph traversal gets it directly, and each edge points back to the notice that asserted it.
The "AI pipeline" piece
The title mentions an AI pipeline, so it helps to know what Microsoft's documentation says about pipelines.
- Pipelines are part of the
azure_aiextension. The docs describe them as a higher-level, AI-shaped surface of sources, steps, sinks and triggers that compiles down to a durable graph automatically. - Execution is built on
pg_durable. The Learn page says runs retry failed steps, checkpoint incremental work and resume long jobs from the last completed step. - Creating a pipeline only registers its definition. It doesn't process existing rows, even with an
on_changetrigger. You callai.run()for the initial load, and later source inserts launch change-triggered runs. - Pipelines run on the primary (read-write) endpoint. Reader nodes can't create or run them.
- Sources are HorizonDB tables. Blob-storage or external content has to be loaded into a staging table first.
- The Learn docs warn that pipeline state isn't portable across major versions of
pg_durableduring preview. Drain or pause running pipelines before upgrading. - The
azure_aiextension needs version 2.5.6 or later for theai.*functions used in these examples.
Prerequisites from the walkthrough
The Community Hub post's walkthrough is aimed at a development database. It lists the following setup steps:
- Add the needed extensions (
pg_textsearch,pg_durableand AGE) to the allowlist and toshared_preload_libraries, keeping existing entries. - Apply the parameter group and restart.
- Have model aliases available, such as the defaults from AI Model Management (
default-chat,default-embeddinganddefault-reranker). - Microsoft also advises checking each preview feature's documentation for supported versions, regions and prerequisites before you try it.
When a model is called, HorizonDB sends the prompt and inputs to your configured Microsoft Foundry endpoint. Permissions, endpoint access, network configuration and data handling remain part of your application design.
Caveats to weigh
- Preview means non-production. That follows from Azure's own definition of the "In preview" status.
- Extracted edges are candidates. Microsoft's own advice is to treat them as candidate evidence and to validate identifiers and relationship types before using them operationally. LLM extraction can misread or invent relationships, so check a sample against the source documents.
- Costs can repeat. The pipeline docs say retries and reruns of changed rows can repeat billable extraction calls. Budget accordingly if you run extraction over a large corpus.
- AGE has limits. Microsoft's Graph-RAG page lists known limitations for Apache AGE. It implements only a subset of openCypher. Deep variable-length traversals beyond three or four hops can blow up. Graphs can't be queried across databases. Those notes were written before this preview announcement, so check them against current documentation.
- Scope of what's confirmed. I could not confirm regional availability, pricing or the exact preview limits from the Azure Updates entry. The Community Hub post doesn't cover them either.
Why it matters
Vector search finds text that sounds like your question. It struggles when the answer depends on relationships, such as which supplier ships from which plant, which plant is down, and which orders depend on both. Microsoft's Learn guidance frames graph-augmented RAG as most useful for multi-hop, relationship-driven questions, and suggests vector-only RAG for simple lookups. The same Learn page cites a benchmark on a 500,000-case U.S. case-law dataset. It reports recall rising from 40% to 70% with citation traversal. That is a Microsoft-published figure on one dataset, and your results will depend on your data.
The more practical selling point is architectural. Graphs, relational tables, vectors and full-text search share one set of transactions and permissions in Postgres. An approved change can commit across relational and graph data as a single unit. For teams already running Postgres, that could mean one less system to sync. Whether the managed extraction quality justifies the move is something only a pilot on your own documents will show.
A sensible way to try it
- Use a development cluster with a small document set you can inspect.
- Define a narrow ontology with two or three entity types.
- Anchor identities on IDs from your operational tables.
- Run the extraction pipeline once with
ai.run()and compare the output with the source notices. - Look for duplicate or misspelled entities and wrong relationship types.
- Only then add change triggers and combine graph traversal with vector or hybrid search.
For admins, the main practical point is that a preview feature is a place to experiment, not something to put in front of planners yet. The quality of your ontology and identity design will matter more than the preview label.
References
- (In preview) Public Preview: Build Knowledge Graphs with AI pipeline in Azure HorizonDB Azure Updates · 2026-11-17T20:00:32Z
- Azure HorizonDB: Data, Transactions, and AI in One Postgres Database | Microsoft Community Hub techcommunity.microsoft.com
- Graph-Augmented RAG Patterns in Azure HorizonDB - Azure HorizonDB | Microsoft Learn learn.microsoft.com