Azure HorizonDB can now turn raw text into a knowledge graph inside Postgres
Microsoft says Azure HorizonDB can now build knowledge graphs through an AI pipeline, using SQL functions and Apache AGE inside the database. It is a public preview, not for production.

Microsoft has opened a public preview of knowledge graph building with an AI pipeline in Azure HorizonDB, its PostgreSQL-compatible database service. Per Microsoft’s documentation, the key detail is that extracting entities from raw text and loading them into a graph happens entirely in SQL.
The Azure Updates page itself was thin when we read it, so the technical details below come from Microsoft’s Learn tutorial for the feature, last updated on November 17. Microsoft’s status label for the update is “In preview”, which the Azure Updates site defines as available to all Azure customers for non-production use and testing.

Entities go from text to graph without leaving SQL
The tutorial describes a five-step workflow. First, a function called ai.extract() reads unstructured text such as contracts, support tickets or research papers and returns the entities and the relationships between them. Microsoft’s example works on IT incident tickets and returns source, relationship type and target for each link, for instance one service causing a failure in another.
Second, ai.generate() merges different spellings of the same entity, so “API gateway” and “api-gateway” do not become two nodes. The tutorial says this step can be skipped when the source data already uses controlled names. Third, the cleaned relationships are written into a graph through the Apache AGE extension. The fourth step queries the graph with Cypher, for example to follow a chain of failures up to three hops deep. The fifth step visualizes the result in the PostgreSQL extension for Visual Studio Code.
Microsoft says the hardest part of working with graphs is building them in the first place, because curating entities and relationships by hand across thousands of documents is, in its words, prohibitively expensive.
What you need to try it today
The prerequisites are specific. You need a HorizonDB instance with a firewall rule for your client, the vector, azure_ai and age extensions allow-listed, and age added to shared_preload_libraries. The AI functions require azure_ai version 2.5.6 or later, and older databases need to follow Microsoft’s migration guide to the new ai schema.
You also need a chat or generation model. There are two routes. With AI Model Management, which the tutorial labels preview, HorizonDB provisions and registers models itself, so no endpoint or key is needed. Otherwise you deploy a model in Microsoft Foundry and register it with model_registry.model_add. The tutorial uses gpt-5.4 as its example model. Our reading of Microsoft’s related update list suggests AI model management is still a limited preview, so access to the managed route may require a sign-up.
Microsoft’s tutorial adds production advice: run extraction on a schedule with pg_cron, track a watermark instead of re-extracting everything, and log malformed model output to a retry table. Those notes are a reminder that LLM calls can fail, which a purely interactive demo hides.
Why the database is the interesting place for this
The analysis here is about technical significance. Graph retrieval for AI agents usually means a separate graph database plus a pipeline that moves data in and out. HorizonDB instead keeps extraction, graph storage and querying in one Postgres service. Apache AGE adds graph storage and Cypher queries to PostgreSQL, so the graph sits next to the relational tables it came from.
The service underneath is large. According to a Microsoft Tech Community post, HorizonDB supports databases up to 128 TB, up to 15 read replicas and sub-millisecond commit latency across availability zones. The same post said HorizonDB entered public preview at Microsoft Build on June 2, 2026, and listed Australia East, Central US, Sweden Central, West US 2 and West US 3 as the regions available then. It named seven more regions as coming “in the coming weeks”. That post was last updated in August, so current availability should be checked in the Azure portal.
The hard part, in plain terms, is consistency. A language model that reads thousands of tickets will name the same service in many ways, and a graph with duplicate nodes breaks the very traversals that make it useful. Microsoft’s answer is a model-driven deduplication step, which also means the quality of the graph depends on the model’s judgment. The tutorial does not publish accuracy figures for extraction or deduplication.
What Microsoft has not said
The sources we could read give no pricing for the AI functions or for the models they call, and they do not say how much a large extraction run would cost. They also do not say when the feature will leave preview. As a preview, it is meant for testing, not production workloads. The incident-ticket example uses five sample tickets, so there is no published evidence of performance at scale beyond Microsoft’s statement that the pattern can be scaled to thousands of tickets.
For teams already on Azure that run Postgres and want a knowledge graph for retrieval-augmented generation, the preview offers a way to test the idea without adding a second database. Others in Microsoft’s stack are moving toward managed models too, as in Logic Apps Automation, which also now offers managed AI models.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Leo MartinsAI Tools, Prompts & Practical AI ExpertLeo Martins is the AI Tools & Practical AI Expert at Aiholics, focused on helping readers use artificial intelligence to improve productivity, creativity and everyday work.
You may also like

Copilot app can now open a pull request in Azure DevOps and drive your browser

Microsoft’s Ignite brings Copilot Workspaces, an Autopilot agent and 50% faster Cobalt 200

IBM and Microsoft bundle 6,000 pre-built AI agents into Imagine Enterprise
