diff --git a/.github/actions/check-files/action.yml b/.github/actions/check-files/action.yml
index a4f1e61..303191a 100644
--- a/.github/actions/check-files/action.yml
+++ b/.github/actions/check-files/action.yml
@@ -24,7 +24,7 @@ runs:
         echo "combined_files=$COMBINED_FILES" >> $GITHUB_OUTPUT
     - name: Get changed files
       id: changed-files
-      uses: tj-actions/changed-files@v47.0.0
+      uses: tj-actions/changed-files@24d32ffd492484c1d75e0c0b894501ddb9d30d62 # v47.0.0
       with:
         files: ${{ steps.combine-files.outputs.combined_files }}
         files_separator: ","
diff --git a/.github/workflows/CI.yml b/.github/workflows/CI.yml
index 7cba0d5..db13712 100644
--- a/.github/workflows/CI.yml
+++ b/.github/workflows/CI.yml
@@ -26,7 +26,7 @@ jobs:
   Get-CI-Image-Tag:
     needs: check-files
     if: needs.check-files.outputs.RUN_BUILD_AND_TEST == 'true'
-    uses: opensearch-project/opensearch-build/.github/workflows/get-ci-image-tag.yml@c2498b758c08fb7bc48476509a5fc1b8dd5f7493 # main
+    uses: opensearch-project/opensearch-build/.github/workflows/get-ci-image-tag.yml@6631dc102f1e64c182eccfc48a319a87bc9079f6 # main
     with:
       product: opensearch
 
diff --git a/nsparse/inverted_index.cpp b/nsparse/inverted_index.cpp
index 7813bca..806e837 100644
--- a/nsparse/inverted_index.cpp
+++ b/nsparse/inverted_index.cpp
@@ -135,6 +135,10 @@ void score_essential_terms(std::vector<DirectTermScorer>& scorers,
             if (doc >= window_end) {
                 break;
             }
+            if (doc < window_base) {
+                ++scorer.current_index;
+                continue;
+            }
             int slot = doc - window_base;
             bitmap[slot >> 6] |= (1ULL << (slot & 63));
             window_scores[slot] +=
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index 2f58059..2d59fa7 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -15,6 +15,7 @@ set(NSPARSE_TEST_SRC
     index_factory_test.cpp
     index_io_test.cpp
     inverted_index_test.cpp
+    inverted_index_recall_test.cpp
     inverted_list_clusters_test.cpp
     inverted_lists_test.cpp
     kmeans_utils_test.cpp
diff --git a/tests/inverted_index_recall_test.cpp b/tests/inverted_index_recall_test.cpp
new file mode 100644
index 0000000..80f0529
--- /dev/null
+++ b/tests/inverted_index_recall_test.cpp
@@ -0,0 +1,266 @@
+/**
+ * Copyright OpenSearch Contributors
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#include "nsparse/inverted_index.h"
+
+#include <gtest/gtest.h>
+
+#include <algorithm>
+#include <cmath>
+#include <map>
+#include <random>
+#include <set>
+#include <vector>
+
+#include "nsparse/types.h"
+
+namespace nsparse {
+namespace {
+
+// Brute-force top-k: compute dot product for all docs and return top-k doc ids.
+std::vector<idx_t> brute_force_top_k(
+    const std::vector<std::vector<std::pair<term_t, float>>>& docs,
+    const std::vector<std::pair<term_t, float>>& query, int k) {
+    std::vector<std::pair<float, idx_t>> scores;
+    for (idx_t d = 0; d < static_cast<idx_t>(docs.size()); ++d) {
+        float score = 0.0F;
+        // Build a quick lookup for doc terms.
+        std::map<term_t, float> doc_map;
+        for (auto& [t, v] : docs[d]) {
+            doc_map[t] = v;
+        }
+        for (auto& [qt, qv] : query) {
+            auto it = doc_map.find(qt);
+            if (it != doc_map.end()) {
+                score += qv * it->second;
+            }
+        }
+        if (score > 0.0F) {
+            scores.emplace_back(score, d);
+        }
+    }
+    // Sort descending by score, then ascending by doc_id for tie-breaking.
+    std::sort(scores.begin(), scores.end(),
+              [](const auto& a, const auto& b) {
+                  if (a.first != b.first) return a.first > b.first;
+                  return a.second < b.second;
+              });
+    std::vector<idx_t> result;
+    for (int i = 0; i < k && i < static_cast<int>(scores.size()); ++i) {
+        result.push_back(scores[i].second);
+    }
+    return result;
+}
+
+// Generate random sparse documents.
+struct RandomSparseData {
+    int dim;
+    int n_docs;
+    int n_queries;
+    int avg_nnz;
+    std::vector<std::vector<std::pair<term_t, float>>> docs;
+    std::vector<std::vector<std::pair<term_t, float>>> queries;
+};
+
+RandomSparseData generate_random_data(int dim, int n_docs, int n_queries,
+                                       int avg_nnz, unsigned seed = 42) {
+    std::mt19937 rng(seed);
+    std::uniform_int_distribution<int> nnz_dist(
+        std::max(1, avg_nnz / 2), avg_nnz * 2);
+    std::uniform_int_distribution<int> term_dist(0, dim - 1);
+    std::uniform_real_distribution<float> val_dist(0.01F, 2.0F);
+
+    RandomSparseData data;
+    data.dim = dim;
+    data.n_docs = n_docs;
+    data.n_queries = n_queries;
+    data.avg_nnz = avg_nnz;
+
+    // Generate documents.
