← back
Connecting the Dots with Context Graphs — Stephen Chin, Neo4j
Takeaway
Use explicit relationships and decision provenance to give agents context that similarity search alone can miss.
Summary
- Context graphs connect siloed enterprise sources with relationships, prior decisions, and tool activity to ground agent reasoning.
- A healthcare example contrasts generic LLM answers and vector retrieval with graph retrieval that incorporates patient history and linked treatments.
- The proposed memory model separates current execution state, long-term domain knowledge, and reasoning traces for provenance and debugging.
- Neo4j combines graph traversal, vector lookup, community algorithms, and an open-source agent memory package to organize and retrieve context.
context-graphsneo4jgraph-rag
Original description
Ask a vector RAG system about a patient's emphysema care plan and it returns generic advice: respiratory therapy, deep breathing. Give it a graph grounded in that patient's actual history and it knows they smoke, knows they've had an operation, and gives recommendations that reflect it. The information existed in both cases. What changed was whether the system could traverse the relationships connecting it. Stephen Chin from Neo4j makes the case that retrieval alone is not enough because agents also lose the reasoning behind past decisions. Context graphs capture not just what was retrieved but what decisions were made, why, which policies applied, and what the outcome was, so that precedent is queryable the next time a similar case comes up. The financial services demo shows this concretely: a loan decision that surfaces a prior rejection, related margin trades, and fraud risk patterns, with the graph traversal visible so the human making the final call can actually see what the system is drawing on. Speaker info: https://x.com/steveonjava / steveonjava