Data Types

Neumann has two layers of types: scalar values for individual fields and tensor values for composite storage.

Scalar Types

ScalarValue represents a single value in the system.

Type	Description	Examples
`Null`	Absence of a value	`NULL`
`Bool`	Boolean	`TRUE`, `FALSE`
`Int`	64-bit signed integer	`42`, `-1`, `0`
`Float`	64-bit floating point	`3.14`, `-0.5`, `1e10`
`String`	UTF-8 text	`'hello'`, `'Alice'`
`Bytes`	Raw binary data	(used internally for blob content)

Literals

Strings: Single-quoted: 'hello world'
Integers: Unquoted numbers: 42, -7
Floats: Numbers with decimal or exponent: 3.14, 1e-5
Booleans: TRUE or FALSE (case-insensitive)
Null: NULL
Arrays: Square brackets: [1, 2, 3] or [0.1, 0.2, 0.3]

Tensor Types

TensorValue wraps scalar values with vector and pointer types for the unified data model.

Type	Description	Use Case
`Scalar(ScalarValue)`	Single scalar value	Table columns, node properties
`Vector(Vec<f32>)`	Dense float vector	Embeddings for similarity search
`Sparse(SparseVector)`	Sparse vector	Memory-efficient high-dimensional embeddings
`Pointer(String)`	Reference to another entity	Graph edges, foreign keys
`Pointers(Vec<String>)`	Multiple references	Multi-valued relationships

Column Types in CREATE TABLE

When creating relational tables, use SQL-style type names. These map to internal scalar types.

SQL Type	Internal Type	Notes
`INT`, `INTEGER`	Int	64-bit signed integer
`BIGINT`	Int	Same as INT (64-bit)
`SMALLINT`	Int	Same as INT (64-bit)
`FLOAT`	Float	64-bit floating point
`DOUBLE`	Float	Same as FLOAT
`REAL`	Float	Same as FLOAT
`DECIMAL(p,s)`	Float	Precision and scale are advisory
`NUMERIC(p,s)`	Float	Same as DECIMAL
`VARCHAR(n)`	String	Max length is advisory
`CHAR(n)`	String	Fixed-width (padded)
`TEXT`	String	Unlimited length
`BOOLEAN`	Bool	TRUE or FALSE
`DATE`	String	Stored as ISO-8601 string
`TIME`	String	Stored as ISO-8601 string
`TIMESTAMP`	String	Stored as ISO-8601 string
`BLOB`	Bytes	Raw binary data
custom name	String	Any unrecognized type stores as String

Type Coercion

Neumann performs implicit type coercion in comparisons:

Int and Float in arithmetic: Int is promoted to Float
String comparisons: lexicographic ordering
Null propagation: any operation with NULL yields NULL
Boolean context: only Bool values are truthy/falsy (no implicit conversion from Int)

Vector Representation

Dense vectors are stored as Vec<f32> and used for similarity search via HNSW indexes. All vectors in a collection must have the same dimensionality.

EMBED STORE 'doc1' [0.1, 0.2, 0.3, 0.4]

Sparse vectors use a compact representation storing only non-zero indices and values, making them efficient for high-dimensional data (e.g., 30,000+ dimensions for bag-of-words models).

Identifiers

Identifiers (table names, column names, labels) follow these rules:

Start with a letter or underscore
Contain letters, digits, and underscores
Case-insensitive for keywords, case-preserving for identifiers
No quoting required for simple names
Use single quotes for string values: 'value'

Keyboard shortcuts

Neumann