Neo4j vs NornicDB Feature Audit

Historical compatibility audit

Last Updated: December 2025 (historical snapshot)

This audit is a December 2025 snapshot of the v0.1.x surface. The current function, procedure, and plugin capability boundaries are maintained in APOC Functions, the APOC procedure capability audit, and SHOW FUNCTIONS / SHOW PROCEDURES; remaining gaps are tracked in #1318.

Scope: Production workloads excluding plugins and multi-database orchestration.

---

Executive Summary

The December 2025 audit recorded the following estimates. These percentages are historical measurements, not current compatibility guarantees:

Category	Status	Notes
Core Data Model	100%	Nodes, relationships, properties, arrays, maps
Cypher Language	100%	Implemented clauses, pattern matching, and subqueries; consult the compatibility guide for limits
Functions	109%	Dec 2025 measurement: 147 functions vs Neo4j's 135. Current built-in surface: SHOW FUNCTIONS; APOC surface: APOC Functions
Indexes	100%	B-tree, full-text, vector, composite, range
Constraints	100%	UNIQUE, EXISTS, NODE KEY, RELATIONSHIP KEY, property type (nodes + relationships) + temporal no-overlap, domain/enum, cardinality, endpoint policy extensions
Transactions	100%	Transactions with BEGIN/COMMIT/ROLLBACK; durability depends on configured sync policy
Built-in Procedures	100%	Dec 2025 measurement: 41 procedures (34 db._ + 7 dbms._). Current surface: SHOW PROCEDURES
APOC	Partial	Dec 2025 measurement: 960+. Current: APOC Functions and the procedure capability audit; remaining gaps tracked in #1318
Protocol/Drivers	95%	Bolt v4.x, all major drivers

New in 0.1.4:

• String query auto-embedding in db.index.vector.queryNodes
• Multi-line SET with arrays support
• Server-side query embedding

---

Feature Parity Scorecard

Category	Weight	Score	Weighted
Core Data Model	20%	100%	20.0
Cypher Language	20%	100%	20.0
Functions	10%	109%	10.9
Indexes	10%	100%	10.0
Constraints	10%	100%	10.0
Transactions	15%	100%	15.0
Procedures	10%	60%	6.0
Protocol/Drivers	5%	95%	4.75
TOTAL	100%	96.65%	96.65%

---

Completed Features

1. Core Data Model - 100%

All 12 features fully implemented: node/relationship creation, multiple labels, all property types (string, int, float, bool, arrays, maps), ID persistence, directed relationships, self-relationships, parallel edges.

2. Cypher Query Language - 100%

Core Clauses (16/16): MATCH, OPTIONAL MATCH, WHERE, RETURN, WITH, CREATE, MERGE, DELETE, DETACH DELETE, SET, REMOVE, ORDER BY, LIMIT, SKIP, UNWIND, UNION/UNION ALL

Pattern Matching (8/8): Fixed-length paths, variable-length paths (bounded/unbounded), shortestPath, allShortestPaths, relationship type filtering, bidirectional matching, named paths

Advanced Features (7/7): List comprehension, pattern comprehension, CASE expressions, map projection, EXISTS subqueries, COUNT subqueries with comparisons

3. Functions - 109% (Dec 2025: 147 vs 135)

Category	Count	Status
String	23	100%
List	17	100%
Mathematical	24	126% (exceeds Neo4j's 19)
Trigonometric	11	100%
Aggregation	12	133% (exceeds Neo4j's 9)
Temporal	25	100%
Spatial	19	127%
Type Conversion	12	100%
Node/Relationship	12	100%
Vector/Similarity	3	100%

4. Indexes - 100%

All index types supported: B-tree, full-text (Lucene-compatible), vector (HNSW), composite (multi-property), token lookup, range indexes. Full CRUD operations including CREATE INDEX, DROP INDEX, SHOW INDEXES, index hints.

5. Constraints & Schema - 100%

Full Neo4j constraint parity on both nodes and relationships:

Node constraints: UNIQUE (single and composite), EXISTS (IS NOT NULL), NODE KEY (composite uniqueness + existence), property type (IS :: TYPE).

Relationship constraints: UNIQUE (single and composite), EXISTS (IS NOT NULL), RELATIONSHIP KEY (composite uniqueness + existence), property type (IS :: TYPE). Uniqueness and key constraints on relationships automatically create owned backing indexes.

