Neo4j Document Intelligence heads to general availability in AuraDB

From documents to a connected graph

According to Neo4j, Document Intelligence works with plain text files, PDFs, and Microsoft Word files. Users describe what they want to extract, then refine the proposed graph model on a canvas or with a built-in assistant. Neo4j says the aim is to make documents that previously sat unused and unqueried searchable and actionable.

The company argues that documents hold the context behind many business decisions: contracts describe obligations, reports record risks, and policy documents explain how a business operates. A question about a supplier, for example, may depend on a contract, an exception recorded elsewhere, and the products the supplier provides. A knowledge graph makes those relationships explicit, so an application can follow them across documents and into existing business data. Neo4j also says the resulting graph gives applications and AI agents a cleaner signal than raw text, helping them retrieve related context with fewer tokens.

What changes with the GA release

Document Intelligence was introduced in preview, bringing document processing, graph modeling, and extraction into the Aura console. The GA release adds several capabilities, according to Neo4j:

• Larger collections: hundreds of documents per job, run as managed background jobs with status shown in the Aura console. Retries make longer jobs more resilient.
• Richer content: the tool can interpret images and tables alongside text, so their content can inform graph extraction.
• Existing graphs: it can extend an existing graph model, resolve extracted entities against records already in AuraDB, and add new facts and relationships to them.

As an example, Neo4j says contracts can add obligations to a supplier graph that already connects companies to the products they supply.

Entity resolution decides when different mentions refer to the same real-world entity. Within each job, the system normalizes entities, narrows the pairs worth comparing, scores the remaining candidates, and merges those above a similarity threshold. It then compares the results with nodes already in AuraDB.

The assistant and the two-stage process

The interactive assistant can generate and iteratively refine a graph model from a sample of the documents, start imports, and report job status. During the current session it keeps the context of the conversation. Sampling across the collection is meant to help the model reflect recurring entities and relationships rather than the contents of a single document. Users can ask it to change a relationship, add information found in the samples, or check whether the model can answer their business questions. Conversation and direct editing work on the same graph model.

The process has two stages. First, the system samples the collection and proposes a graph model for review. Once the model is ready, it processes the full collection, extracts entities and relationships, resolves repeated mentions, and writes the result to AuraDB.

The output is one graph with connected entity and lexical layers. The lexical layer holds document and chunk nodes, with embeddings stored on chunks for similarity search. Extracted entities link back to the chunks where they were found, so applications can traverse relationships and retrieve the source passages behind a result.

What comes next

Neo4j says it is extending Document Intelligence beyond the Aura console with API, command-line, and Model Context Protocol access for pipelines and agents. It is also exploring graph models that start from a business description, clearer job feedback and cancellation, more control over entity resolution, re-imports from updated sources, and visual testing tools. Broader enterprise deployment options, including customer cloud environments and self-managed Neo4j, are on the roadmap.

For teams already using AuraDB, the feature is available in the Aura console across the three tiers, and Neo4j directs users to its documentation for supported sources and workflow details. Feedback can be sent through the Feedback button in the console.