Context with boundaries
Retrieve the neighbourhood relevant to a task, retain the evidence path and exclude information outside the user’s permissions.
Technical depth · Graph engineering
Knowledge graphs at million-entity scale.
Large knowledge graphs succeed when identity, topology, retrieval and analytical workloads are engineered together. The entity count matters—but relationship density, update rate, query shape and memory often matter more.
Million-entity scale
Scale the workload—not just the node count.
“Millions of entities” does not by itself determine the architecture. Relationship count, property payload, update rate, query shape, concurrency, latency targets and analytical memory usually matter more.
01 · ModelIdentity & meaning
Stable identifiers, entity types, relationship semantics, time and provenance.
02 · IngestResolve & connect
Batch or streaming pipelines, entity resolution, deduplication and validation.
03 · RetrieveIndex & constrain
Uniqueness constraints plus range, text, full-text and vector indexes where justified.
04 · AnalyseProject the graph
Load the task-relevant labels, relationships and properties into an analytical projection.
05 · OperateMeasure & evolve
Profile queries, estimate memory, monitor quality, update incrementally and partition only when needed.
Operational graph
- Start from access patternsModel the relationships the system must traverse, not every relationship that could exist.
- Protect identity and qualityUse keys, constraints, provenance and deterministic merge rules to prevent silent duplication.
- Index entry pointsIndexes help locate starting entities; traversal handles the connected neighbourhood.
- Separate heavy workloadsUse replicas, analytical services or workload isolation when operational and analytical demands compete.
- Partition deliberatelyFederation or sharding can help when data ownership, geography, tenancy or scale provides a clean boundary.
Analytical graph
- Project only what the question needsSelect labels, relationship types, directions, weights, properties and time windows.
- Estimate before executionMemory estimation and graph dimensions should inform algorithm choice and concurrency.
- Choose the right execution modeStream exploratory results; mutate intermediate features; write back only durable, governed outputs.
- Use incremental recomputationRefresh affected communities, embeddings or features when possible instead of rebuilding everything.
- Test representative topologyA million sparse entities and a million densely connected entities behave very differently.
Benefits
Why the graph becomes more valuable as the system grows.
The value comes from reusing connected structure across applications, not from drawing more nodes on a screen.
Explainable relationships
Show how an answer, risk signal or recommendation connects to entities, rules, events and source evidence.
Reusable intelligence
The same identity, ontology, provenance and relationship layer can support agents, causal work, search, analytics and ML features.
Graph algorithms
Choose the family that matches the decision.
Algorithm names are not a strategy. Define the business question, graph projection, direction, weights, time window and validation method before selecting one.
Who or what matters most?
Rank influential, connected or bridging entities.
- Degree centrality
- PageRank and article rank
- Betweenness centrality
- Eigenvector and harmonic centrality
Which entities form groups?
Find dense clusters, components and hierarchical communities.
- Louvain and Leiden
- Label propagation
- Weakly and strongly connected components
- Modularity-based clustering
How are entities connected?
Find efficient routes, reachable entities and dependency paths.
- Dijkstra and A*
- Yen’s k-shortest paths
- Breadth-first search
- Minimum spanning tree and Steiner tree
What resembles—or may connect to—what?
Measure neighbourhood similarity and estimate missing or future links.
- Node similarity and k-nearest neighbours
- Common neighbours and Adamic–Adar
- Link prediction pipelines
- Topological similarity functions
How can topology become features?
Represent nodes as vectors for search, clustering and downstream ML.
- FastRP
- Node2Vec
- GraphSAGE
- HashGNN and feature-aware embeddings
What can the graph help predict?
Train and evaluate models using topology, properties and learned embeddings.
- Node classification and regression
- Link prediction
- Graph features for external models
- DAG algorithms and topological ordering
Algorithm availability and exact names vary by graph platform and version. The families above describe the analytical choices; production use still requires data-quality checks, representative evaluation and domain review.
Design principle
Scale the questions—not only the database.
A production graph should serve bounded operational traversals and task-specific analytical projections. Not every query or algorithm needs every entity, relationship or property.
See how graphs support agents, causality and context →Primary references
Further technical reading.
- Neo4j Graph Data Science: algorithm families ↗Current categories for centrality, community detection, similarity, paths, embeddings, DAG algorithms and link prediction.
- GDS memory estimation and system requirements ↗Planning graph projections and algorithm memory.
- Neo4j indexes and constraints ↗Operational entry points, search and data integrity.
- Neo4j scaling strategies ↗Read scaling, federation, sharding and distributed options.
Designing connected intelligence at enterprise scale?
Start with the decisions, graph topology, evidence and workload—not the technology label.