+    for (int d = 0; d < n_docs; ++d) {
+        int nnz = nnz_dist(rng);
+        std::set<term_t> used_terms;
+        std::vector<std::pair<term_t, float>> doc;
+        for (int j = 0; j < nnz; ++j) {
+            term_t t = static_cast<term_t>(term_dist(rng));
+            if (used_terms.insert(t).second) {
+                doc.emplace_back(t, val_dist(rng));
+            }
+        }
+        // Sort by term for consistent ordering.
+        std::sort(doc.begin(), doc.end());
+        data.docs.push_back(std::move(doc));
+    }
+
+    // Generate queries.
+    for (int q = 0; q < n_queries; ++q) {
+        int nnz = nnz_dist(rng);
+        std::set<term_t> used_terms;
+        std::vector<std::pair<term_t, float>> query;
+        for (int j = 0; j < nnz; ++j) {
+            term_t t = static_cast<term_t>(term_dist(rng));
+            if (used_terms.insert(t).second) {
+                query.emplace_back(t, val_dist(rng));
+            }
+        }
+        std::sort(query.begin(), query.end());
+        data.queries.push_back(std::move(query));
+    }
+
+    return data;
+}
+
+void add_docs_from_data(
+    InvertedIndex& index,
+    const std::vector<std::vector<std::pair<term_t, float>>>& docs) {
+    std::vector<idx_t> indptr;
+    std::vector<term_t> indices;
+    std::vector<float> values;
+
+    indptr.push_back(0);
+    for (const auto& doc : docs) {
+        for (const auto& [term, value] : doc) {
+            indices.push_back(term);
+            values.push_back(value);
+        }
+        indptr.push_back(static_cast<idx_t>(indices.size()));
+    }
+
+    index.add(static_cast<idx_t>(docs.size()), indptr.data(), indices.data(),
+              values.data());
+}
+
+float compute_recall(const std::vector<idx_t>& predicted,
+                     const std::vector<idx_t>& ground_truth) {
+    std::set<idx_t> gt_set(ground_truth.begin(), ground_truth.end());
+    int hits = 0;
+    for (auto id : predicted) {
+        if (id != INVALID_IDX && gt_set.count(id)) {
+            ++hits;
+        }
+    }
+    return ground_truth.empty()
+               ? 1.0F
+               : static_cast<float>(hits) / static_cast<float>(gt_set.size());
+}
+
+// Test recall with various dataset sizes and k values.
+class InvertedIndexRecallTest
+    : public ::testing::TestWithParam<std::tuple<int, int, int, int>> {};
+
+TEST_P(InvertedIndexRecallTest, recall_should_be_perfect) {
+    auto [n_docs, dim, avg_nnz, k] = GetParam();
+    int n_queries = 50;
+
+    auto data = generate_random_data(dim, n_docs, n_queries, avg_nnz);
+
+    InvertedIndex index(dim);
+    add_docs_from_data(index, data.docs);
+    index.build();
+    Index* idx = &index;
+
+    float total_recall = 0.0F;
+    int n_valid_queries = 0;
+
+    for (int q = 0; q < n_queries; ++q) {
+        const auto& query = data.queries[q];
+        if (query.empty()) continue;
+
+        // Build query CSR.
+        std::vector<idx_t> q_indptr = {0,
+                                        static_cast<idx_t>(query.size())};
+        std::vector<term_t> q_indices;
+        std::vector<float> q_values;
+        for (auto& [t, v] : query) {
+            q_indices.push_back(t);
+            q_values.push_back(v);
+        }
+
+        std::vector<idx_t> labels(k, INVALID_IDX);
+        std::vector<float> distances(k, -1.0F);
+
+        idx->search(1, q_indptr.data(), q_indices.data(), q_values.data(), k,
+                     distances.data(), labels.data());
+
+        // Brute-force ground truth.
+        auto gt = brute_force_top_k(data.docs, query, k);
+
+        float recall = compute_recall(
+            std::vector<idx_t>(labels.begin(), labels.end()), gt);
+
+        total_recall += recall;
+        ++n_valid_queries;
+
+        // Print failing queries for debugging.
+        if (recall < 1.0F) {
+            std::cout << "Query " << q << ": recall=" << recall
+                      << " predicted=[";
+            for (int i = 0; i < k; ++i) {
+                if (i > 0) std::cout << ",";
+                std::cout << labels[i];
+            }
+            std::cout << "] gt=[";
+            for (size_t i = 0; i < gt.size(); ++i) {
+                if (i > 0) std::cout << ",";
+                std::cout << gt[i];
+            }
+            std::cout << "]" << std::endl;
+
+            // Print scores for debugging.
