add and verify filter option

split out common subroutines

13 files changed, 853 insertions, 259 deletions

diff --git a/eval/src/tests/ann/CMakeLists.txt b/eval/src/tests/ann/CMakeLists.txt
index 52b4d675d9c..34babf1412f 100644
--- a/eval/src/tests/ann/CMakeLists.txt
+++ b/eval/src/tests/ann/CMakeLists.txt
@@ -10,6 +10,16 @@ vespa_add_executable(eval_sift_benchmark_app
     vespaeval
 )
 
+vespa_add_executable(eval_gist_benchmark_app
+    SOURCES
+    gist_benchmark.cpp
+    xp-annoy-nns.cpp
+    xp-hnswlike-nns.cpp
+    xp-lsh-nns.cpp
+    DEPENDS
+    vespaeval
+)
+
 vespa_add_executable(eval_remove_bm_app
     SOURCES
     remove-bm.cpp
diff --git a/eval/src/tests/ann/find-with-nns.h b/eval/src/tests/ann/find-with-nns.h
new file mode 100644
index 00000000000..3481b403f86
--- /dev/null
+++ b/eval/src/tests/ann/find-with-nns.h
@@ -0,0 +1,12 @@
+// Copyright 2020 Oath Inc. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+
+TopK find_with_nns(uint32_t sk, NNS_API &nns, uint32_t qid) {
+    TopK result;
+    const PointVector &qv = generatedQueries[qid];
+    vespalib::ConstArrayRef<float> query(qv.v, NUM_DIMS);
+    auto rv = nns.topK(result.K, query, sk);
+    for (size_t i = 0; i < result.K; ++i) {
+        result.hits[i] = Hit(rv[i].docid, rv[i].sq.distance);
+    }
+    return result;
+}
diff --git a/eval/src/tests/ann/for-sift-top-k.h b/eval/src/tests/ann/for-sift-top-k.h
index ba91cb2aebc..8a659a507bc 100644
--- a/eval/src/tests/ann/for-sift-top-k.h
+++ b/eval/src/tests/ann/for-sift-top-k.h
@@ -6,7 +6,7 @@ struct TopK {
     static constexpr size_t K = 100;
     Hit hits[K];
 
-    size_t recall(const TopK &other) {
+    size_t recall(const TopK &other) const {
         size_t overlap = 0;
         size_t i = 0;
         size_t j = 0;
diff --git a/eval/src/tests/ann/gist_benchmark.cpp b/eval/src/tests/ann/gist_benchmark.cpp
new file mode 100644
index 00000000000..45559fc2557
--- /dev/null
+++ b/eval/src/tests/ann/gist_benchmark.cpp
@@ -0,0 +1,295 @@
+// Copyright 2020 Oath Inc. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+
+#include <vespa/vespalib/testkit/test_kit.h>
+#include <vespa/vespalib/util/priority_queue.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <unistd.h>
+#include <stdio.h>
+#include <chrono>
+
+#define NUM_DIMS 960
+#define NUM_DOCS 200000
+#define NUM_REACH 10000
+#define NUM_Q 1000
+
+#include "doc_vector_access.h"
+#include "nns.h"
+#include "for-sift-hit.h"
+#include "for-sift-top-k.h"
+
+std::vector<TopK> bruteforceResults;
+
+struct PointVector {
+    float v[NUM_DIMS];
+    using ConstArr = vespalib::ConstArrayRef<float>;
+    operator ConstArr() const { return ConstArr(v, NUM_DIMS); }
+};
+
+static PointVector *aligned_alloc(size_t num) {
+    size_t num_bytes = num * sizeof(PointVector);
+    double mega_bytes = num_bytes / (1024.0*1024.0);
+    fprintf(stderr, "allocate %.2f MB of vectors\n", mega_bytes);
+    char *mem = (char *)malloc(num_bytes + 512);
+    mem += 512;
+    size_t val = (size_t)mem;
+    size_t unalign = val % 512;
+    mem -= unalign;
+    return reinterpret_cast<PointVector *>(mem);
+}
+
+static PointVector *generatedQueries = aligned_alloc(NUM_Q);
+static PointVector *generatedDocs = aligned_alloc(NUM_DOCS);
+
+struct DocVectorAdapter : public DocVectorAccess<float>
+{
+    vespalib::ConstArrayRef<float> get(uint32_t docid) const override {
+        ASSERT_TRUE(docid < NUM_DOCS);
+        return generatedDocs[docid];
+    }
+};
+
+double computeDistance(const PointVector &query, uint32_t docid) {
+    const PointVector &docvector = generatedDocs[docid];
+    return l2distCalc.l2sq_dist(query, docvector);
+}
+
+void read_queries(std::string fn) {
+    int fd = open(fn.c_str(), O_RDONLY);
+    ASSERT_TRUE(fd > 0);
+    int d;
+    size_t rv;
+    fprintf(stderr, "reading %u queries from %s\n", NUM_Q, fn.c_str());
+    for (uint32_t qid = 0; qid < NUM_Q; ++qid) {
+        rv = read(fd, &d, 4);
+        ASSERT_EQUAL(rv, 4u);
+        ASSERT_EQUAL(d, NUM_DIMS);
+        rv = read(fd, &generatedQueries[qid].v, NUM_DIMS*sizeof(float));
+        ASSERT_EQUAL(rv, sizeof(PointVector));
+    }
+    close(fd);
+}
+
+void read_docs(std::string fn) {
+    int fd = open(fn.c_str(), O_RDONLY);
+    ASSERT_TRUE(fd > 0);
+    int d;
+    size_t rv;
+    fprintf(stderr, "reading %u doc vectors from %s\n", NUM_DOCS, fn.c_str());
+    for (uint32_t docid = 0; docid < NUM_DOCS; ++docid) {
+        rv = read(fd, &d, 4);
+        ASSERT_EQUAL(rv, 4u);
+        ASSERT_EQUAL(d, NUM_DIMS);
+        rv = read(fd, &generatedDocs[docid].v, NUM_DIMS*sizeof(float));
+        ASSERT_EQUAL(rv, sizeof(PointVector));
+    }
+    close(fd);
+}
+
+using TimePoint = std::chrono::steady_clock::time_point;
+using Duration = std::chrono::steady_clock::duration;
+
+double to_ms(Duration elapsed) {
+    std::chrono::duration<double, std::milli> ms(elapsed);
+    return ms.count();
+}
+
+void read_data(std::string dir) {
+    TimePoint bef = std::chrono::steady_clock::now();
+    read_queries(dir + "/gist_query.fvecs");
+    TimePoint aft = std::chrono::steady_clock::now();
+    fprintf(stderr, "read queries: %.3f ms\n", to_ms(aft - bef));
+    bef = std::chrono::steady_clock::now();
+    read_docs(dir + "/gist_base.fvecs");
+    aft = std::chrono::steady_clock::now();
+    fprintf(stderr, "read docs: %.3f ms\n", to_ms(aft - bef));
+}
+
+
+struct BfHitComparator {
+    bool operator() (const Hit &lhs, const Hit& rhs) const {
+        if (lhs.distance < rhs.distance) return false;
+        if (lhs.distance > rhs.distance) return true;
+        return (lhs.docid > rhs.docid);
+    }
+};
+
+class BfHitHeap {
+private:
+    size_t _size;
+    vespalib::PriorityQueue<Hit, BfHitComparator> _priQ;
+public:
+    explicit BfHitHeap(size_t maxSize) : _size(maxSize), _priQ() {
+        _priQ.reserve(maxSize);
+    }
+    ~BfHitHeap() {}
+    void maybe_use(const Hit &hit) {
+        if (_priQ.size() < _size) {
+            _priQ.push(hit);
+        } else if (hit.distance < _priQ.front().distance) {
+            _priQ.front() = hit;
+            _priQ.adjust();
+        }
+    }
+    std::vector<Hit> bestHits() {
+        std::vector<Hit> result;
+        size_t i = _priQ.size();
+        result.resize(i);
+        while (i-- > 0) {
+            result[i] = _priQ.front();
+            _priQ.pop_front();
+        }
+        return result;
+    }
+};
+
+TopK bruteforce_nns(const PointVector &query) {
+    TopK result;
+    BfHitHeap heap(result.K);
+    for (uint32_t docid = 0; docid < NUM_DOCS; ++docid) {
+        const PointVector &docvector = generatedDocs[docid];
+        double d = l2distCalc.l2sq_dist(query, docvector);
+        Hit h(docid, d);
+        heap.maybe_use(h);
+    }
+    std::vector<Hit> best = heap.bestHits();
+    for (size_t i = 0; i < result.K; ++i) {
+        result.hits[i] = best[i];
+    }
+    return result;
+}
+
+void verifyBF(uint32_t qid) {
+    const PointVector &query = generatedQueries[qid];
+    TopK &result = bruteforceResults[qid];
+    double min_distance = result.hits[0].distance;
+    std::vector<double> all_c2;
+    for (uint32_t i = 0; i < NUM_DOCS; ++i) {
+        double dist = computeDistance(query, i);
+        if (dist < min_distance) {
+            fprintf(stderr, "WARN dist %.9g < mindist %.9g\n", dist, min_distance);
+        }
+        EXPECT_FALSE(dist+0.000001 < min_distance);
+        if (min_distance > 0.0) all_c2.push_back(dist / min_distance);
+    }
+    if (all_c2.size() != NUM_DOCS) return;
+    std::sort(all_c2.begin(), all_c2.end());
+    for (uint32_t idx : { 1, 3, 10, 30, 100, 300, 1000, 3000, NUM_DOCS/2, NUM_DOCS-1}) {
+        fprintf(stderr, "c2-factor[%u] = %.3f\n", idx, all_c2[idx]);
+    }
+}
+
+using NNS_API = NNS<float>;
+
+TEST("require that brute force works") {
+    TimePoint bef = std::chrono::steady_clock::now();
+    fprintf(stderr, "generating %u brute force results\n", NUM_Q);
+    bruteforceResults.reserve(NUM_Q);
+    for (uint32_t cnt = 0; cnt < NUM_Q; ++cnt) {
+        const PointVector &query = generatedQueries[cnt];
+        bruteforceResults.emplace_back(bruteforce_nns(query));
+    }
+    TimePoint aft = std::chrono::steady_clock::now();
+    fprintf(stderr, "timing for brute force: %.3f ms = %.3f ms per query\n",
+            to_ms(aft - bef), to_ms(aft - bef)/NUM_Q);
+    for (int cnt = 0; cnt < NUM_Q; cnt = (cnt+1)*2) {
+        verifyBF(cnt);
+    }
+}
+
+#include "find-with-nns.h"
+#include "verify-top-k.h"
+
+void timing_nns(const char *name, NNS_API &nns, std::vector<uint32_t> sk_list) {
+    for (uint32_t search_k : sk_list) {
+        TimePoint bef = std::chrono::steady_clock::now();
+        for (int cnt = 0; cnt < NUM_Q; ++cnt) {
+            find_with_nns(search_k, nns, cnt);
+        }
+        TimePoint aft = std::chrono::steady_clock::now();
+        fprintf(stderr, "timing for %s search_k=%u: %.3f ms = %.3f ms/q\n",
+                name, search_k, to_ms(aft - bef), to_ms(aft - bef)/NUM_Q);
+    }
+}
+
+#include "quality-nns.h"
+
+template <typename FUNC>
+void bm_nns_simple(const char *name, FUNC creator, std::vector<uint32_t> sk_list) {
+    std::unique_ptr<NNS_API> nnsp = creator();
+    NNS_API &nns = *nnsp;
+    fprintf(stderr, "trying %s indexing...\n", name);
+    TimePoint bef = std::chrono::steady_clock::now();
+    for (uint32_t i = 0; i < NUM_DOCS; ++i) {
+        nns.addDoc(i);
+    }
+    TimePoint aft = std::chrono::steady_clock::now();
+    fprintf(stderr, "build %s index with %u docs: %.3f ms\n", name, NUM_DOCS, to_ms(aft - bef));
+    timing_nns(name, nns, sk_list);
+    fprintf(stderr, "Quality for %s [A] clean build with %u documents:\n", name, NUM_DOCS);
+    quality_nns(nns, sk_list);
+}
+
+template <typename FUNC>
+void benchmark_nns(const char *name, FUNC creator, std::vector<uint32_t> sk_list) {
+    bm_nns_simple(name, creator, sk_list);
+}
+
+#if 0
+TEST("require that Locality Sensitive Hashing mostly works") {
+    DocVectorAdapter adapter;
+    auto creator = [&adapter]() { return make_rplsh_nns(NUM_DIMS, adapter); };
+    benchmark_nns("RPLSH", creator, { 200, 1000 });
+}
+#endif
+
+#if 0
+TEST("require that Annoy via NNS api mostly works") {
+    DocVectorAdapter adapter;
+    auto creator = [&adapter]() { return make_annoy_nns(NUM_DIMS, adapter); };
+    benchmark_nns("Annoy", creator, { 8000, 10000 });
+}
+#endif
+
+#if 1
+TEST("require that HNSW via NNS api mostly works") {
+    DocVectorAdapter adapter;
+    auto creator = [&adapter]() { return make_hnsw_nns(NUM_DIMS, adapter); };
+    benchmark_nns("HNSW-like", creator, { 100, 150, 200 });
+}
+#endif
+
+#if 0
+TEST("require that HNSW wrapped api mostly works") {
+    DocVectorAdapter adapter;
+    auto creator = [&adapter]() { return make_hnsw_wrap(NUM_DIMS, adapter); };
+    benchmark_nns("HNSW-wrap", creator, { 100, 150, 200 });
+}
+#endif
+
+/**
+ * Before running the benchmark the ANN_GIST1M data set must be downloaded and extracted:
+ *   wget ftp://ftp.irisa.fr/local/texmex/corpus/gist.tar.gz
+ *   tar -xf gist.tar.gz
+ *
+ * The benchmark program will load the data set from $HOME/gist if no directory is specified.
+ *
+ * More information about the dataset is found here: http://corpus-texmex.irisa.fr/.
+ */
+int main(int argc, char **argv) {
+    TEST_MASTER.init(__FILE__);
+    std::string gist_dir = ".";
+    if (argc > 1) {
+        gist_dir = argv[1];
+    } else {
+        char *home = getenv("HOME");
+        if (home) {
+            gist_dir = home;
+            gist_dir += "/gist";
+        }
+    }
+    read_data(gist_dir);
+    TEST_RUN_ALL();
+    return (TEST_MASTER.fini() ? 0 : 1);
+}
diff --git a/eval/src/tests/ann/nns.h b/eval/src/tests/ann/nns.h
index ffe2882188e..ef3e4b5d69c 100644
--- a/eval/src/tests/ann/nns.h
+++ b/eval/src/tests/ann/nns.h
@@ -37,6 +37,31 @@ struct NnsHitComparatorLessDocid {
     }
 };
 
