Skip to content

Solution for long term memory for agent coding CLIs and to facilitate handoff between different agent vendors

Rust

fix: align hybrid authority candidate windows

AkitaOnRailsf55e32c07194
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 35fdf753..1b000e68 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -39,6 +39,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
including subdirectories and linked worktrees. (#259)
### Fixed
+- Hybrid search now gives its FTS, vector, and graph streams the same bounded
+ candidate window used by authority-aware FTS search. Small result limits no
+ longer exclude a canonical page before post-fusion authority ranking can
+ promote it, and candidate-limit arithmetic is saturating throughout. (#276)
- Forced workspace deletion now removes the immutable managed-workstream
segment directories whose database rows are removed by the workspace
cascade. Its admin report includes workstream/run counts and IDs, and raw
diff --git a/crates/ai-memory-store/src/lib.rs b/crates/ai-memory-store/src/lib.rs
index 5baa8095..7a4e9204 100644
--- a/crates/ai-memory-store/src/lib.rs
+++ b/crates/ai-memory-store/src/lib.rs
@@ -2214,7 +2214,7 @@ mod tests {
String::new(),
String::new(),
0,
- 10,
+ 1,
)
.await
.unwrap();
diff --git a/crates/ai-memory-store/src/reader.rs b/crates/ai-memory-store/src/reader.rs
index d9ec7b3e..8ed08601 100644
--- a/crates/ai-memory-store/src/reader.rs
+++ b/crates/ai-memory-store/src/reader.rs
@@ -2139,14 +2139,15 @@ impl ReaderPool {
dim: u32,
limit: usize,
) -> StoreResult<Vec<PageHit>> {
+ // Authority is applied after RRF, so every stream needs the same
+ // bounded candidate window used by FTS-only search. A `limit * 2`
+ // window was too narrow at small limits for a canonical page to enter
+ // the fused pool and receive its bounded promotion.
+ let candidate_limit = authority_candidate_limit(limit);
+
// Fetch FTS5 hits first.
let fts_candidates = self
- .search_page_candidates_for_project(
- workspace_id,
- project_id,
- query,
- limit.saturating_mul(2),
- )
+ .search_page_candidates_for_project(workspace_id, project_id, query, candidate_limit)
.await?;
let mut authorities: std::collections::HashMap<PageId, PageAuthority> = fts_candidates
.iter()
@@ -2166,7 +2167,7 @@ impl ReaderPool {
provider,
model,
dim,
- limit * 2,
+ candidate_limit,
)
.await?;
}
@@ -2183,7 +2184,7 @@ impl ReaderPool {
}
}
let graph_hits = self
- .graph_neighbors_for_project(workspace_id, project_id, seed_ids, limit * 2)
+ .graph_neighbors_for_project(workspace_id, project_id, seed_ids, candidate_limit)
.await?;
// RRF fuse: score(d) = Σ 1/(k + rank_i(d)) over rankers.
@@ -5436,6 +5437,17 @@ mod tests {
AgentKind, Handoff, HandoffId, HandoffState, NewHandoff, ProjectId, SessionId, WorkspaceId,
};
+ #[test]
+ fn authority_candidate_window_is_bounded_and_saturating() {
+ use super::authority_candidate_limit;
+
+ assert_eq!(authority_candidate_limit(0), 0);
+ assert_eq!(authority_candidate_limit(1), 20);
+ assert_eq!(authority_candidate_limit(10), 40);
+ assert_eq!(authority_candidate_limit(1_000), 1_300);
+ assert_eq!(authority_candidate_limit(usize::MAX), usize::MAX);
+ }
+
/// Build an open handoff for the pure-selection tests. `manual` toggles
/// the manual (`from_session_id == None`) vs auto distinction; `t` is the
/// creation time in microseconds; `summary` doubles as a label to assert
diff --git a/docs/vector-backend-policy.md b/docs/vector-backend-policy.md
index 868480d1..8a4c5297 100644
--- a/docs/vector-backend-policy.md
+++ b/docs/vector-backend-policy.md
@@ -11,9 +11,10 @@ pages per project. At that size, brute-force cosine is simple,
inspectable, and fast enough, especially because vector retrieval is only
one signal in the query path. `memory_query` already combines FTS5,
graph-neighbor expansion, optional vector RRF, and raw observation
-fallback. A bounded page-authority adjustment runs after those candidate
-streams, so semantic similarity cannot by itself declare a session page more
-canonical than a maintained decision.
+fallback. Each stream contributes through the same bounded candidate window,
+then a page-authority adjustment runs after fusion, so semantic similarity
+cannot by itself declare a session page more canonical than a maintained
+decision.
Adding `sqlite-vec` would add operational surface before it has proven
value: