What is Neo4j Generative AI assistance?

Neo4j's Generative AI Assistance features are available in the Aura console and are designed to assist users in creating, exploring, and querying their graphs.

Currently we provide assistance in three areas:

• Explore Copilot provides assistance generating search phrases from natural language requests.
• Query Copilot helps beginners learn and construct Cypher queries while also saving time for experienced users by providing a solid starting point for their queries.
• Import modeling and mapping assistance help turn data sources into graph models and map date to those models.

How Does It Work?

All of these features are powered by Google's Vertex AI platform using Large Language models (LLM). You can learn more about Google's data governance and responsible AI commitments in the Google Cloud Documentation.

In the case of copilots, users enter a natural language query and receive search phrases or Cypher code generated by the model. The LLM prompt includes: the user's natural language query, a description of the database schema, and few-shot examples and guidelines. This means every request will have a reasonable understanding of your database. In the case of Query Copilot, users can edit and adjust the generated Cypher code before executing it. Users can also adjust their natural language question to refine the Cypher code generated.

In the case of import, information such as source table, column names, and types is used to approximate a graph model and data mapping.

How are Generative AI assistance features enabled?

Generative AI assistance features are customer-controlled features at the organization level. Organization admins can configure an organization-wide setting in the Aura Console Organization settings:

On first use of a Generative AI assisted feature, all end users are made aware of how to identify these features, how the feature operates, how results should be treated with caution, and how it can be disabled:

Is my data used to train AI models?

No, your data is not used to train AI models.

Should I worry about compliance issues?

No. Neo4j maintains a certified Information Security Management System based on the ISO 27001 standard, is SOC 2 Type 2 compliant, and has been assessed against HIPAA requirements. Neo4j also has implemented policy and process in alignment with the NIST Artificial Intelligence Risk Management Framework and has self-assessed against ISO 20243 requirements that address specific threats to the integrity of software products throughout the product life cycle. You can obtain more details on Neo4j's compliance efforts here.

What are some limitations of the Generative AI features?

Query copilot features provide very valuable assistance but some generated Cypher queries may not be entirely correct. Here is a list of issues you may find with the generated queries:

• Incorrect relationship direction: The database schema may not always be fully observed, leading to reversed relationship directions.
• Hallucinating "model shortcuts": The Copilot may create relationships that do not exist in the actual database schema.
• Deprecated syntax: Generated queries might use outdated Cypher syntax that fails to run.
• Variable naming clashes: The Copilot might unintentionally reuse variable names, resulting in incorrect values.
• Inefficient queries: Some queries might be computationally expensive due to unbounded traversals or unlimited queries.
• Lack of relationship returns: Queries may not include relationships in the returns, requiring additional prompting to include them.

Assisted Import features attempt to convert your source data into a graph model and map your data to it automatically. This may not always create an optimal graph model and can result in incorrect mappings. 

Where can I learn more about these features?

Check out this blog post: Introducing Your New Cypher Copilot.

It features step-by-step instructions to turn on the Query Copilot and walks through a couple of examples showing how to use it.