+class BitVector {
+private:
+    std::vector<uint64_t> _bits;
+public:
+    BitVector(size_t sz) : _bits((sz+63)/64) {}
+    BitVector& setBit(size_t idx) {
+        uint64_t mask = 1;
+        mask <<= (idx%64);
+        _bits[idx/64] |= mask;
+        return *this;
+    }
+    bool isSet(size_t idx) const {
+        uint64_t mask = 1;
+        mask <<= (idx%64);
+        uint64_t word = _bits[idx/64];
+        return (word & mask) != 0;
+    }
+    BitVector& clearBit(size_t idx) {
+        uint64_t mask = 1;
+        mask <<= (idx%64);
+        _bits[idx/64] &= ~mask;
+        return *this;
+    }
+};
+
 template <typename FltType = float>
 class NNS
 {
@@ -50,6 +75,7 @@ public:
 
     using Vector = vespalib::ConstArrayRef<FltType>;
     virtual std::vector<NnsHit> topK(uint32_t k, Vector vector, uint32_t search_k) = 0;
+    virtual std::vector<NnsHit> topKfilter(uint32_t k, Vector vector, uint32_t search_k, const BitVector &blacklist) = 0;
     virtual ~NNS() {}
 protected:
     uint32_t _numDims;
diff --git a/eval/src/tests/ann/quality-nns.h b/eval/src/tests/ann/quality-nns.h
new file mode 100644
index 00000000000..9ac37f0ef04
--- /dev/null
+++ b/eval/src/tests/ann/quality-nns.h
@@ -0,0 +1,42 @@
+// Copyright 2020 Oath Inc. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+
+bool reach_with_nns_k(NNS_API &nns, uint32_t docid, uint32_t k) {
+    const PointVector &qv = generatedDocs[docid];
+    vespalib::ConstArrayRef<float> query(qv.v, NUM_DIMS);
+    auto rv = nns.topK(k, query, k);
+    if (rv.size() != k) {
+        fprintf(stderr, "Result/K=%u from query for %u is %zu hits\n",
+                k, docid, rv.size());
+        return false;
+    }
+    if (rv[0].docid != docid) {
+      if (rv[0].sq.distance != 0.0)
+        fprintf(stderr, "Expected/K=%u to find %u but got %u with sq distance %.3f\n",
+                k, docid, rv[0].docid, rv[0].sq.distance);
+    }
+    return (rv[0].docid == docid || rv[0].sq.distance == 0.0);
+}
+
+void quality_nns(NNS_API &nns, std::vector<uint32_t> sk_list) {
+    for (uint32_t search_k : sk_list) {
+        double sum_recall = 0;
+        for (int cnt = 0; cnt < NUM_Q; ++cnt) {
+            sum_recall += verify_nns_quality(search_k, nns, cnt);
+        }
+        fprintf(stderr, "Overall average recall: %.2f\n", sum_recall / NUM_Q);
+    }
+    for (uint32_t search_k : { 1, 10, 100, 1000 }) {
+        TimePoint bef = std::chrono::steady_clock::now();
+        uint32_t reached = 0;
+        for (uint32_t i = 0; i < NUM_REACH; ++i) {
+            uint32_t target = i * (NUM_DOCS / NUM_REACH);
+            if (reach_with_nns_k(nns, target, search_k)) ++reached;
+        }
+        fprintf(stderr, "Could reach %u of %u documents with k=%u\n",
+                reached, NUM_REACH, search_k);
+        TimePoint aft = std::chrono::steady_clock::now();
+        fprintf(stderr, "reach time k=%u: %.3f ms = %.3f ms/q\n",
+                search_k, to_ms(aft - bef), to_ms(aft - bef)/NUM_REACH);
+        if (reached == NUM_REACH) break;
+    }
+}
diff --git a/eval/src/tests/ann/remove-bm.cpp b/eval/src/tests/ann/remove-bm.cpp
index be010552ab8..005f3804af9 100644
--- a/eval/src/tests/ann/remove-bm.cpp
+++ b/eval/src/tests/ann/remove-bm.cpp
@@ -13,6 +13,7 @@
 #define NUM_DOCS 250000
 #define NUM_DOCS_REMOVE 50000
 #define EFFECTIVE_DOCS (NUM_DOCS - NUM_DOCS_REMOVE)
+#define NUM_REACH 10000
 #define NUM_Q 1000
 
