1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
// SPDX-License-Identifier: Apache-2.0
//! Vector storage precision tag.
//!
//! Selects the on-disk + in-memory dtype for vector storage on a per-collection
//! basis. Independent of quantization (see [`crate::vector_ann::VectorQuantization`]):
//! a collection can be `(F32, None)`, `(BF16, None)`, `(F32, RaBitQ)`,
//! `(BF16, RaBitQ)`, etc. Storage dtype controls the durable form; quantization
//! is an optional search-time overlay on top.
/// Vector storage dtype for HNSW + flat indexes.
///
/// `F32` is the default and the historical NodeDB storage form. `F16` and `BF16`
/// give 2x memory + disk savings with negligible recall loss for typical
/// embedding workloads, at the cost of a slightly more expensive distance kernel
/// (F16/BF16 must up-convert to F32 for arithmetic on hardware without native
/// half-precision FMA, e.g., pre-AVX-512-FP16 x86).
///
/// `FP8` (E4M3 / E5M2) is deliberately omitted from this release; it is rare in
/// vector-search workloads relative to the conversion-surface cost of supporting
/// it, and the recall hit on typical embeddings (1.5-bit-ish effective mantissa
/// precision) is severe. Reconsider when there is concrete user demand.
#[repr(u8)]
#[derive(
Debug,
Clone,
Copy,
Default,
PartialEq,
Eq,
Hash,
serde::Serialize,
serde::Deserialize,
zerompk::ToMessagePack,
zerompk::FromMessagePack,
)]
#[msgpack(c_enum)]
#[non_exhaustive]
pub enum VectorStorageDtype {
/// 32-bit IEEE 754 single precision. Default; 4 bytes per dim.
#[default]
F32 = 0,
/// 16-bit IEEE 754 half precision. 2 bytes per dim. ~3 decimal digits of
/// precision; ~6e-5 to 65504 range.
F16 = 1,
/// 16-bit Brain Float (Google bfloat16). 2 bytes per dim. Same exponent
/// range as F32 (~1e-38 to 3.4e38) but only ~7-bit mantissa. Better
/// dynamic range than F16; preferred for embedding workloads.
BF16 = 2,
}
impl VectorStorageDtype {
/// Bytes occupied per vector dimension at this dtype.
pub const fn bytes_per_dim(self) -> usize {
match self {
Self::F32 => 4,
Self::F16 => 2,
Self::BF16 => 2,
}
}
/// Total bytes needed to store `dim`-dimensional vector in this dtype.
pub const fn bytes_for_dim(self, dim: usize) -> usize {
dim * self.bytes_per_dim()
}
/// Stable lowercase string identifier — used in DDL parsing
/// (`WITH (storage_dtype='bf16')`) and in error messages.
pub const fn as_str(self) -> &'static str {
match self {
Self::F32 => "f32",
Self::F16 => "f16",
Self::BF16 => "bf16",
}
}
/// Parse from the lowercase identifier. Returns `None` for unknown values;
/// the caller wraps that in a typed error (e.g., `NodeDbError::bad_request`)
/// with a precise message naming the offending value.
pub fn parse(s: &str) -> Option<Self> {
match s {
"f32" => Some(Self::F32),
"f16" => Some(Self::F16),
"bf16" => Some(Self::BF16),
_ => None,
}
}
}
impl core::str::FromStr for VectorStorageDtype {
type Err = ();
fn from_str(s: &str) -> Result<Self, Self::Err> {
Self::parse(s).ok_or(())
}
}
impl core::fmt::Display for VectorStorageDtype {
fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
f.write_str(self.as_str())
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn default_is_f32() {
assert_eq!(VectorStorageDtype::default(), VectorStorageDtype::F32);
}
#[test]
fn bytes_per_dim_matches_iec_widths() {
assert_eq!(VectorStorageDtype::F32.bytes_per_dim(), 4);
assert_eq!(VectorStorageDtype::F16.bytes_per_dim(), 2);
assert_eq!(VectorStorageDtype::BF16.bytes_per_dim(), 2);
}
#[test]
fn bytes_for_dim_is_dim_times_width() {
assert_eq!(VectorStorageDtype::F32.bytes_for_dim(128), 512);
assert_eq!(VectorStorageDtype::BF16.bytes_for_dim(1536), 3072);
assert_eq!(VectorStorageDtype::F16.bytes_for_dim(256), 512);
}
#[test]
fn as_str_roundtrips_from_str() {
for v in [
VectorStorageDtype::F32,
VectorStorageDtype::F16,
VectorStorageDtype::BF16,
] {
assert_eq!(VectorStorageDtype::parse(v.as_str()), Some(v));
}
}
#[test]
fn from_str_unknown_returns_none() {
assert_eq!(VectorStorageDtype::parse("fp8"), None);
assert_eq!(VectorStorageDtype::parse("F32"), None);
assert_eq!(VectorStorageDtype::parse(""), None);
}
#[test]
fn display_matches_as_str() {
for v in [
VectorStorageDtype::F32,
VectorStorageDtype::F16,
VectorStorageDtype::BF16,
] {
assert_eq!(format!("{}", v), v.as_str());
}
}
#[test]
fn msgpack_roundtrip() {
for v in [
VectorStorageDtype::F32,
VectorStorageDtype::F16,
VectorStorageDtype::BF16,
] {
let bytes = zerompk::to_msgpack_vec(&v).unwrap();
let restored: VectorStorageDtype = zerompk::from_msgpack(&bytes).unwrap();
assert_eq!(restored, v);
}
}
}