SUMM

Stephen Chin introduces himself as head of developer relations at Neo4j and mentions writing a new book on graph RAG with O'Reilly.
The talk will cover agentic GraphRAG and how to leverage it.
Attendees are polled about their experience with graph databases and graph RAG.

Current agentic systems often fail to meet use cases and are prone to hallucinations and biases.
An example shows an LLM making incorrect assumptions and gender-biased decisions when reasoning about a classroom scenario.
Such issues can lead to incorrect business outcomes, especially in high-stakes domains like life sciences or supply chain management.
LLMs provide surface-level intelligence but lack true reasoning capabilities.

Knowledge graphs can provide deeper reasoning and structure lacking in standard LLM responses.
Agentic systems improve result quality by using multiple agents (LLMs) in orchestrated workflows.
However, traditional agentic architectures are monolithic, difficult to maintain, and hard to secure.

MCP (Multi-Channel Platform) serves as a tool orchestration layer, with agents interacting through data sources and servers.
Neo4j has developed tools on MCP, including:
- A Cypher tool to generate queries from prompts.
- A memory module for agent memory integration.
- Cloud API connectors for database provisioning.
Tools can be plugged into agent architectures for enhanced memory and logic.
Many agent frameworks (e.g., Zep, Cognney, Memz) run on top of Neo4j and use it to provide graph-based memory.

Systems using GraphRAG show lower hallucination rates and provide more relevant, structured answers than vector-only RAGs or direct LLM outputs.
Vector similarity-based RAG improves over direct LLM, but relevance and actual understanding are still limited.
GraphRAG pattern involves:
- Initial vector search to map questions to relevant nodes.
- Retrieval of connected nodes (context) from the knowledge graph.
- Effective for combining structured and unstructured data in responses.

Typical GraphRAG pipeline uses both traversal (graph) and vector similarity.
Input questions undergo vector or graph search; graph analytics can further enhance results.
Context derived from graph is provided to LLM to improve answer quality.
Neo4j’s MCP server provides text-to-Cypher functionality, although generated queries from LLMs can sometimes be inaccurate.
Best pattern is performing a vector search, mapping results to graph nodes, and re-ranking for context relevance.

Example: CLA replaced SaaS systems with a GraphRAG solution, using enterprise wikis, HR systems, and documentation.
Processed 250,000 employee questions in the first year, serving 2,000 daily queries, with 85% employee adoption.

Neo4j Certified Developer Program is recommended for hands-on learning; it offers certification and community support.
Neo4j’s free annual Nodes Conference provides global, 24-hour content on graph technology and applications.

Standard pattern: LLM translates user queries to vector search; top results are mapped to graph nodes to provide context.
When importing unstructured data, Neo4j creates nodes and attaches embeddings as node properties.
Various plugins and integrations exist, including Neo4j Python library, LangChain, LlamaIndex, and Haystack, supporting flexible associations and workflows.
MCP's memory functionality specifics are developed by Michael Hunger's team (more details in a subsequent session).

Neo4j integrates with multiple agent and memory frameworks.
On LangChain vs. LangGraph: LangGraph is seen as an agent extension built by the LangChain team.
Recommended to use whichever framework best suits needs, as Neo4j supports broad integration.

Agentic GraphRAG: AI’s Logical Edge — Stephen Chin, Neo4j