NornicDB extensions: Temporal no-overlap (IS TEMPORAL NO OVERLAP) prevents overlapping time intervals on nodes or relationships grouped by a composite key. Domain/enum (IN ['val1', 'val2']) restricts a property to a fixed set of allowed values on nodes or relationships. Cardinality (REQUIRE MAX COUNT N) limits outgoing or incoming edge count per node for a given relationship type, with direction encoded in the FOR clause arrows (FOR ()-[r:TYPE]->() for outgoing, FOR ()<-[r:TYPE]-() for incoming). Endpoint policies (REQUIRE ALLOWED / REQUIRE DISALLOWED) restrict which label pairs may be connected by a relationship type; ALLOWED policies form a union whitelist and DISALLOWED policies are a blacklist with precedence.

All constraint DDL supports IF NOT EXISTS for idempotent creation. Constraints are validated against existing data at creation time and enforced on all write paths (CREATE, MERGE, SET, REMOVE) including transactions and bulk operations. Label mutations (SET n:Label, REMOVE n:Label) are also validated against endpoint policy constraints. SHOW CONSTRAINTS reports entity type (NODE or RELATIONSHIP), constraint type, owned backing indexes, and additional columns for cardinality (direction, maxCount) and policy (sourceLabel, targetLabel, policyMode) constraints. Block-style constraint contracts are exposed separately via SHOW CONSTRAINT CONTRACTS.

6. Transactions - 100%

Transaction behavior through native storage; see durability boundaries:

• Atomicity: All operations commit together or none
• Consistency: Constraint validation before commit
• Isolation: Snapshot isolation via MVCC at the storage transaction layer, including read-your-writes and write-write conflict detection at commit
• Durability: WAL-based recovery with configurable sync/acknowledgement boundaries

Supports: BEGIN/COMMIT/ROLLBACK, implicit transactions, automatic rollback on error, transaction metadata.

7. Protocol & Drivers - 95%

Driver	Status
Python (neo4j-driver)	Full
JavaScript/TypeScript	Full
Go (neo4j-go-driver)	Full
Java (neo4j-java-driver)	Full
.NET, Ruby	Untested (should work)

Bolt v4.x fully supported. v5.x backward compatible.

---

Built-in Procedures (18+ Implemented)

db.* Procedures

db.labels, db.propertyKeys, db.relationshipTypes, db.info, db.ping
db.index.vector.queryNodes, db.index.vector.createNodeIndex
db.index.vector.queryRelationships
db.index.fulltext.queryNodes, db.index.fulltext.queryRelationships
db.index.fulltext.listAvailableAnalyzers
db.awaitIndex, db.awaitIndexes, db.resampleIndex
db.stats.clear/collect/retrieve/status/stop
db.clearQueryCaches
db.create.setNodeVectorProperty, db.create.setRelationshipVectorProperty

dbms.* Procedures

dbms.info, dbms.listConfig, dbms.clientConfig
tx.setMetaData

NEW: Vector Search Enhancements

-- String query (auto-embedded by NornicDB)
CALL db.index.vector.queryNodes('idx', 10, 'machine learning tutorial')
YIELD node, score

-- Direct vector (Neo4j compatible)
CALL db.index.vector.queryNodes('idx', 10, [0.1, 0.2, 0.3])
YIELD node, score

-- Multi-line SET with arrays (optional user metadata)
MATCH (n:Node {id: 'abc'})
SET n.embedding = [0.7, 0.2, 0.05, 0.05],
    n.embedding_model = 'mxbai-embed-large',
    n.has_embedding = true

Note: NornicDB-managed embedding metadata is stored internally (not in Properties) to avoid property namespace pollution.

---

APOC Procedures (52 Implemented)

Core Utilities

Category	Functions
Path/Graph	apoc.path.subgraphNodes, apoc.path.expand, apoc.path.spanningTree
Map	merge, setKey, removeKey, removeKeys, fromPairs, fromLists, clean
Collection	flatten, toSet, sum, avg, min, max, sort, reverse, union, unionAll, intersection, subtract, contains, containsAll
Text	apoc.text.join
Conversion	toJson, fromJsonMap, fromJsonList
Meta	apoc.meta.type, apoc.meta.isType
UUID	apoc.create.uuid

