Neo4j Virtual Graph enters public preview for Snowflake, Databricks and BigQuery

Graph reasoning without moving the data

According to Neo4j, copying large enterprise datasets into secondary databases for graph analytics brings unnecessary costs, data drift and governance complexity. Virtual Graph is meant to avoid that trade-off by running graph reasoning where the data already sits. In Public Preview, its zero-copy architecture now supports Snowflake, Databricks and Google BigQuery.

Neo4j says teams can create working knowledge graphs in minutes. Built-in AI tooling analyzes existing data warehouse tables and automatically proposes a graph schema. When an AI agent or a user runs a Cypher query, Virtual Graph translates the request deterministically into optimized SQL and pushes the heavy compute down to the cloud warehouse engine. Data stays under existing perimeter controls, and results come back in graph form over the Bolt protocol.

As the product approaches General Availability, Neo4j says it gives system integrators and partners a zero-ETL entry point for GraphRAG and relational reasoning without costly data migrations.

Per-user identity for AI agents

Neo4j also published a multi-cloud reference architecture that connects Microsoft Copilot Studio to Neo4j using per-user Okta SSO. It targets what the company calls a "confused deputy" problem: an AI agent that queries a database through a single shared service account may hold broader authorization than the end user and return confidential data in response to an unpredictable prompt.

In the design, agent requests go through a Neo4j Model Context Protocol (MCP) server hosted on Google Cloud Run. Individual Okta OAuth access tokens are passed down to the database, where Neo4j validates the signed token against Okta and maps user groups to database-level roles. According to Neo4j, access is enforced at the database level even if an LLM generates a wide, multi-hop Cypher query, so users retrieve only the context they are permitted to see.

Neocarta benchmark on token costs

A benchmark on the Neo4j Developer Blog looked at Text2SQL agents pointed at large data catalogs. Neo4j describes two problems: heavy token use from pasting raw schemas into prompts, and hallucinations when agents guess which tables to join. Its answer is a semantic layer built with Neocarta, which collects catalog schemas, business terms and real join paths into a metadata knowledge graph. Agents then use semantic search over MCP to retrieve only the tables, columns and certified join paths they need.

Tested on a 278-table BigQuery schema and a 264-table Databricks Unity Catalog, the semantic layer, according to Neo4j:

• Cut token usage by 45% to 81% across multiple model families.
• Lifted cheap models such as GPT-4o-mini from failing dead-end loops to 100% accuracy.
• Doubled execution speed while saving hundreds of dollars per month in API token costs.

Customer deployments and partner activity

Neo4j reported co-sell activity with Google Cloud, AWS and Microsoft in Q3. On Google Cloud, a global property data company deployed 2 TB of Aura VDC to unify fragmented property and tax datasets, and an AI software provider expanded from about 1.9 TB across seven environments to an enterprise production deployment that models software dependencies for AI agents.

On AWS, a US mortgage and financial services lender scaled to an 8 TiB Aura VDC deployment and pairs Graph Data Science and GraphRAG with Amazon Bedrock. Neo4j says the setup supports sub-second context retrieval for autonomous sales agents and achieved a 20% lift in borrower follow-ups. On Azure, a major telecommunications carrier renewed and expanded its footprint for security workflows and customer churn prediction.

For Q4, Neo4j plans regional hands-on AI workshops through October and November with AWS, Google Cloud, Microsoft, Snowflake and Databricks.

What it means for enterprise teams

The announcements center on two concerns Neo4j names for production AI: token cost and data security across clouds. Virtual Graph is still in Public Preview, so General Availability and any pricing details have not been stated in the source. Teams already holding data in Snowflake, Databricks or BigQuery can test graph queries without migrating it, while the Okta and Copilot Studio design gives security teams a reference for enforcing user-level access. The token savings figures come from Neo4j's own benchmark and should be checked against each team's catalogs and models.