New! Relate by Neo4j | An embedded database for AI agents Learn More

NODES 26 — November 12, 2026

Document Intelligence: Agentic Workflows for Knowledge Graphs Speakers

Session track: Modern Applications

Session time:

Session description:

urning documents into a useful knowledge graph requires a coherent extraction model, consistent entity identities and links to supporting source passages. An agentic experience brings these operations into a workflow you can guide and inspect. Firat Tekiner and Estelle Scifo will demonstrate how the Document Intelligence agent in Neo4j Aura uses tools to inspect sampled documents, propose and refine graph models, assess their fit for business questions, start imports and retrieve job status. You will see how conversation, canvas editing and extraction previews work together to give you control over the process. They will explain the data engineering behind this experience, from parsing text, images and tables to chunking, embeddings, model-guided extraction and background import. Entity resolution will be a central focus. You will see how the pipeline normalizes names, compares candidate matches and resolves repeated mentions across documents and against existing Aura nodes. They will discuss how duplicate entities and incorrect merges affect downstream queries. Using Cypher and vector search, they will query the connected entity and lexical layers to retrieve facts and supporting passages for downstream AI agents. You will learn how to guide the agent, design the extraction model, inspect entity matches and preserve source provenance when enriching an existing knowledge graph.

Meta Knowledge Graph: A Self Learning Agentic System

Session track: Data Intelligence

Session time:

Session description:

This talk presents a practical architecture for building harness-agnostic intelligent agent systems that can persist context, retrieve knowledge, and improve across sessions. Rather than treating memory as a standalone database, it frames agent intelligence as a loop: tools provide on-demand access to enterprise data and memory, hooks passively capture execution traces, metadata layers help agents discover and reason over available knowledge, and background consolidation turns raw activity into durable learnings and prompt improvements. The session will cover the core design patterns, tradeoffs between agent-driven and harness-driven orchestration, and how a knowledge graph can act as the shared substrate for provenance, retrieval, memory, and self-improvement.

Speaker

photo of Firat Tekiner

Firat Tekiner

Director AI Product Management

Firat Tekiner, PhD, MBA, is Senior Director of AI Product Management at Neo4j. Previously Senior Staff Product Manager for Data Analytics and AI at Google Cloud, he led Data Agents work in BigQuery. He has over 25 years’ experience building products and information systems for major research, education, telecommunications, finance and retail organisations. Earlier roles include National Supercomputing Services, the National Centre for Text Mining and Ab Initio Software. He has over 50 publications across parallel computing, big data, AI and computer communications, and co-authored Architecting Data and Machine Learning Platforms (O’Reilly, 2023) and Architecting Data Systems for AI (O’Reilly, exp 2026).