 #include "doc_vector_access.h"
@@ -30,10 +31,10 @@ struct PointVector {
 };
 
 static PointVector *aligned_alloc(size_t num) {
-    size_t sz = num * sizeof(PointVector);
-    double mega_bytes = sz / (1024.0*1024.0);
+    size_t num_bytes = num * sizeof(PointVector);
+    double mega_bytes = num_bytes / (1024.0*1024.0);
     fprintf(stderr, "allocate %.2f MB of vectors\n", mega_bytes);
-    char *mem = (char *)malloc(sz + 512);
+    char *mem = (char *)malloc(num_bytes + 512);
     mem += 512;
     size_t val = (size_t)mem;
     size_t unalign = val % 512;
@@ -221,83 +222,8 @@ TEST("require that brute force works") {
     }
 }
 
-bool reach_with_nns_1(NNS_API &nns, uint32_t docid) {
-    const PointVector &qv = generatedDocs[docid];
-    vespalib::ConstArrayRef<float> query(qv.v, NUM_DIMS);
-    auto rv = nns.topK(1, query, 1);
-    if (rv.size() != 1) {
-        fprintf(stderr, "Result/A from query for %u is %zu hits\n", docid, rv.size());
-        return false;
-    }
-    if (rv[0].docid != docid) {
-      if (rv[0].sq.distance != 0.0)
-        fprintf(stderr, "Expected/A to find %u but got %u with sq distance %.3f\n",
-                docid, rv[0].docid, rv[0].sq.distance);
-    }
-    return (rv[0].docid == docid || rv[0].sq.distance == 0.0);
-}
-
-bool reach_with_nns_100(NNS_API &nns, uint32_t docid) {
-    const PointVector &qv = generatedDocs[docid];
-    vespalib::ConstArrayRef<float> query(qv.v, NUM_DIMS);
-    auto rv = nns.topK(10, query, 100);
-    if (rv.size() != 10) {
-        fprintf(stderr, "Result/B from query for %u is %zu hits\n", docid, rv.size());
-    }
-    if (rv[0].docid != docid) {
-      if (rv[0].sq.distance != 0.0)
-        fprintf(stderr, "Expected/B to find %u but got %u with sq distance %.3f\n",
-                docid, rv[0].docid, rv[0].sq.distance);
-    }
-    return (rv[0].docid == docid || rv[0].sq.distance == 0.0);
-}
-
-bool reach_with_nns_1k(NNS_API &nns, uint32_t docid) {
-    const PointVector &qv = generatedDocs[docid];
-    vespalib::ConstArrayRef<float> query(qv.v, NUM_DIMS);
-    auto rv = nns.topK(10, query, 1000);
-    if (rv.size() != 10) {
-        fprintf(stderr, "Result/C from query for %u is %zu hits\n", docid, rv.size());
-    }
-    if (rv[0].docid != docid) {
-      if (rv[0].sq.distance != 0.0)
-        fprintf(stderr, "Expected/C to find %u but got %u with sq distance %.3f\n",
-                docid, rv[0].docid, rv[0].sq.distance);
-    }
-    return (rv[0].docid == docid || rv[0].sq.distance == 0.0);
-}
-
-TopK find_with_nns(uint32_t sk, NNS_API &nns, uint32_t qid) {
-    TopK result;
-    const PointVector &qv = generatedQueries[qid];
-    vespalib::ConstArrayRef<float> query(qv.v, NUM_DIMS);
-    auto rv = nns.topK(result.K, query, sk);
-    for (size_t i = 0; i < result.K; ++i) {
-        result.hits[i] = Hit(rv[i].docid, rv[i].sq.distance);
-    }
-    return result;
-}
-
-void verify_nns_quality(uint32_t sk, NNS_API &nns, uint32_t qid) {
-    TopK perfect = bruteforceResults[qid];
-    TopK result = find_with_nns(sk, nns, qid);
-    int recall = perfect.recall(result);
-    EXPECT_TRUE(recall > 40);
-    double sum_error = 0.0;
-    double c_factor = 1.0;
-    for (size_t i = 0; i < result.K; ++i) {
-        double factor = (result.hits[i].distance / perfect.hits[i].distance);
-        if (factor < 0.99 || factor > 25) {
-            fprintf(stderr, "hit[%zu] got distance %.3f, expected %.3f\n",
-                    i, result.hits[i].distance, perfect.hits[i].distance);
-        }
-        sum_error += factor;
-        c_factor = std::max(c_factor, factor);
-    }
-    EXPECT_TRUE(c_factor < 1.5);
-    fprintf(stderr, "quality sk=%u: query %u: recall %d  c2-factor %.3f  avg c2: %.3f\n",
-            sk, qid, recall, c_factor, sum_error / result.K);
-}
+#include "find-with-nns.h"
+#include "verify-top-k.h"
 
 void timing_nns(const char *name, NNS_API &nns, std::vector<uint32_t> sk_list) {
     for (uint32_t search_k : sk_list) {
@@ -311,64 +237,22 @@ void timing_nns(const char *name, NNS_API &nns, std::vector<uint32_t> sk_list) {
     }
 }
 
-void quality_nns(NNS_API &nns, std::vector<uint32_t> sk_list) {
-    for (uint32_t search_k : sk_list) {
-        for (int cnt = 0; cnt < NUM_Q; ++cnt) {
-            verify_nns_quality(search_k, nns, cnt);
-        }
-    }
-    uint32_t reached = 0;
-    for (uint32_t i = 0; i < 20000; ++i) {
-        if (reach_with_nns_1(nns, i)) ++reached;
-    }
-    fprintf(stderr, "Could reach %u of 20000 first documents with k=1\n", reached);
-    reached = 0;
-    for (uint32_t i = 0; i < 20000; ++i) {
-        if (reach_with_nns_100(nns, i)) ++reached;
-    }
-    fprintf(stderr, "Could reach %u of 20000 first documents with k=100\n", reached);
-    reached = 0;
-    for (uint32_t i = 0; i < 20000; ++i) {
-        if (reach_with_nns_1k(nns, i)) ++reached;
-    }
-    fprintf(stderr, "Could reach %u of 20000 first documents with k=1000\n", reached);
-}
+#include "quality-nns.h"
 
