Neurosymbolic Architecture in Production with Neo4j, GDS, Node2Vec & LLM Grounding

Session Track: Graph Memory & Agents

Session Time: 01:00 pm - 01:30 pm April 15

Session description

In this session, João Cunha will present the neurosymbolic architecture used by Kipon to operate a production-grade performance management system built on Neo4j. The talk will show how a highly relational domain—composed of people, skills, tasks, feedback, projects, and operational interactions—is modeled as an operational graph capable of supporting multi-hop queries, structural inference, and low-latency retrieval.

João will detail how Kipon runs Jaccard Similarity, Node Similarity, PageRank, and Louvain/Leiden via AuraDS directly against the live graph to infer structural similarity between profiles, detect functional communities, identify knowledge hubs, and drive allocation and development decisions.

He will also describe Kipon’s hybrid embedding pipeline:

(1) structural embeddings generated with Node2Vec (biased random walks) to capture graph topology;
(2) semantic embeddings generated by large language models to encode textual attributes of skills, roles, and feedback;
(3) a unified vector space that integrates structural and semantic signals for concept-level proximity, similarity search, and clustering of heterogeneous entities.

Building on these structural and semantic layers, the session will then show how Kipon uses a proprietary RDF-inspired ontology as a symbolic grounding layer that constrains the inference space of LLMs. This layer anchors model outputs to the graph’s schema and permissible relations, reducing hallucinations, increasing factual consistency, and enabling full explainability through explicit path tracing. João will present how this ontology interacts with the embedding pipeline and the GDS outputs, along with the underlying data model, Cypher query patterns, engineering decisions, latency benchmarks, and best practices for deploying graph-powered neurosymbolic AI pipelines in production with robustness, predictability, and auditability.