Dynamic Cypher

CALL apoc.cypher.run('MATCH (n) RETURN count(n)', {})
CALL apoc.cypher.runMany('CREATE (n); MATCH (n) RETURN n', {})

Batch Processing

CALL apoc.periodic.iterate(
  'MATCH (n:Node) RETURN n',
  'SET n.processed = true',
  {batchSize: 1000}
)

Graph Algorithms

Algorithm	Procedure	Status
Dijkstra	apoc.algo.dijkstra	APOC name
A*	apoc.algo.aStar	APOC name
All Simple Paths	apoc.algo.allSimplePaths	APOC name
PageRank	apoc.algo.pageRank	NornicDB extension (GDS: gds.pageRank.stream)
Betweenness	apoc.algo.betweenness	NornicDB extension (GDS: gds.betweenness.stream)
Closeness	apoc.algo.closeness	NornicDB extension (GDS: gds.closeness.stream)
Neighbors	apoc.neighbors.tohop, apoc.neighbors.byhop	APOC names

Community Detection

Algorithm	Procedure	Status
Louvain	apoc.algo.louvain	NornicDB extension (GDS: gds.louvain.stream)
Label Propagation	apoc.algo.labelPropagation	NornicDB extension (GDS: gds.labelPropagation.stream)
Connected Components	apoc.algo.wcc	NornicDB extension (GDS: gds.wcc.stream)

A NornicDB extension isn't in APOC or GDS: a query that uses it runs only on NornicDB (see NornicDB extensions).

Graph Data Science (GDS)

These GDS names work as in GDS 2.13 (Neo4j 5.26), with its parameters, columns, values and error messages:

Area	Names
Graph catalog	gds.graph.project (label / type projections, with orientation, properties, aggregation and the configuration map), gds.graph.list, gds.graph.drop, gds.graph.exists (procedure and function); one catalog per database
FastRP	gds.fastRP.stream, gds.fastRP.stats
Link prediction	gds.alpha.linkprediction.adamicAdar, commonNeighbors, preferentialAttachment, resourceAllocation, totalNeighbors, sameCommunity
Utilities	gds.util.asNode, asNodes, nodeProperty, NaN, infinity, isFinite, isInfinite, gds.version

NornicDB's differences: nodeId is the node's id (a STRING, #1165); FastRP's randomSeed defaults to 42, so the same call gives the same embeddings, and embeddingDimension is at most 8,192; memoryUsage and sizeInBytes are reported as unmeasured ("", -1); the catalog isn't per user; concurrency and readConcurrency are accepted and change nothing (NornicDB runs these single-threaded).

Not implemented (a query using them fails with an unknown procedure or function):

• every other GDS algorithm, in any mode: centrality (gds.pageRank, gds.betweenness, gds.closeness, gds.degree, gds.articleRank, …), community detection (gds.louvain, gds.leiden, gds.labelPropagation, gds.wcc, gds.scc, gds.triangleCount, gds.localClusteringCoefficient, gds.kmeans, …), similarity (gds.nodeSimilarity, gds.knn, …), path finding (gds.shortestPath.*, gds.allShortestPaths.*, gds.bfs, gds.dfs, gds.spanningTree, …), embeddings other than FastRP (gds.node2vec, gds.beta.graphSage, gds.hashgnn);
• the mutate, write and estimate modes, including FastRP's;
• Cypher projections (gds.graph.project as an aggregation, gds.graph.project.cypher) and other catalog operations (gds.graph.filter, gds.graph.nodeProperties.*, gds.graph.relationships.*, gds.graph.sample.*, graph export);
• ML pipelines and models (gds.beta.pipeline.*, gds.model.*).

Some of these algorithms exist as NornicDB extensions under apoc.algo.* (tables above); NornicDB's own gds.linkPrediction.*.stream procedures are extensions too.