-void benchmark_nns(const char *name, NNS_API &nns, std::vector<uint32_t> sk_list) {
+template <typename FUNC>
+void bm_nns_simple(const char *name, FUNC creator, std::vector<uint32_t> sk_list) {
+    std::unique_ptr<NNS_API> nnsp = creator();
+    NNS_API &nns = *nnsp;
     fprintf(stderr, "trying %s indexing...\n", name);
-
-#if 0
-    TimePoint bef = std::chrono::steady_clock::now();
-    for (uint32_t i = 0; i < NUM_DOCS_REMOVE; ++i) {
-        nns.addDoc(EFFECTIVE_DOCS + i);
-    }
-    for (uint32_t i = 0; i < EFFECTIVE_DOCS - NUM_DOCS_REMOVE; ++i) {
-        nns.addDoc(i);
-    }
-    for (uint32_t i = 0; i < NUM_DOCS_REMOVE; ++i) {
-        nns.removeDoc(EFFECTIVE_DOCS + i);
-        nns.addDoc(EFFECTIVE_DOCS - NUM_DOCS_REMOVE + i);
-    }
-    TimePoint aft = std::chrono::steady_clock::now();
-    fprintf(stderr, "build %s index with %u docs: %.3f ms\n", name, EFFECTIVE_DOCS, to_ms(aft - bef));
-
-    timing_nns(name, nns, sk_list);
-    fprintf(stderr, "Quality for %s realistic build with %u documents:\n", name, EFFECTIVE_DOCS);
-    quality_nns(nns, sk_list);
-#endif
-
-#if 1
     TimePoint bef = std::chrono::steady_clock::now();
     for (uint32_t i = 0; i < EFFECTIVE_DOCS; ++i) {
         nns.addDoc(i);
     }
     TimePoint aft = std::chrono::steady_clock::now();
     fprintf(stderr, "build %s index with %u docs: %.3f ms\n", name, EFFECTIVE_DOCS, to_ms(aft - bef));
-
     timing_nns(name, nns, sk_list);
-    fprintf(stderr, "Quality for %s clean build with %u documents:\n", name, EFFECTIVE_DOCS);
+    fprintf(stderr, "Quality for %s [A] clean build with %u documents:\n", name, EFFECTIVE_DOCS);
     quality_nns(nns, sk_list);
-
     bef = std::chrono::steady_clock::now();
     for (uint32_t i = 0; i < NUM_DOCS_REMOVE; ++i) {
         nns.addDoc(EFFECTIVE_DOCS + i);
@@ -379,111 +263,115 @@ void benchmark_nns(const char *name, NNS_API &nns, std::vector<uint32_t> sk_list
     aft = std::chrono::steady_clock::now();
     fprintf(stderr, "build %s index add then remove %u docs: %.3f ms\n",
             name, NUM_DOCS_REMOVE, to_ms(aft - bef));
-
     timing_nns(name, nns, sk_list);
-    fprintf(stderr, "Quality for %s remove-damaged build with %u documents:\n", name, EFFECTIVE_DOCS);
+    fprintf(stderr, "Quality for %s [B] remove-damaged build with %u documents:\n", name, EFFECTIVE_DOCS);
     quality_nns(nns, sk_list);
-#endif
+}
 
-#if 0
+template <typename FUNC>
+void bm_nns_remove_old(const char *name, FUNC creator, std::vector<uint32_t> sk_list) {
+    std::unique_ptr<NNS_API> nnsp = creator();
+    NNS_API &nns = *nnsp;
     TimePoint bef = std::chrono::steady_clock::now();
+    for (uint32_t i = 0; i < NUM_DOCS_REMOVE; ++i) {
+        nns.addDoc(EFFECTIVE_DOCS + i);
+    }
     for (uint32_t i = 0; i < EFFECTIVE_DOCS; ++i) {
         nns.addDoc(i);
     }
+    for (uint32_t i = 0; i < NUM_DOCS_REMOVE; ++i) {
+        nns.removeDoc(EFFECTIVE_DOCS + i);
+    }
     TimePoint aft = std::chrono::steady_clock::now();
     fprintf(stderr, "build %s index with %u docs: %.3f ms\n", name, EFFECTIVE_DOCS, to_ms(aft - bef));
-
     timing_nns(name, nns, sk_list);
-    fprintf(stderr, "Quality for %s clean build with %u documents:\n", name, EFFECTIVE_DOCS);
+    fprintf(stderr, "Quality for %s [C] remove-oldest build with %u documents:\n", name, EFFECTIVE_DOCS);
     quality_nns(nns, sk_list);
+}
 
-    bef = std::chrono::steady_clock::now();
-    for (uint32_t i = 0; i < EFFECTIVE_DOCS; ++i) {
-        nns.removeDoc(i);
-    }
-    aft = std::chrono::steady_clock::now();
-    fprintf(stderr, "build %s index removed %u docs: %.3f ms\n", name, EFFECTIVE_DOCS, to_ms(aft - bef));
-
-    const uint32_t addFirst = NUM_DOCS - (NUM_DOCS_REMOVE * 3);
-    const uint32_t addSecond = NUM_DOCS - (NUM_DOCS_REMOVE * 2);
-
-    bef = std::chrono::steady_clock::now();
-    for (uint32_t i = 0; i < addFirst; ++i) {
-        nns.addDoc(i);
-    }
-    aft = std::chrono::steady_clock::now();
-    fprintf(stderr, "build %s index with %u docs: %.3f ms\n", name, addFirst, to_ms(aft - bef));
-
-    bef = std::chrono::steady_clock::now();
+template <typename FUNC>
+void bm_nns_interleave(const char *name, FUNC creator, std::vector<uint32_t> sk_list) {
+    std::unique_ptr<NNS_API> nnsp = creator();
+    NNS_API &nns = *nnsp;
+    TimePoint bef = std::chrono::steady_clock::now();
     for (uint32_t i = 0; i < NUM_DOCS_REMOVE; ++i) {
         nns.addDoc(EFFECTIVE_DOCS + i);
-        nns.addDoc(addFirst + i);
     }
-    aft = std::chrono::steady_clock::now();
-    fprintf(stderr, "build %s index added %u docs: %.3f ms\n",
-            name, 2 * NUM_DOCS_REMOVE, to_ms(aft - bef));
-
-    bef = std::chrono::steady_clock::now();
+    for (uint32_t i = 0; i < EFFECTIVE_DOCS - NUM_DOCS_REMOVE; ++i) {
+        nns.addDoc(i);
+    }
     for (uint32_t i = 0; i < NUM_DOCS_REMOVE; ++i) {
         nns.removeDoc(EFFECTIVE_DOCS + i);
-        nns.addDoc(addSecond + i);
+        nns.addDoc(EFFECTIVE_DOCS - NUM_DOCS_REMOVE + i);
     }
-    aft = std::chrono::steady_clock::now();
-    fprintf(stderr, "build %s index added %u and removed %u docs: %.3f ms\n",
-            name, NUM_DOCS_REMOVE, NUM_DOCS_REMOVE, to_ms(aft - bef));
-
+    TimePoint aft = std::chrono::steady_clock::now();
+    fprintf(stderr, "build %s index with %u docs: %.3f ms\n", name, EFFECTIVE_DOCS, to_ms(aft - bef));
     timing_nns(name, nns, sk_list);
-    fprintf(stderr, "Quality for %s with %u documents some churn:\n", name, EFFECTIVE_DOCS);
+    fprintf(stderr, "Quality for %s [D] realistic build with %u documents:\n", name, EFFECTIVE_DOCS);
     quality_nns(nns, sk_list);
+}
 
-#endif
-
-#if 0
-    bef = std::chrono::steady_clock::now();
-    fprintf(stderr, "removing and adding %u documents...\n", EFFECTIVE_DOCS);
-    for (uint32_t i = 0; i < EFFECTIVE_DOCS; ++i) {
-        nns.removeDoc(i);
+template <typename FUNC>
+void bm_nns_remove_old_add_new(const char *name, FUNC creator, std::vector<uint32_t> sk_list) {
+    std::unique_ptr<NNS_API> nnsp = creator();
+    NNS_API &nns = *nnsp;
+    TimePoint bef = std::chrono::steady_clock::now();
+    for (uint32_t i = 0; i < NUM_DOCS_REMOVE; ++i) {
+        nns.addDoc(EFFECTIVE_DOCS + i);
+    }
+    for (uint32_t i = 0; i < EFFECTIVE_DOCS - NUM_DOCS_REMOVE; ++i) {
         nns.addDoc(i);
     }
-    aft = std::chrono::steady_clock::now();
-    fprintf(stderr, "build %s index rem/add %u docs: %.3f ms\n",
-            name, EFFECTIVE_DOCS, to_ms(aft - bef));
-
+    for (uint32_t i = 0; i < NUM_DOCS_REMOVE; ++i) {
+        nns.removeDoc(EFFECTIVE_DOCS + i);
+    }
+    for (uint32_t i = 0; i < NUM_DOCS_REMOVE; ++i) {
+        nns.addDoc(EFFECTIVE_DOCS - NUM_DOCS_REMOVE + i);
+    }
+    TimePoint aft = std::chrono::steady_clock::now();
+    fprintf(stderr, "build %s index with %u docs: %.3f ms\n", name, EFFECTIVE_DOCS, to_ms(aft - bef));
     timing_nns(name, nns, sk_list);
-    fprintf(stderr, "Quality for %s with %u documents full churn:\n", name, EFFECTIVE_DOCS);
+    fprintf(stderr, "Quality for %s [E] remove old, add new build with %u documents:\n", name, EFFECTIVE_DOCS);
     quality_nns(nns, sk_list);
-#endif
+}
+
+template <typename FUNC>
+void benchmark_nns(const char *name, FUNC creator, std::vector<uint32_t> sk_list) {
+    bm_nns_simple(name, creator, sk_list);
+    bm_nns_remove_old(name, creator, sk_list);
+    bm_nns_interleave(name, creator, sk_list);
+    bm_nns_remove_old_add_new(name, creator, sk_list);
 }
 
 #if 0
 TEST("require that Locality Sensitive Hashing mostly works") {
     DocVectorAdapter adapter;
-    std::unique_ptr<NNS_API> nns = make_rplsh_nns(NUM_DIMS, adapter);
-    benchmark_nns("RPLSH", *nns, { 200, 1000 });
+    auto creator = [&adapter]() { return make_rplsh_nns(NUM_DIMS, adapter); };
+    benchmark_nns("RPLSH", creator, { 200, 1000 });
 }
 #endif
 