+            std::cout << "  predicted_scores=[";
+            for (int i = 0; i < k; ++i) {
+                if (i > 0) std::cout << ",";
+                std::cout << distances[i];
+            }
+            std::cout << "]" << std::endl;
+
+            // Print ground truth scores.
+            std::cout << "  gt_scores=[";
+            for (size_t i = 0; i < gt.size(); ++i) {
+                if (i > 0) std::cout << ",";
+                // Compute score for gt[i].
+                float score = 0.0F;
+                std::map<term_t, float> doc_map;
+                for (auto& [t, v] : data.docs[gt[i]]) doc_map[t] = v;
+                for (auto& [qt, qv] : query) {
+                    auto it = doc_map.find(qt);
+                    if (it != doc_map.end()) score += qv * it->second;
+                }
+                std::cout << score;
+            }
+            std::cout << "]" << std::endl;
+        }
+    }
+
+    float avg_recall =
+        n_valid_queries > 0 ? total_recall / n_valid_queries : 1.0F;
+    std::cout << "Average recall: " << avg_recall << " (n_docs=" << n_docs
+              << ", dim=" << dim << ", avg_nnz=" << avg_nnz << ", k=" << k
+              << ")" << std::endl;
+
+    // InvertedIndex should have PERFECT recall (it's an exact algorithm).
+    EXPECT_FLOAT_EQ(avg_recall, 1.0F)
+        << "InvertedIndex recall should be 1.0 but got " << avg_recall;
+}
+
+INSTANTIATE_TEST_SUITE_P(
+    RecallTests, InvertedIndexRecallTest,
+    ::testing::Values(
+        // (n_docs, dim, avg_nnz, k)
+        // Small: single window, no pruning.
+        std::make_tuple(100, 500, 10, 10),
+        // Medium: multiple windows, pruning active.
+        std::make_tuple(5000, 1000, 20, 10),
+        // Large: many windows, aggressive pruning.
+        std::make_tuple(10000, 2000, 30, 10),
+        // Large with higher k.
+        std::make_tuple(10000, 2000, 30, 50),
+        // Dense queries, many terms.
+        std::make_tuple(5000, 500, 50, 10),
+        // Very sparse queries and docs.
+        std::make_tuple(10000, 5000, 5, 10)));
+
+}  // namespace
+}  // namespace nsparse
diff --git a/tests/inverted_index_test.cpp b/tests/inverted_index_test.cpp
index 427d9d2..95f379d 100644
--- a/tests/inverted_index_test.cpp
+++ b/tests/inverted_index_test.cpp
@@ -411,5 +411,58 @@ TEST(InvertedIndexIO, roundtrip_preserves_fourcc) {
     delete loaded;
 }
 
+// Regression test: score_essential_terms must skip docs before window_base.
+// When scorers span multiple windows (>4096 docs) and re-partitioning changes
+// first_essential, a newly-essential scorer may have its cursor pointing at
+// docs before the current window. Without the bounds check, this produces a
+// negative slot index causing SIGBUS / out-of-bounds write.
+TEST(InvertedIndexSearch, search_multi_window_no_crash) {
+    // Use 2 terms. term0 has docs in early range, term1 has docs spread across
+    // multiple windows (>4096). This forces the BMW algorithm to process
+    // multiple scoring windows and potentially re-partition essential terms.
+    constexpr int kDim = 2;
+    constexpr int kNumDocs = 6000;  // > kScoreWindowSize (4096)
+    InvertedIndex index(kDim);
+
+    // Build docs: every doc has term0, every 3rd doc also has term1.
+    // This creates different posting list lengths that affect max_score
+    // partitioning.
+    std::vector<std::map<int, float>> docs;
+    docs.reserve(kNumDocs);
+    for (int i = 0; i < kNumDocs; ++i) {
+        std::map<int, float> doc;
+        doc[0] = 0.1F + static_cast<float>(i % 10) * 0.01F;
+        if (i % 3 == 0) {
+            doc[1] = 0.5F + static_cast<float>(i % 7) * 0.05F;
+        }
+        docs.push_back(doc);
+    }
+    add_docs(index, docs);
+    index.build();
+
+    // Query both terms — forces multi-term scoring across multiple windows.
+    std::vector<idx_t> query_indptr = {0, 2};
+    std::vector<term_t> query_indices = {0, 1};
+    std::vector<float> query_values = {1.0F, 1.0F};
+
+    constexpr int kTopK = 10;
+    std::vector<idx_t> labels(kTopK, -1);
+    std::vector<float> distances(kTopK, -1.0F);
+
+    Index* idx = &index;
+    // The main assertion is that this does not crash (SIGBUS).
+    idx->search(1, query_indptr.data(), query_indices.data(),
+                query_values.data(), kTopK, distances.data(), labels.data());
+
+    // Verify we got valid results.
+    EXPECT_GE(labels[0], 0);
+    EXPECT_GT(distances[0], 0.0F);
+
+    // Results should be sorted by score descending.
+    for (int i = 1; i < kTopK; ++i) {
+        EXPECT_GE(distances[i - 1], distances[i]);
+    }
+}
+
 }  // namespace
 }  // namespace nsparse