Data Import/Export

Operation	Procedure	Status
Load JSON	apoc.load.json, apoc.load.jsonArray	Yes
Load CSV	apoc.load.csv	Yes
Export JSON	apoc.export.json.all, apoc.export.json.query	Yes
Export CSV	apoc.export.csv.all, apoc.export.csv.query	Yes
Import JSON	apoc.import.json	Yes

---

NornicDB-Exclusive Features

Features NornicDB has that Neo4j doesn't:

Feature	Description
Automatic Vector Index	All node embeddings indexed automatically, no setup required (with managed embeddings enabled)
String Query Embedding	db.index.vector.queryNodes accepts strings, auto-embeds server-side
Hybrid Search REST API	/nornicdb/search with RRF fusion of vector + BM25
Knowledge-Layer Scoring	Declarative decay profiles and retention bindings via Cypher DDL
Auto-Relationships	Automatic edge creation via embedding similarity
GPU Acceleration	Metal/CUDA/OpenCL/Vulkan for vector ops
Embedded Mode	Use as library without server
Link Prediction	ML-based relationship prediction (TLP algorithms)
MCP Server	Native Model Context Protocol for LLM tools
Temporal No-Overlap	Prevents overlapping time intervals on nodes or relationships
Domain/Enum Constraints	Restricts property values to a declared set of allowed values
Cardinality Constraints	Limits outgoing or incoming edge count per node for a given relationship type
Endpoint Policies	Restricts which (source, target) label pairs may be connected by a relationship type

Measured Performance and Resource Comparison

The current same-hardware Northwind benchmark records the following results for its 48,000-product / 48,000-order workload:

Metric	Neo4j	NornicDB	Difference
Overall mean query latency	98.64 ms	0.23 ms	-99.8% (432.38x ratio)
Overall throughput	7.76 ops/s	17.70 ops/s	+128.1% (2.28x)
Peak system memory used	33.4 GiB	32.5 GiB	-2.7%
Total benchmark data directory	204 MiB	134 MiB	-34.3%

See the full Northwind comparison for workload, hardware, correctness fingerprints, power, and storage details. Treat these as workload-specific measurements rather than universal deployment requirements.

---

Recently Completed

Feature	Implementation
Bookmarks (causal consistency)	Returns nornicdb:bookmark:* on commit, accepts in BEGIN
String query auto-embedding	db.index.vector.queryNodes accepts text strings
Multi-line SET with arrays	Full support for embedding storage workflow
db.index.fulltext.createNodeIndex	Create fulltext indexes on node labels
db.index.fulltext.createRelationshipIndex	Create fulltext indexes on relationship types
db.index.vector.createRelationshipIndex	Create vector indexes on relationships
db.index.fulltext.drop	Drop fulltext indexes
db.index.vector.drop	Drop vector indexes
Prometheus /metrics endpoint	Full metrics export (requests, nodes, edges, embeddings, slow queries)
Slow query logging	Configurable threshold (default 100ms), file or stderr output

Not Applicable

Feature	Reason
Cluster management (dbms.*)	Single-node design
Enterprise security	Use external auth
Multi-database	Use separate instances

---

Use Case Compatibility

Recommended (95-100% Compatible)

• LLM/AI Agent Memory (primary design target)
• Knowledge Graphs
• Semantic Search (GPU-accelerated, MMR diversification)
• Graph Analysis (shortestPath, traversals, subgraphs)
• Recommendation Engines
• Financial/Transactional (full ACID)
• Multi-tenant Systems (constraint enforcement)
• Development/Testing (fast, lightweight)
• Enterprise Monitoring (Prometheus /metrics endpoint)

---

Roadmap

Recently Completed (v0.1.4)

• String query auto-embedding in vector search
• Multi-line SET with arrays
• Server-side query embedding via Cypher executor
• Prometheus /metrics endpoint
• Slow query logging (configurable threshold)
• MMR diversification for search results
• Eval harness for search quality validation
• Cross-encoder reranking for Stage 2 retrieval (local GGUF or external API: Cohere, TEI, Ollama)

Next Priority

Task	Effort	Status
Prometheus metrics	2 days	Done
Slow query logging	1 day	Done
MMR diversification	1 day	Done
Cross-encoder rerank	3 days	Done
Plugin system for APOC	3 days	Done
Eval harness	2 days	Done

---

Conclusion

NornicDB v0.1.4 provides production-grade Neo4j replacement capability:

• 96% overall feature parity
• 100% core data model compatibility
• 109% function parity at the December 2025 audit (147 vs Neo4j's 135); see APOC Functions for the current catalog
• Full ACID transactions
• All constraint types enforced
• Neo4j driver compatibility (Bolt v4.x)
• LLM-native features (memory decay, vector search, auto-embedding)

For LLM/AI workloads: Recommended (99% effective parity) For general Neo4j replacement: Suitable (96% feature parity)

---

Last Updated: December 1, 2025Previous Audit: November 27, 2025