 #if 0
 TEST("require that Annoy via NNS api mostly works") {
     DocVectorAdapter adapter;
-    std::unique_ptr<NNS_API> nns = make_annoy_nns(NUM_DIMS, adapter);
-    benchmark_nns("Annoy", *nns, { 8000, 10000 });
+    auto creator = [&adapter]() { return make_annoy_nns(NUM_DIMS, adapter); };
+    benchmark_nns("Annoy", creator, { 8000, 10000 });
 }
 #endif
 
 #if 1
 TEST("require that HNSW via NNS api mostly works") {
     DocVectorAdapter adapter;
-    std::unique_ptr<NNS_API> nns = make_hnsw_nns(NUM_DIMS, adapter);
-    benchmark_nns("HNSW-like", *nns, { 100, 150, 200 });
+    auto creator = [&adapter]() { return make_hnsw_nns(NUM_DIMS, adapter); };
+    benchmark_nns("HNSW-like", creator, { 100, 150, 200 });
 }
 #endif
 
 #if 0
 TEST("require that HNSW wrapped api mostly works") {
     DocVectorAdapter adapter;
-    std::unique_ptr<NNS_API> nns = make_hnsw_wrap(NUM_DIMS, adapter);
-    benchmark_nns("HNSW-wrap", *nns, { 100, 150, 200 });
+    auto creator = [&adapter]() { return make_hnsw_wrap(NUM_DIMS, adapter); };
+    benchmark_nns("HNSW-wrap", creator, { 100, 150, 200 });
 }
 #endif
 
diff --git a/eval/src/tests/ann/sift_benchmark.cpp b/eval/src/tests/ann/sift_benchmark.cpp
index 022c9404f5d..5f3c16e127d 100644
--- a/eval/src/tests/ann/sift_benchmark.cpp
+++ b/eval/src/tests/ann/sift_benchmark.cpp
@@ -13,14 +13,15 @@
 #define NUM_DIMS 128
 #define NUM_DOCS 1000000
 #define NUM_Q 1000
+#define NUM_REACH 10000
 
 #include "doc_vector_access.h"
 #include "nns.h"
 #include "for-sift-hit.h"
 #include "for-sift-top-k.h"
+#include "std-random.h"
 
 std::vector<TopK> bruteforceResults;
-std::vector<float> tmp_v(NUM_DIMS);
 
 struct PointVector {
     float v[NUM_DIMS];
@@ -29,10 +30,10 @@ struct PointVector {
 };
 
 static PointVector *aligned_alloc(size_t num) {
-    size_t sz = num * sizeof(PointVector);
-    double mega_bytes = sz / (1024.0*1024.0);
+    size_t num_bytes = num * sizeof(PointVector);
+    double mega_bytes = num_bytes / (1024.0*1024.0);
     fprintf(stderr, "allocate %.2f MB of vectors\n", mega_bytes);
-    char *mem = (char *)malloc(sz + 512);
+    char *mem = (char *)malloc(num_bytes + 512);
     mem += 512;
     size_t val = (size_t)mem;
     size_t unalign = val % 512;
@@ -53,7 +54,7 @@ struct DocVectorAdapter : public DocVectorAccess<float>
 
 double computeDistance(const PointVector &query, uint32_t docid) {
     const PointVector &docvector = generatedDocs[docid];
-    return l2distCalc.l2sq_dist(query, docvector, tmp_v);
+    return l2distCalc.l2sq_dist(query, docvector);
 }
 
 void read_queries(std::string fn) {
@@ -151,7 +152,7 @@ TopK bruteforce_nns(const PointVector &query) {
     BfHitHeap heap(result.K);
     for (uint32_t docid = 0; docid < NUM_DOCS; ++docid) {
         const PointVector &docvector = generatedDocs[docid];
-        double d = l2distCalc.l2sq_dist(query, docvector, tmp_v);
+        double d = l2distCalc.l2sq_dist(query, docvector);
         Hit h(docid, d);
         heap.maybe_use(h);
     }
@@ -162,24 +163,58 @@ TopK bruteforce_nns(const PointVector &query) {
     return result;
 }
 
+TopK bruteforce_nns_filter(const PointVector &query, const BitVector &blacklist) {
+    TopK result;
+    BfHitHeap heap(result.K);
+    for (uint32_t docid = 0; docid < NUM_DOCS; ++docid) {
+        if (blacklist.isSet(docid)) continue;
+        const PointVector &docvector = generatedDocs[docid];
+        double d = l2distCalc.l2sq_dist(query, docvector);
+        Hit h(docid, d);
+        heap.maybe_use(h);
+    }
+    std::vector<Hit> best = heap.bestHits();
+    EXPECT_EQUAL(best.size(), result.K);
+    for (size_t i = 0; i < result.K; ++i) {
+        result.hits[i] = best[i];
+    }
+    return result;
+}
+
+
 void verifyBF(uint32_t qid) {
     const PointVector &query = generatedQueries[qid];
     TopK &result = bruteforceResults[qid];
     double min_distance = result.hits[0].distance;
-    std::vector<double> all_c2;
     for (uint32_t i = 0; i < NUM_DOCS; ++i) {
         double dist = computeDistance(query, i);
         if (dist < min_distance) {
             fprintf(stderr, "WARN dist %.9g < mindist %.9g\n", dist, min_distance);
         }
         EXPECT_FALSE(dist+0.000001 < min_distance);
-        if (min_distance > 0) all_c2.push_back(dist / min_distance);
     }
-    if (all_c2.size() != NUM_DOCS) return;
-    std::sort(all_c2.begin(), all_c2.end());
-    for (uint32_t idx : { 1, 3, 10, 30, 100, 300, 1000, 3000, NUM_DOCS/2, NUM_DOCS-1}) {
-        fprintf(stderr, "c2-factor[%u] = %.3f\n", idx, all_c2[idx]);
+}
+
+void timing_bf_filter(int percent)
+{
+    BitVector blacklist(NUM_DOCS);
+    RndGen rnd;
+    for (uint32_t idx = 0; idx < NUM_DOCS; ++idx) {
+        if (rnd.nextUniform() < 0.01 * percent) {
+            blacklist.setBit(idx);
+        } else {
+            blacklist.clearBit(idx);
+        }
+    }
+    TimePoint bef = std::chrono::steady_clock::now();
+    for (int cnt = 0; cnt < NUM_Q; ++cnt) {
+        const PointVector &qv = generatedQueries[cnt];
+        auto res = bruteforce_nns_filter(qv, blacklist);
+        EXPECT_TRUE(res.hits[res.K - 1].distance > 0.0);
     }
+    TimePoint aft = std::chrono::steady_clock::now();
+    fprintf(stderr, "timing for bruteforce filter %d %%: %.3f ms = %.3f ms/q\n",
+            percent, to_ms(aft - bef), to_ms(aft - bef)/NUM_Q);
 }
 
 TEST("require that brute force works") {
@@ -195,52 +230,90 @@ TEST("require that brute force works") {
     for (int cnt = 0; cnt < NUM_Q; cnt = (cnt+1)*2) {
         verifyBF(cnt);
     }
+#if 1
+    for (uint32_t filter_percent : { 0, 1, 10, 50, 90, 95, 99 }) {
+        timing_bf_filter(filter_percent);
+    }
+#endif
 }
 
 using NNS_API = NNS<float>;
 
-TopK find_with_nns(uint32_t sk, NNS_API &nns, uint32_t qid) {
-    TopK result;
+size_t search_with_filter(uint32_t sk, NNS_API &nns, uint32_t qid,
+                          const BitVector &blacklist)
+{
     const PointVector &qv = generatedQueries[qid];
     vespalib::ConstArrayRef<float> query(qv.v, NUM_DIMS);
-    auto rv = nns.topK(result.K, query, sk);
-    for (size_t i = 0; i < result.K; ++i) {
-        result.hits[i] = Hit(rv[i].docid, rv[i].sq.distance);
+    auto rv = nns.topKfilter(100, query, sk, blacklist);
+    return rv.size();
+}
+
+#include "find-with-nns.h"
+#include "verify-top-k.h"
+
+void verify_with_filter(uint32_t sk, NNS_API &nns, uint32_t qid,
+                        const BitVector &blacklist)
+{
+    const PointVector &qv = generatedQueries[qid];
+    auto expected = bruteforce_nns_filter(qv, blacklist);
+    vespalib::ConstArrayRef<float> query(qv.v, NUM_DIMS);
+    auto rv = nns.topKfilter(expected.K, query, sk, blacklist);
+    TopK actual;
+    for (size_t i = 0; i < actual.K; ++i) {
+        actual.hits[i] = Hit(rv[i].docid, rv[i].sq.distance);
     }
-    return result;
+    verify_top_k(expected, actual, sk, qid);
 }
 
