candding 0.1.0

Pure candle embeddings for Rust with no ONNX: dense, sparse, multi-vector and rerank models today, image models on the roadmap
Documentation 
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

use crate::models::SparseHead;
use crate::registry::DtypeSupport;
use crate::registry::Support;
use crate::registry::WeightsFormat;
use crate::registry::sparse::entry::SparseComputation;
use crate::registry::sparse::entry::SparseEntry;

/// `Qdrant/bm25`: no neural network, tokenizer or weight file, since the repository's `config.json` is `{}`; the entry
/// is `Computed`, with no family, `vocab_size` 0 and `NoWeights`.
pub(super) const BM25: SparseEntry = SparseEntry {
  id: "Qdrant/bm25",
  family: None,
  head: SparseHead::Bm25,
  vocab_size: 0,
  default_max_length: usize::MAX,
  cpu: Support::NotApplicable,
  metal: Support::NotApplicable,
  cuda: Support::NotApplicable,
  dtypes: DtypeSupport::not_applicable(),
  license: "Apache-2.0",
  size_gb: 0.0001,
  phase: 3,
  weights: WeightsFormat::NoWeights,
  gated: false,
  accel_only_dtypes: &[],
  head_source: None,
  head_revision: None,
  computation: SparseComputation::Computed,
  cell_exceptions: &[],
};