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
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
// Copyright 2024-2026 WritersLogic Contributors
// SPDX-License-Identifier: AGPL-3.0-or-later
use fxhash::{FxHashMap, FxHashSet};
use super::HmsCore;
use crate::core::entangled::EntangledHVec;
use crate::core::types::ConceptCandidate;
impl HmsCore {
/// Cluster similar vectors and return concept candidates with centroids and coherence scores.
///
/// Uses posting-list candidate generation: only vector pairs sharing at least
/// one active index are compared, avoiding O(n²) all-pairs at high sparsity.
pub fn synthesize_concepts(&self) -> Vec<ConceptCandidate> {
let cfg = &self.config.concepts;
let mut all_ids = Vec::new();
let mut all_vectors = Vec::new();
let shards = self.shards.read();
shards.for_each_shard(|shard| {
let (ids, vectors) = shard.load_all_vectors();
all_ids.extend(ids);
all_vectors.extend(vectors);
});
if all_ids.len() < cfg.min_cluster_size {
return vec![];
}
let n = all_vectors.len();
let mut used = vec![false; n];
let mut concepts = Vec::new();
// For large collections, use posting-list candidate generation
// to avoid O(n²). For small collections, brute-force is fine.
let use_postings = n > 500;
let neighbors: Vec<FxHashSet<usize>> = if use_postings {
let mut postings: FxHashMap<u32, Vec<usize>> = FxHashMap::default();
for (i, vec) in all_vectors.iter().enumerate() {
for &idx in vec.indices() {
postings.entry(idx).or_default().push(i);
}
}
let mut nbrs = vec![FxHashSet::default(); n];
for list in postings.values() {
for (pos, &a) in list.iter().enumerate() {
for &b in &list[pos + 1..] {
nbrs[a].insert(b);
nbrs[b].insert(a);
}
}
}
nbrs
} else {
Vec::new()
};
for i in 0..n {
if used[i] {
continue;
}
let mut cluster = vec![i];
if use_postings {
for &j in &neighbors[i] {
if !used[j]
&& all_vectors[i].similarity(&all_vectors[j]) > cfg.similarity_threshold
{
cluster.push(j);
}
}
} else {
for j in (i + 1)..n {
if !used[j]
&& all_vectors[i].similarity(&all_vectors[j]) > cfg.similarity_threshold
{
cluster.push(j);
}
}
}
if cluster.len() >= cfg.min_cluster_size {
for &idx in &cluster {
used[idx] = true;
}
let cluster_vecs: Vec<&EntangledHVec> =
cluster.iter().map(|&idx| &all_vectors[idx]).collect();
let centroid = self.bundle(&cluster_vecs);
let coherence: f64 = cluster_vecs
.iter()
.map(|v| v.similarity(¢roid))
.sum::<f64>()
/ cluster_vecs.len() as f64;
let member_ids: Vec<String> =
cluster.iter().map(|&idx| all_ids[idx].clone()).collect();
concepts.push(ConceptCandidate {
centroid_id: member_ids.first().cloned().unwrap_or_default(),
member_count: cluster.len() as u32,
coherence,
member_ids,
stable: false,
});
}
}
// Centroid refinement: k-means-style reassignment (up to 3 iterations).
// Recompute centroids and reassign vectors to closest centroid.
if concepts.len() >= 2 {
let cluster_indices: Vec<usize> = (0..n).filter(|&i| used[i]).collect();
// Build initial assignment: vector index -> concept index
let mut assignment: FxHashMap<usize, usize> = FxHashMap::default();
for (ci, concept) in concepts.iter().enumerate() {
for mid in &concept.member_ids {
if let Some(pos) = all_ids.iter().position(|id| id == mid) {
assignment.insert(pos, ci);
}
}
}
let mut centroids: Vec<EntangledHVec> = concepts
.iter()
.enumerate()
.map(|(ci, _)| {
let members: Vec<&EntangledHVec> = cluster_indices
.iter()
.filter(|&&vi| assignment.get(&vi) == Some(&ci))
.map(|&vi| &all_vectors[vi])
.collect();
self.bundle(&members)
})
.collect();
let mut stable = true;
for _iter in 0..3 {
let mut changed = false;
for &vi in &cluster_indices {
let mut best_ci = assignment[&vi];
let mut best_sim = all_vectors[vi].similarity(¢roids[best_ci]);
for (ci, centroid) in centroids.iter().enumerate() {
let sim = all_vectors[vi].similarity(centroid);
if sim > best_sim {
best_sim = sim;
best_ci = ci;
}
}
if best_ci != assignment[&vi] {
assignment.insert(vi, best_ci);
changed = true;
}
}
if !changed {
stable = true;
break;
}
stable = false;
// Recompute centroids after reassignment
for (ci, centroid) in centroids.iter_mut().enumerate() {
let members: Vec<&EntangledHVec> = cluster_indices
.iter()
.filter(|&&vi| assignment.get(&vi) == Some(&ci))
.map(|&vi| &all_vectors[vi])
.collect();
if !members.is_empty() {
*centroid = self.bundle(&members);
}
}
}
// Rebuild concepts from refined assignments
for (ci, concept) in concepts.iter_mut().enumerate() {
let members: Vec<usize> = cluster_indices
.iter()
.filter(|&&vi| assignment.get(&vi) == Some(&ci))
.copied()
.collect();
if members.is_empty() {
concept.member_count = 0;
concept.member_ids.clear();
concept.coherence = 0.0;
concept.stable = stable;
continue;
}
let member_vecs: Vec<&EntangledHVec> =
members.iter().map(|&vi| &all_vectors[vi]).collect();
let coherence: f64 = member_vecs
.iter()
.map(|v| v.similarity(¢roids[ci]))
.sum::<f64>()
/ member_vecs.len() as f64;
concept.member_ids = members.iter().map(|&vi| all_ids[vi].clone()).collect();
concept.centroid_id = concept.member_ids.first().cloned().unwrap_or_default();
concept.member_count = members.len() as u32;
concept.coherence = coherence;
concept.stable = stable;
}
// Remove empty concepts
concepts.retain(|c| c.member_count > 0);
} else {
// Single or zero concepts: trivially stable
for concept in &mut concepts {
concept.stable = true;
}
}
concepts
}
}