-void verify_nns_quality(uint32_t sk, NNS_API &nns, uint32_t qid) {
-    TopK perfect = bruteforceResults[qid];
-    TopK result = find_with_nns(sk, nns, qid);
-    int recall = perfect.recall(result);
-    EXPECT_TRUE(recall > 40);
-    double sum_error = 0.0;
-    double c_factor = 1.0;
-    for (size_t i = 0; i < result.K; ++i) {
-        double factor = (result.hits[i].distance / perfect.hits[i].distance);
-        if (factor < 0.99 || factor > 25) {
-            fprintf(stderr, "hit[%zu] got distance %.3f, expected %.3f\n",
-                    i, result.hits[i].distance, perfect.hits[i].distance);
+void timing_nns_filter(const char *name, NNS_API &nns,
+                       std::vector<uint32_t> sk_list, int percent)
+{
+    BitVector blacklist(NUM_DOCS);
+    RndGen rnd;
+    for (uint32_t idx = 0; idx < NUM_DOCS; ++idx) {
+        if (rnd.nextUniform() < 0.01 * percent) {
+            blacklist.setBit(idx);
+        } else {
+            blacklist.clearBit(idx);
+        }
+    }
+    for (uint32_t search_k : sk_list) {
+        TimePoint bef = std::chrono::steady_clock::now();
+        for (int cnt = 0; cnt < NUM_Q; ++cnt) {
+            uint32_t nh = search_with_filter(search_k, nns, cnt, blacklist);
+            EXPECT_EQUAL(nh, 100u);
+        }
+        TimePoint aft = std::chrono::steady_clock::now();
+        fprintf(stderr, "timing for %s filter %d %% search_k=%u: %.3f ms = %.3f ms/q\n",
+                name, percent, search_k, to_ms(aft - bef), to_ms(aft - bef)/NUM_Q);
+#if 0
+        fprintf(stderr, "Quality check for %s filter %d %%:\n", name, percent);
+        for (int cnt = 0; cnt < NUM_Q; ++cnt) {
+            verify_with_filter(search_k, nns, cnt, blacklist);
         }
-        sum_error += factor;
-        c_factor = std::max(c_factor, factor);
+#endif
     }
-    EXPECT_TRUE(c_factor < 1.5);
-    fprintf(stderr, "quality sk=%u: query %u: recall %d, c2-factor %.3f, avg c2: %.3f\n",
-            sk, qid, recall, c_factor, sum_error / result.K);
-    if (qid == 6) {
-        for (size_t i = 0; i < 10; ++i) {
-            fprintf(stderr, "topk[%zu] BF{%u %.3f} index{%u %.3f}\n",
-                    i,
-                    perfect.hits[i].docid, perfect.hits[i].distance,
-                    result.hits[i].docid, result.hits[i].distance);
+}
+
+void timing_nns(const char *name, NNS_API &nns, std::vector<uint32_t> sk_list) {
+    for (uint32_t search_k : sk_list) {
+        TimePoint bef = std::chrono::steady_clock::now();
+        for (int cnt = 0; cnt < NUM_Q; ++cnt) {
+            find_with_nns(search_k, nns, cnt);
         }
+        TimePoint aft = std::chrono::steady_clock::now();
+        fprintf(stderr, "timing for %s search_k=%u: %.3f ms = %.3f ms/q\n",
+                name, search_k, to_ms(aft - bef), to_ms(aft - bef)/NUM_Q);
     }
 }
 
-void benchmark_nns(const char *name, NNS_API &nns, std::vector<uint32_t> sk_list) {
+#include "quality-nns.h"
+
+template <typename FUNC>
+void benchmark_nns(const char *name, FUNC creator, std::vector<uint32_t> sk_list) {
     fprintf(stderr, "trying %s indexing...\n", name);
+    std::unique_ptr<NNS_API> nnsp = creator();
+    NNS_API &nns = *nnsp;
     TimePoint bef = std::chrono::steady_clock::now();
     for (uint32_t i = 0; i < NUM_DOCS; ++i) {
         nns.addDoc(i);
@@ -250,50 +323,44 @@ void benchmark_nns(const char *name, NNS_API &nns, std::vector<uint32_t> sk_list
     TimePoint aft = std::chrono::steady_clock::now();
     fprintf(stderr, "build %s index: %.3f ms\n", name, to_ms(aft - bef));
 
-    for (uint32_t search_k : sk_list) {
-        bef = std::chrono::steady_clock::now();
-        for (int cnt = 0; cnt < NUM_Q; ++cnt) {
-            find_with_nns(search_k, nns, cnt);
-        }
-        aft = std::chrono::steady_clock::now();
-        fprintf(stderr, "timing for %s search_k=%u: %.3f ms = %.3f ms/q\n",
-                name, search_k, to_ms(aft - bef), to_ms(aft - bef)/NUM_Q);
-        for (int cnt = 0; cnt < NUM_Q; ++cnt) {
-            verify_nns_quality(search_k, nns, cnt);
-        }
+    fprintf(stderr, "Timings for %s :\n", name);
+    timing_nns(name, nns, sk_list);
+    for (uint32_t filter_percent : { 0, 1, 10, 50, 90, 95, 99 }) {
+        timing_nns_filter(name, nns, sk_list, filter_percent);
     }
+    fprintf(stderr, "Quality for %s :\n", name);
+    quality_nns(nns, sk_list);
 }
 
-
-#if 1
+#if 0
 TEST("require that Locality Sensitive Hashing mostly works") {
     DocVectorAdapter adapter;
-    std::unique_ptr<NNS_API> nns = make_rplsh_nns(NUM_DIMS, adapter);
-    benchmark_nns("RPLSH", *nns, { 200, 1000 });
+    auto creator = [&adapter]() { return make_rplsh_nns(NUM_DIMS, adapter); };
+    benchmark_nns("RPLSH", creator, { 200, 1000 });
 }
 #endif
 
 #if 1
 TEST("require that Annoy via NNS api mostly works") {
     DocVectorAdapter adapter;
-    std::unique_ptr<NNS_API> nns = make_annoy_nns(NUM_DIMS, adapter);
-    benchmark_nns("Annoy", *nns, { 8000, 10000 });
+    auto creator = [&adapter]() { return make_annoy_nns(NUM_DIMS, adapter); };
+    benchmark_nns("Annoy", creator, { 8000, 10000 });
 }
 #endif
 
 #if 1
 TEST("require that HNSW via NNS api mostly works") {
     DocVectorAdapter adapter;
-    std::unique_ptr<NNS_API> nns = make_hnsw_nns(NUM_DIMS, adapter);
-    benchmark_nns("HNSW-like", *nns, { 100, 150, 200 });
+    auto creator = [&adapter]() { return make_hnsw_nns(NUM_DIMS, adapter); };
+    benchmark_nns("HNSW-like", creator, { 100, 150, 200 });
 }
 #endif
 
 #if 0
 TEST("require that HNSW wrapped api mostly works") {
     DocVectorAdapter adapter;
-    std::unique_ptr<NNS_API> nns = make_hnsw_wrap(NUM_DIMS, adapter);
-    benchmark_nns("HNSW-wrap", *nns, { 100, 150, 200 });
+    auto creator = [&adapter]() { return make_hnsw_wrap(NUM_DIMS, adapter); };
+    benchmark_nns("HNSW-wrap", creator, { 100, 150, 200 });
 }
 #endif
 
diff --git a/eval/src/tests/ann/verify-top-k.h b/eval/src/tests/ann/verify-top-k.h
new file mode 100644
index 00000000000..220c273d017
--- /dev/null
+++ b/eval/src/tests/ann/verify-top-k.h
@@ -0,0 +1,27 @@
+// Copyright 2020 Oath Inc. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
+
+int verify_top_k(const TopK &perfect, const TopK &result, uint32_t sk, uint32_t qid) {
+    int recall = perfect.recall(result);
+    EXPECT_TRUE(recall > 40);
+    double sum_error = 0.0;
+    double c_factor = 1.0;
+    for (size_t i = 0; i < result.K; ++i) {
+        double factor = (result.hits[i].distance / perfect.hits[i].distance);
+        if (factor < 0.99 || factor > 25) {
+            fprintf(stderr, "hit[%zu] got distance %.3f, expected %.3f\n",
+                    i, result.hits[i].distance, perfect.hits[i].distance);
+        }
+        sum_error += factor;
+        c_factor = std::max(c_factor, factor);
+    }
+    EXPECT_TRUE(c_factor < 1.5);
+    fprintf(stderr, "quality sk=%u: query %u: recall %d  c2-factor %.3f  avg c2: %.3f\n",
+            sk, qid, recall, c_factor, sum_error / result.K);
+    return recall;
+}
+
+int verify_nns_quality(uint32_t sk, NNS_API &nns, uint32_t qid) {
+    TopK perfect = bruteforceResults[qid];
+    TopK result = find_with_nns(sk, nns, qid);
+    return verify_top_k(perfect, result, sk, qid);
+}
diff --git a/eval/src/tests/ann/xp-annoy-nns.cpp b/eval/src/tests/ann/xp-annoy-nns.cpp
index f022aae5974..213e583d95a 100644
--- a/eval/src/tests/ann/xp-annoy-nns.cpp
+++ b/eval/src/tests/ann/xp-annoy-nns.cpp
@@ -27,6 +27,7 @@ struct Node {
     virtual Node *addDoc(uint32_t docid, V vector, AnnoyLikeNns &meta) = 0;
     virtual int remove(uint32_t docid, V vector) = 0;
     virtual void findCandidates(std::set<uint32_t> &cands, V vector, NodeQueue &queue, double minDist) const = 0;
+    virtual void filterCandidates(std::set<uint32_t> &cands, V vector, NodeQueue &queue, double minDist, const BitVector &blacklist) const = 0;
     virtual void stats(std::vector<uint32_t> &depths) = 0;
 };
 
