a3s-memory 0.1.3

A3S Memory - Pluggable memory storage for AI agents
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
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

use super::in_memory::{IndexSnapshot, PartitionBlock};
use super::{
    VectorIndexDescriptor, VectorIndexError, VectorResult, VectorSearchHit, VectorSearchRequest,
    VectorSearchResult,
};
use std::cmp::Ordering;
use std::collections::{BTreeMap, BinaryHeap};
use std::sync::Arc;

pub(super) fn search_snapshot(
    snapshot: Arc<IndexSnapshot>,
    descriptor: &VectorIndexDescriptor,
    query: Vec<f32>,
    request: VectorSearchRequest,
) -> VectorResult<VectorSearchResult> {
    let mut candidates = BinaryHeap::new();
    let mut searched_records = 0usize;
    let filter_labels = !request.labels.is_empty();

    for (partition_name, partition) in &snapshot.partitions {
        if !request.partitions.is_empty() && !request.partitions.contains(partition_name) {
            continue;
        }
        for (record_index, vector) in partition
            .vectors
            .chunks_exact(descriptor.dimension)
            .enumerate()
        {
            if filter_labels && !labels_match(&partition.labels[record_index], &request.labels) {
                continue;
            }
            searched_records = searched_records
                .checked_add(1)
                .ok_or(VectorIndexError::SizeOverflow)?;
            let score = dot_product_f64(&query, vector);
            if !score.is_finite() || score > f64::from(f32::MAX) || score < f64::from(f32::MIN) {
                return Err(VectorIndexError::ScoreOverflow {
                    partition: partition_name.clone(),
                    id: partition.ids[record_index].clone(),
                });
            }
            let candidate = ScoredRecord {
                partition,
                record_index,
                score,
            };
            if candidates.len() < request.limit {
                candidates.push(candidate);
            } else if candidates
                .peek()
                .is_some_and(|worst| compare_best_first(&candidate, worst).is_lt())
            {
                candidates.pop();
                candidates.push(candidate);
            }
        }
    }

    let mut ranked = candidates.into_vec();
    ranked.sort_by(compare_best_first);
    let hits = ranked
        .into_iter()
        .map(|candidate| VectorSearchHit {
            id: candidate.partition.ids[candidate.record_index].clone(),
            partition: candidate.partition.name.clone(),
            score: candidate.score as f32,
            labels: candidate.partition.labels[candidate.record_index].clone(),
        })
        .collect::<Vec<_>>();

    Ok(VectorSearchResult {
        truncated: searched_records > hits.len(),
        hits,
        status: snapshot.status(),
        searched_records,
    })
}

fn labels_match(
    record_labels: &BTreeMap<String, String>,
    required: &BTreeMap<String, String>,
) -> bool {
    required
        .iter()
        .all(|(key, value)| record_labels.get(key) == Some(value))
}

fn dot_product_f64(left: &[f32], right: &[f32]) -> f64 {
    left.iter().zip(right).fold(0.0f64, |score, (left, right)| {
        score + f64::from(*left) * f64::from(*right)
    })
}

struct ScoredRecord<'a> {
    partition: &'a PartitionBlock,
    record_index: usize,
    score: f64,
}

impl ScoredRecord<'_> {
    fn id(&self) -> &str {
        &self.partition.ids[self.record_index]
    }
}

impl PartialEq for ScoredRecord<'_> {
    fn eq(&self, other: &Self) -> bool {
        self.score.to_bits() == other.score.to_bits()
            && self.partition.name == other.partition.name
            && self.id() == other.id()
    }
}

impl Eq for ScoredRecord<'_> {}

impl PartialOrd for ScoredRecord<'_> {
    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
        Some(self.cmp(other))
    }
}

impl Ord for ScoredRecord<'_> {
    fn cmp(&self, other: &Self) -> Ordering {
        // BinaryHeap keeps its greatest value at the root. Reverse score order
        // so the worst retained candidate is removed when the heap exceeds k.
        other
            .score
            .total_cmp(&self.score)
            .then_with(|| self.partition.name.cmp(&other.partition.name))
            .then_with(|| self.id().cmp(other.id()))
    }
}

fn compare_best_first(left: &ScoredRecord<'_>, right: &ScoredRecord<'_>) -> Ordering {
    right
        .score
        .total_cmp(&left.score)
        .then_with(|| left.partition.name.cmp(&right.partition.name))
        .then_with(|| left.id().cmp(right.id()))
}