hermes-server 1.8.21

gRPC search server for Hermes
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

# Document index schema for hermes
# Fields: id, title, abstract, content, uris, sparse/dense vectors with spans

index documents {
    # Unique document identifier
    field id: text<raw> [indexed, stored]

    # Title with English stemming for full-text search
    field title: text<en_stem> [indexed, stored]

    # Abstract with English stemming for full-text search
    field abstract: text<en_stem> [indexed, stored]

    # Main content with English stemming for full-text search
    field content: text<en_stem> [indexed, stored]

    # List of URIs - multi-valued, lowercase tokenizer for exact matching
    field uris: text<default> [indexed, stored<multi>]

    # Multi sparse vectors for title (e.g., SPLADE embeddings)
    # uint8 quantization reduces memory ~4x vs float32, weight_threshold prunes low values
    field title_sparse_vectors: sparse_vector<u32> [indexed<quantization: uint8, weight_threshold: 0.01>, stored<multi>]

    # Span metadata for title sparse vectors - stored JSON list, not indexed
    field title_sparse_span: json [stored<multi>]

    # Multi dense vectors (e.g., sentence embeddings)
    # f16 storage quantization halves memory with <0.1% recall loss
    # IVF-RaBitQ uses binary quantization (~32x compression) for ANN search
    # Index starts as Flat (brute-force), call build_vector_index() to train IVF-RaBitQ
    field dense_vectors: dense_vector<1024, f16> [indexed<ivf_rabitq, num_clusters: 256, nprobe: 32>, stored<multi>]

    # Span metadata for dense vectors - stored JSON list, not indexed
    field dense_span: json [stored<multi>]
}