@@ -38,6 +39,7 @@ struct LeafNode : public Node {
     Node *addDoc(uint32_t docid, V vector, AnnoyLikeNns &meta) override;
     int remove(uint32_t docid, V vector) override;
     void findCandidates(std::set<uint32_t> &cands, V vector, NodeQueue &queue, double minDist) const override;
+    void filterCandidates(std::set<uint32_t> &cands, V vector, NodeQueue &queue, double minDist, const BitVector &blacklist) const override;
 
     Node *split(AnnoyLikeNns &meta);
     virtual void stats(std::vector<uint32_t> &depths) override { depths.push_back(1); }
@@ -55,6 +57,7 @@ struct SplitNode : public Node {
     Node *addDoc(uint32_t docid, V vector, AnnoyLikeNns &meta) override;
     int remove(uint32_t docid, V vector) override;
     void findCandidates(std::set<uint32_t> &cands, V vector, NodeQueue &queue, double minDist) const override;
+    void filterCandidates(std::set<uint32_t> &cands, V vector, NodeQueue &queue, double minDist, const BitVector &blacklist) const override;
 
     double planeDistance(V vector) const;
     virtual void stats(std::vector<uint32_t> &depths) override {
@@ -106,6 +109,8 @@ public:
     }
     std::vector<NnsHit> topK(uint32_t k, Vector vector, uint32_t search_k) override;
 
+    std::vector<NnsHit> topKfilter(uint32_t k, Vector vector, uint32_t search_k, const BitVector &bitvector) override;
+
     V getVector(uint32_t docid) const { return _dva.get(docid); }
     double uniformRnd() { return _rndGen.nextUniform(); } 
     uint32_t dims() const { return _numDims; }
@@ -304,6 +309,16 @@ LeafNode::findCandidates(std::set<uint32_t> &cands, V, NodeQueue &, double) cons
     }
 }
 
+void
+LeafNode::filterCandidates(std::set<uint32_t> &cands, V, NodeQueue &, double, const BitVector &blacklist) const
+{
+    for (uint32_t d : docids) {
+        if (blacklist.isSet(d)) continue;
+        cands.insert(d);
+    }
+}
+
+
 SplitNode::~SplitNode()
 {
     delete leftChildren;
@@ -344,6 +359,15 @@ SplitNode::findCandidates(std::set<uint32_t> &, V vector, NodeQueue &queue, doub
     queue.push(std::make_pair(std::min(d, minDist), rightChildren));
 }
 
+void
+SplitNode::filterCandidates(std::set<uint32_t> &, V vector, NodeQueue &queue, double minDist, const BitVector &) const
+{
+    double d = planeDistance(vector);
+    // fprintf(stderr, "push 2 nodes dist %g\n", d);
+    queue.push(std::make_pair(std::min(-d, minDist), leftChildren));
+    queue.push(std::make_pair(std::min(d, minDist), rightChildren));
+}
+
 std::vector<NnsHit>
 AnnoyLikeNns::topK(uint32_t k, Vector vector, uint32_t search_k)
 {
@@ -387,6 +411,40 @@ AnnoyLikeNns::topK(uint32_t k, Vector vector, uint32_t search_k)
     return r;
 }
 
+std::vector<NnsHit>
+AnnoyLikeNns::topKfilter(uint32_t k, Vector vector, uint32_t search_k, const BitVector &blacklist)
+{
+    ++find_top_k_cnt;
+    std::vector<NnsHit> r;
+    r.reserve(k);
+    std::set<uint32_t> candidates;
+    NodeQueue queue;
+    for (Node *root : _roots) {
+        double dist = std::numeric_limits<double>::max();
+        queue.push(std::make_pair(dist, root));
+    }
+    while ((candidates.size() < std::max(k, search_k)) && (queue.size() > 0)) {
+        const QueueNode& top = queue.top();
+        double md = top.first;
+        // fprintf(stderr, "find candidates: node with min distance %g\n", md);
+        Node *n = top.second;
+        queue.pop();
+        n->filterCandidates(candidates, vector, queue, md, blacklist);
+        ++find_cand_cnt;
+    }
+    for (uint32_t docid : candidates) {
+        if (blacklist.isSet(docid)) continue;
+        double dist = l2distCalc.l2sq_dist(vector, _dva.get(docid));
+        NnsHit hit(docid, SqDist(dist));
+        r.push_back(hit);
+    }
+    std::sort(r.begin(), r.end(), NnsHitComparatorLessDistance());
+    while (r.size() > k) r.pop_back();
+    return r;
+}
+
+
+
 void
 AnnoyLikeNns::dumpStats() {
     fprintf(stderr, "stats for AnnoyLikeNns:\n");
diff --git a/eval/src/tests/ann/xp-hnsw-wrap.cpp b/eval/src/tests/ann/xp-hnsw-wrap.cpp
index 3eb01142dcd..45c7a974254 100644
--- a/eval/src/tests/ann/xp-hnsw-wrap.cpp
+++ b/eval/src/tests/ann/xp-hnsw-wrap.cpp
@@ -46,6 +46,34 @@ public:
         }
         return result;
     }
+
+    std::vector<NnsHit> topKfilter(uint32_t k, Vector vector, uint32_t search_k, const BitVector &blacklist) override {
+        std::vector<NnsHit> reversed;
+        uint32_t adjusted_k = k+4;
+        uint32_t adjusted_sk = search_k+4;
+        for (int retry = 0; (retry < 5) && (reversed.size() < k); ++retry) {
+            reversed.clear();
+            _hnsw.setEf(adjusted_sk);
+            auto priQ = _hnsw.searchKnn(vector.cbegin(), adjusted_k);
+            while (! priQ.empty()) {
+                auto pair = priQ.top();
+                if (! blacklist.isSet(pair.second)) {
+                    reversed.emplace_back(pair.second, SqDist(pair.first));
+                }
+                priQ.pop();
+            }
+            double got = 1 + reversed.size();
+            double factor = 1.25 * k / got;
+            adjusted_k *= factor;
+            adjusted_sk *= factor;
+        }
+        std::vector<NnsHit> result;
+        while (result.size() < k && !reversed.empty()) {
+            result.push_back(reversed.back());
+            reversed.pop_back();
+        }
+        return result;
+    }
 };
 
 std::unique_ptr<NNS<float>>
diff --git a/eval/src/tests/ann/xp-hnswlike-nns.cpp b/eval/src/tests/ann/xp-hnswlike-nns.cpp
index 5cdbdd8efa3..90fc0fe2e92 100644
--- a/eval/src/tests/ann/xp-hnswlike-nns.cpp
+++ b/eval/src/tests/ann/xp-hnswlike-nns.cpp
@@ -32,6 +32,11 @@ static size_t distcalls_heuristic;
 static size_t distcalls_shrink;
 static size_t distcalls_refill;
 static size_t refill_needed_calls;
+static size_t shrink_needed_calls;
+static size_t disconnected_weak_links;
+static size_t disconnected_for_symmetry;
+static size_t select_n_full;
+static size_t select_n_partial;
 
 struct LinkList : std::vector<uint32_t>
 {
@@ -76,6 +81,7 @@ struct VisitedSet
         ptr = (Mark *)malloc(size * sizeof(Mark));
         curval = -1;
         sz = size;
+        clear();
     }
     void clear() {
         ++curval;
@@ -99,8 +105,9 @@ struct VisitedSetPool
     VisitedSet &get(size_t size) {
         if (size > lastUsed->sz) {
             lastUsed = std::make_unique<VisitedSet>(size*2);
+        } else {
+            lastUsed->clear();
         }
-        lastUsed->clear();
         return *lastUsed;
     }
 };
@@ -214,6 +221,10 @@ public:
     void search_layer(Vector vector, FurthestPriQ &w,
                       uint32_t ef, uint32_t searchLevel);
 
