Stop juggling five databases for one AI app.
Neumann stores your tables, graphs, and vectors in one place. Query across all three in a single statement.
-- Find engineers similar to Alice who report to Bob
FIND NODE person
WHERE role = 'engineer'
SIMILAR TO 'user:alice'
CONNECTED TO 'user:bob'One query. Relational filter + vector similarity + graph traversal.
Three engines, one system. Store a table, connect entities in a graph, and search by vector similarity -- without moving data between systems. No ETL, no sync, no glue code.
Semantic consensus. Concurrent writes to different fields auto-merge. The consensus layer classifies conflicts geometrically rather than treating all concurrent writes as errors.
Alice updates email, Bob updates photo (same user, same time).
Traditional DB: conflict, manual resolution.
Neumann: auto-merges (different fields = orthogonal changes).
AI-native by design. Built-in embedding storage, semantic caching for LLM responses, and encrypted vault for secrets. The query language understands similarity, not just equality.
RAG -- Store documents with embeddings and relationships. Semantic search follows graph links automatically.
Agent memory -- Conversation history with vector recall across sessions. Cache repeated LLM calls to cut API costs.
Knowledge graphs -- Combine structured data with semantic similarity. Find entities by what they mean, not just what they match.
Access control -- Graph-based permissions. Query results respect who's asking.
See Use Cases for worked examples with Python, TypeScript, and the CLI.
curl -sSfL https://raw.githubusercontent.com/Shadylukin/Neumann/main/install.sh | bash
neumann-- Relational
CREATE TABLE users (id INT, name TEXT, role TEXT);
INSERT users id=1, name='Alice', role='engineer';
SELECT * FROM users WHERE role = 'engineer';
-- Graph
NODE CREATE person {name: 'Alice'};
NODE CREATE person {name: 'Bob'};
EDGE CREATE 1 -> 2 : reports_to;
-- Vector
EMBED STORE 'user:alice' [0.1, 0.2, 0.3];
SIMILAR 'user:alice' TOP 5;More install methods (Homebrew, Cargo, Docker, source) in the Installation Guide. Full walkthrough in the Quick Start Tutorial.
Python
pip install neumann-dbfrom neumann import NeumannClient
client = NeumannClient.connect("localhost:9200")
result = client.execute("SELECT * FROM users")TypeScript
npm install @scrunchee/clientimport { NeumannClient } from '@scrunchee/client';
const client = await NeumannClient.connect("localhost:9200");
const result = await client.execute("SELECT * FROM users");Full tutorials: Python | TypeScript
Web dashboard with system status and query terminal
Interactive graph visualization with force-directed layout
| I want to... | Go to |
|---|---|
| Follow a tutorial | Quick Start, RAG in 5 Minutes, Knowledge Graph |
| Solve a specific problem | How-to Guides (30+ guides covering storage, graphs, vectors, security, deployment) |
| Look up syntax or config | Query Language, Configuration, Error Types |
| Understand the architecture | Design Overview, Consensus, HNSW |
Full table of contents | Rustdoc API Reference
Benchmarked on Apple M-series silicon:
- 3.8M reads/sec, 2.0M writes/sec (in-memory, no durability)
- 22K durable writes/sec with group commit, 260/sec with fsync
- 150us vector similarity (HNSW, 10K embeddings, 128-dim)
- 52M conflict checks/sec via sparse delta comparison
Neumann is pre-1.0 (v0.3.1). Core engines have 95%+ test coverage and 139 fuzz targets. Single-node is production-ready. Multi-node consensus has comprehensive testing (loom, proptest, deterministic simulation) but needs more real-world validation.
See the roadmap for where we're headed -- native embeddings, natural language queries, and AI-native analytics.
John von Neumann unified code and data in the stored-program architecture. Neumann unifies structure, relationships, and semantics.
Licensed under the Business Source License 1.1.
Built by Lukin Ackroyd in Auckland, New Zealand.
Neumann is the infrastructure layer for Scrunchee.