From 9592e064486ed632b5ad339425e77fcc7f4d4266 Mon Sep 17 00:00:00 2001
From: root <root@RuVector.wally.local>
Date: Tue, 12 May 2026 19:53:45 +0000
Subject: [PATCH] fix(ruvector-router-core): rebuild index from storage on open
 + fix HNSW result heap

  - VectorDB::new() now loads all persisted vectors into the HNSW index
    on startup, fixing search() always returning 0 results after restart
    (closes #430)

  - Fix result heap in search_knn_internal: was min-heap, must be max-heap
    so peek() returns the furthest result for correct pruning at scale

  - Remove ef_construction.min(m*2) beam clamp in insert() so the full
    candidate pool is used during graph construction
---
 crates/ruvector-router-core/src/index.rs     | 17 ++++++++---------
 crates/ruvector-router-core/src/vector_db.rs | 16 ++++++++++++++--
 2 files changed, 22 insertions(+), 11 deletions(-)

diff --git a/crates/ruvector-router-core/src/index.rs b/crates/ruvector-router-core/src/index.rs
index 559ddf566..83f2904ad 100644
--- a/crates/ruvector-router-core/src/index.rs
+++ b/crates/ruvector-router-core/src/index.rs
@@ -116,8 +116,7 @@ impl HnswIndex {
         }
 
         // Find nearest neighbors (safe now - no locks held)
-        let neighbors =
-            self.search_knn_internal(&vector, self.config.ef_construction.min(self.config.m * 2));
+        let neighbors = self.search_knn_internal(&vector, self.config.ef_construction);
 
         // Re-acquire graph lock for modifications
         let mut graph = self.graph.write();
@@ -187,7 +186,7 @@ impl HnswIndex {
         let entry_id = entry_point.as_ref().unwrap();
         let mut visited = HashSet::new();
         let mut candidates = BinaryHeap::new();
-        let mut result = BinaryHeap::new();
+        let mut result: BinaryHeap<std::cmp::Reverse<Neighbor>> = BinaryHeap::new();
 
         // Calculate distance to entry point
         if let Some(entry_vec) = vectors.get(entry_id) {
@@ -199,14 +198,14 @@ impl HnswIndex {
             };
 
             candidates.push(neighbor.clone());
-            result.push(neighbor);
+            result.push(std::cmp::Reverse(neighbor));
             visited.insert(entry_id.clone());
         }
 
         // Search phase
         while let Some(current) = candidates.pop() {
             // Check if we should continue
-            if let Some(furthest) = result.peek() {
+            if let Some(std::cmp::Reverse(furthest)) = result.peek() {
                 if current.distance > furthest.distance && result.len() >= ef {
                     break;
                 }
@@ -235,11 +234,11 @@ impl HnswIndex {
 
                         // Add to results if better than current worst
                         if result.len() < ef {
-                            result.push(neighbor);
-                        } else if let Some(worst) = result.peek() {
+                            result.push(std::cmp::Reverse(neighbor));
+                        } else if let Some(std::cmp::Reverse(worst)) = result.peek() {
                             if dist < worst.distance {
                                 result.pop();
-                                result.push(neighbor);
+                                result.push(std::cmp::Reverse(neighbor));
                             }
                         }
                     }
@@ -248,7 +247,7 @@ impl HnswIndex {
         }
 
         // Convert to sorted vector
-        let mut sorted_results: Vec<Neighbor> = result.into_iter().collect();
+        let mut sorted_results: Vec<Neighbor> = result.into_iter().map(|r| r.0).collect();
         sorted_results.sort_by(|a, b| {
             a.distance
                 .partial_cmp(&b.distance)
diff --git a/crates/ruvector-router-core/src/vector_db.rs b/crates/ruvector-router-core/src/vector_db.rs
index eacb8035b..88a58d7b2 100644
--- a/crates/ruvector-router-core/src/vector_db.rs
+++ b/crates/ruvector-router-core/src/vector_db.rs
@@ -32,8 +32,20 @@ impl VectorDB {
 
         let index = Arc::new(HnswIndex::new(hnsw_config));
 
-        let stats = Arc::new(RwLock::new(VectorDbStats {
-            total_vectors: 0,
+        // Rebuild index from persisted vectors (fixes search returning 0 results after restart)
+        let stored_ids = storage.get_all_ids()?;
+          if !stored_ids.is_empty() {
+              let mut entries = Vec::with_capacity(stored_ids.len());
+              for id in &stored_ids {
+                  if let Some(vector) = storage.get(id)? {
+                      entries.push((id.clone(), vector));
+                  }
+              }
+              index.insert_batch(entries)?;
+          }
+
+          let stats = Arc::new(RwLock::new(VectorDbStats {
+              total_vectors: stored_ids.len(),
             index_size_bytes: 0,
             storage_size_bytes: 0,
             avg_query_latency_us: 0.0,