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
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
// SPDX-License-Identifier: BUSL-1.1
//! Hybrid (vector + text) search handler for the Data Plane CoreLoop.
use tracing::debug;
use nodedb_fts::FtsSearchParams;
use nodedb_fts::posting::QueryMode;
use crate::bridge::envelope::{ErrorCode, Response};
use crate::data::executor::core_loop::CoreLoop;
use crate::data::executor::task::ExecutionTask;
use crate::types::TenantId;
/// Default hybrid search weight: 0.5 = equal vector + text.
const DEFAULT_VECTOR_WEIGHT: f32 = 0.5;
/// Parameters for [`CoreLoop::execute_hybrid_search`].
pub(in crate::data::executor) struct HybridSearchParams<'a> {
pub tid: u64,
pub collection: &'a str,
pub query_vector: &'a [f32],
pub query_text: &'a str,
pub top_k: usize,
pub ef_search: usize,
pub fuzzy: bool,
pub vector_weight: f32,
pub filter_bitmap: Option<&'a nodedb_types::SurrogateBitmap>,
pub rls_filters: &'a [u8],
pub score_alias: Option<&'a str>,
}
impl CoreLoop {
/// Execute a hybrid search: vector + text, fused via weighted RRF.
///
/// `score_alias` overrides the response field name for the RRF score
/// column. When `None` the executor uses the fixed default `rrf_score`.
pub(in crate::data::executor) fn execute_hybrid_search(
&self,
task: &ExecutionTask,
params: HybridSearchParams<'_>,
) -> Response {
let HybridSearchParams {
tid,
collection,
query_vector,
query_text,
top_k,
ef_search,
fuzzy,
vector_weight,
filter_bitmap,
rls_filters,
score_alias,
} = params;
let tenant_id = TenantId::new(tid);
debug!(
core = self.core_id,
tid,
%collection,
%query_text,
top_k,
vector_weight,
"hybrid search"
);
// Scan-quiesce gate.
let _scan_guard = match self.acquire_scan_guard(task, tid, collection) {
Ok(g) => g,
Err(resp) => return resp,
};
let weight = if vector_weight <= 0.0 || vector_weight >= 1.0 {
DEFAULT_VECTOR_WEIGHT
} else {
vector_weight
};
let text_weight = 1.0 - weight;
// Fetch more candidates than top_k from each engine so RRF has
// enough material to fuse. 3x is a good balance.
let fetch_k = top_k.saturating_mul(3).max(20);
// 1. Vector search.
let index_key =
CoreLoop::vector_index_key(task.request.database_id.as_u64(), tid, collection, "");
let vector_collection = self.vector_collections.get(&index_key);
let vector_results = if let Some(index) = vector_collection {
if index.is_empty() {
Vec::new()
} else {
let ef = if ef_search > 0 {
ef_search.max(fetch_k)
} else {
fetch_k.saturating_mul(4).max(64)
};
match filter_bitmap {
Some(surrogate_bm) => {
let mut buf = Vec::with_capacity(surrogate_bm.0.serialized_size());
if surrogate_bm.0.serialize_into(&mut buf).is_ok() {
index.search_with_bitmap_bytes(query_vector, fetch_k, ef, &buf)
} else {
index.search(query_vector, fetch_k, ef)
}
}
None => index.search(query_vector, fetch_k, ef),
}
}
} else {
Vec::new()
};
// 2. Text search (no surrogate prefilter for the text leg of hybrid search).
let text_results = self
.inverted
.search(
task.request.database_id.as_u64(),
tenant_id,
collection,
FtsSearchParams {
query: query_text,
top_k: fetch_k,
fuzzy_enabled: fuzzy,
mode: QueryMode::And,
prefilter: None,
},
)
.unwrap_or_default();
// 3. Build ranked lists for weighted RRF.
// Higher weight → lower k → steeper rank discount → more influence.
use crate::query::fusion::{RankedResult, reciprocal_rank_fusion_weighted};
let base_k = 60.0_f64;
let k_vector = if weight > 0.01 {
base_k / weight as f64
} else {
base_k * 100.0
};
let k_text = if text_weight > 0.01 {
base_k / text_weight as f64
} else {
base_k * 100.0
};
// Translate vector local-hnsw IDs to surrogate-hex doc_ids so the
// vector and text legs share the same RRF key space. Headless rows
// (no surrogate binding) fall back to a non-fusable sentinel —
// they cannot match any FTS doc_id, which is the correct behavior.
//
// Inside a transaction, read-your-own-writes: the vector and text legs
// must also observe this transaction's staged document writes, folded
// in via the shared overlay splice (which reuses the single-source
// vector/FTS overlay merges). Outside a transaction the committed-only
// construction below runs unchanged.
let (vector_ranked, text_ranked): (Vec<RankedResult>, Vec<RankedResult>) =
if let Some(txn_id) = task.request.txn_id {
self.hybrid_ranked_with_overlay(
super::hybrid_overlay::HybridOverlayParams {
txn_id,
database_id: task.request.database_id,
tid: tenant_id,
collection,
query_vector,
query_text,
fetch_k,
filter_bitmap,
},
&vector_results,
vector_collection,
&text_results,
)
} else {
let vector_ranked: Vec<RankedResult> = vector_results
.iter()
.enumerate()
.map(|(rank, r)| RankedResult {
document_id: super::vector_search::vector_leg_doc_id(
vector_collection,
r.id,
),
rank,
score: r.distance,
source: "vector",
})
.collect();
let text_ranked: Vec<RankedResult> = text_results
.iter()
.enumerate()
.map(|(rank, r)| RankedResult {
document_id: crate::engine::document::store::surrogate_to_doc_id(r.doc_id),
rank,
score: r.score,
source: "text",
})
.collect();
(vector_ranked, text_ranked)
};
let fused = reciprocal_rank_fusion_weighted(
&[vector_ranked, text_ranked],
&[k_vector, k_text],
top_k,
);
// Build response with per-engine rank diagnostics.
// RLS post-fusion: filter fused results by looking up each document.
let results: Vec<_> = fused
.iter()
.filter(|f| {
if rls_filters.is_empty() {
return true;
}
match self.sparse.get(
task.request.database_id.as_u64(),
tid,
collection,
&f.document_id,
) {
Ok(Some(bytes)) => {
super::rls_eval::rls_check_msgpack_bytes(rls_filters, &bytes)
}
_ => false,
}
})
.map(|f| {
let vector_rank = vector_results.iter().position(|r| {
let doc_id = vector_collection
.and_then(|c| c.get_surrogate(r.id))
.map(crate::engine::document::store::surrogate_to_doc_id)
.unwrap_or_else(|| format!("__local_{}", r.id));
doc_id == f.document_id
});
let text_rank = text_results.iter().position(|r| {
crate::engine::document::store::surrogate_to_doc_id(r.doc_id) == f.document_id
});
super::super::response_codec::HybridSearchHit {
doc_id: &f.document_id,
score_field: score_alias.unwrap_or("rrf_score"),
rrf_score: f.rrf_score,
vector_rank,
text_rank,
}
})
.collect();
if let Some(ref m) = self.metrics {
m.record_fts_search(0);
}
match super::super::response_codec::encode(&results) {
Ok(payload) => self.response_with_payload(task, payload),
Err(e) => self.response_error(
task,
ErrorCode::Internal {
detail: e.to_string(),
},
),
}
}
}