+    void search_layer_with_filter(Vector vector, FurthestPriQ &w,
+                                  uint32_t ef, uint32_t searchLevel,
+                                  const BitVector &blacklist);
+
     bool haveCloserDistance(HnswHit e, const LinkList &r) const {
         for (uint32_t prevId : r) {
             double dist = distance(e.docid, prevId);
@@ -278,6 +289,10 @@ public:
             fprintf(stderr, "distance calls for refill: %zu is %.3f per op\n", distcalls_refill, distcalls_refill / div);
             fprintf(stderr, "distance calls for other: %zu is %.3f per op\n", distcalls_other, distcalls_other / div);
             fprintf(stderr, "refill needed calls: %zu is %.3f per op\n", refill_needed_calls, refill_needed_calls / div);
+            fprintf(stderr, "shrink needed calls: %zu is %.3f per op\n", shrink_needed_calls, shrink_needed_calls / div);
+            fprintf(stderr, "disconnected weak links: %zu is %.3f per op\n", disconnected_weak_links, disconnected_weak_links / div);
+            fprintf(stderr, "disconnected for symmetry: %zu is %.3f per op\n", disconnected_for_symmetry, disconnected_for_symmetry / div);
+            fprintf(stderr, "select neighbors: partial %zu vs full %zu\n", select_n_partial, select_n_full);
         }
     }                    
 
@@ -315,10 +330,19 @@ public:
         LinkList lostLinks;
         LinkList oldLinks = links;
         links = remove_weakest(distances, maxLinks, lostLinks);
+#define KEEP_SYM
+#ifdef KEEP_SYM
         for (uint32_t lost_id : lostLinks) {
+            ++disconnected_for_symmetry;
             remove_link_from(lost_id, shrink_id, level);
+        }
+#define DO_REFILL_AFTER_KEEP_SYM
+#ifdef DO_REFILL_AFTER_KEEP_SYM
+        for (uint32_t lost_id : lostLinks) {
             refill_ifneeded(lost_id, oldLinks, level);
         }
+#endif
+#endif
     }
 
     void each_shrink_ifneeded(const LinkList &neighbors, uint32_t level);
@@ -337,7 +361,9 @@ public:
                 }
                 remove_link_from(n_id, docid, level);
             }
-            for (uint32_t n_id : my_links) {
+            while (! my_links.empty()) {
+                uint32_t n_id = my_links.back();
+                my_links.pop_back();
                 refill_ifneeded(n_id, my_links, level);
             }
         }
@@ -363,12 +389,12 @@ public:
         ++distcalls_other;
         HnswHit entryPoint(_entryId, SqDist(entryDist));
         int searchLevel = _entryLevel;
-        FurthestPriQ w;
-        w.push(entryPoint);
         while (searchLevel > 0) {
-            search_layer(vector, w, std::min(k, search_k), searchLevel);
+            entryPoint = search_layer_simple(vector, entryPoint, searchLevel);
             --searchLevel;
         }
+        FurthestPriQ w;
+        w.push(entryPoint);
         search_layer(vector, w, std::max(k, search_k), 0);
         while (w.size() > k) {
             w.pop();
@@ -381,8 +407,11 @@ public:
         }
         return result;
     }
+
+    std::vector<NnsHit> topKfilter(uint32_t k, Vector vector, uint32_t search_k, const BitVector &blacklist) override;
 };
 
+
 double
 HnswLikeNns::distance(Vector v, uint32_t b) const
 {
@@ -390,12 +419,40 @@ HnswLikeNns::distance(Vector v, uint32_t b) const
     return l2distCalc.l2sq_dist(v, w);
 }
 
+std::vector<NnsHit> 
+HnswLikeNns::topKfilter(uint32_t k, Vector vector, uint32_t search_k, const BitVector &blacklist)
+{
+    std::vector<NnsHit> result;
+    if (_entryLevel < 0) return result;
+    double entryDist = distance(vector, _entryId);
+    ++distcalls_other;
+    HnswHit entryPoint(_entryId, SqDist(entryDist));
+    int searchLevel = _entryLevel;
+    while (searchLevel > 0) {
+        entryPoint = search_layer_simple(vector, entryPoint, searchLevel);
+        --searchLevel;
+    }
+    FurthestPriQ w;
+    w.push(entryPoint);
+    search_layer_with_filter(vector, w, std::max(k, search_k), 0, blacklist);
+    NearestList tmp = w.steal();
+    std::sort(tmp.begin(), tmp.end(), LesserDist());
+    result.reserve(std::min((size_t)k, tmp.size()));
+    for (const auto & hit : tmp) {
+        if (blacklist.isSet(hit.docid)) continue;
+        result.emplace_back(hit.docid, SqDist(hit.dist));
+        if (result.size() == k) break;
+    }
+    return result;
+}
+
 void
 HnswLikeNns::each_shrink_ifneeded(const LinkList &neighbors, uint32_t level) {
     uint32_t maxLinks = (level > 0) ? _M : (2 * _M);
     for (uint32_t old_id : neighbors) {
         LinkList &oldLinks = getLinkList(old_id, level);
         if (oldLinks.size() > maxLinks) {
+            ++shrink_needed_calls;
             shrink_links(old_id, maxLinks, level);
         }
     }
@@ -437,6 +494,44 @@ HnswLikeNns::search_layer(Vector vector, FurthestPriQ &w,
     return;
 }
 
+void
+HnswLikeNns::search_layer_with_filter(Vector vector, FurthestPriQ &w,
+                                      uint32_t ef, uint32_t searchLevel,
+                                      const BitVector &blacklist)
+{
+    NearestPriQ candidates;
+    VisitedSet &visited = _visitedSetPool.get(_nodes.size());
+
+    for (const HnswHit & entry : w.peek()) {
+        candidates.push(entry);
+        visited.mark(entry.docid);
+        if (blacklist.isSet(entry.docid)) ++ef;
+    }
+    double limd = std::numeric_limits<double>::max();
+    while (! candidates.empty()) {
+        HnswHit cand = candidates.top();
+        if (cand.dist > limd) {
+            break;
+        }
+        candidates.pop();
+        for (uint32_t e_id : getLinkList(cand.docid, searchLevel)) {
+            if (visited.isMarked(e_id)) continue;
+            visited.mark(e_id);
+            double e_dist = distance(vector, e_id);
+            ++distcalls_search_layer;
+            if (e_dist < limd) {
+                candidates.emplace(e_id, SqDist(e_dist));
+                if (blacklist.isSet(e_id)) continue;
+                w.emplace(e_id, SqDist(e_dist));
+                if (w.size() > ef) {
+                    w.pop();
+                    limd = w.top().dist;
+                }
+            }
+        }
+    }
+}
+
 LinkList
 HnswLikeNns::remove_weakest(const NearestList &neighbors, uint32_t curMax, LinkList &lost) const
 {
@@ -458,13 +553,13 @@ HnswLikeNns::remove_weakest(const NearestList &neighbors, uint32_t curMax, LinkL
     return result;
 }
 
+#define NO_BACKFILL
 #ifdef NO_BACKFILL
 LinkList
 HnswLikeNns::select_neighbors(const NearestList &neighbors, uint32_t curMax) const
 {
     LinkList result;
     result.reserve(curMax+1);
-    bool needFiltering = (neighbors.size() > curMax);
     NearestPriQ w;
     for (const auto & entry : neighbors) {
         w.push(entry);
@@ -472,12 +567,16 @@ HnswLikeNns::select_neighbors(const NearestList &neighbors, uint32_t curMax) con
     while (! w.empty()) {
         HnswHit e = w.top();
         w.pop();
-        if (needFiltering && haveCloserDistance(e, result)) {
+        if (haveCloserDistance(e, result)) {
             continue;
         }
         result.push_back(e.docid);
-        if (result.size() == curMax) return result;
+        if (result.size() == curMax) {
+            ++select_n_full;
+            return result;
+        }
     }
+    ++select_n_partial;
     return result;
 }
 #else
@@ -502,10 +601,10 @@ HnswLikeNns::select_neighbors(const NearestList &neighbors, uint32_t curMax) con
         result.push_back(e.docid);
         if (result.size() == curMax) return result;
     }
-    if (result.size() * 4 < curMax) {
+    if (result.size() * 4 < _M) {
         for (uint32_t fill_id : backfill) {
             result.push_back(fill_id);
-            if (result.size() * 4 >= curMax) break;
+            if (result.size() * 2 >= _M) break;
         }
     }
     return result;
@@ -576,7 +675,9 @@ HnswLikeNns::dumpStats() const {
         for (uint32_t n_id : link_list) {
             const LinkList &neigh_list = getLinkList(n_id, 0);
             if (! neigh_list.has_link_to(id)) {
+#ifdef KEEP_SYM
                 fprintf(stderr, "BAD: %u has link to neighbor %u, but backlink is missing\n", id, n_id);
+#endif
                 all_sym = false;
             }
         }
diff --git a/eval/src/tests/ann/xp-lsh-nns.cpp b/eval/src/tests/ann/xp-lsh-nns.cpp
index 0ea119a9c70..c028a07a9d7 100644
--- a/eval/src/tests/ann/xp-lsh-nns.cpp
+++ b/eval/src/tests/ann/xp-lsh-nns.cpp
@@ -118,6 +118,7 @@ public:
         }
     }
     std::vector<NnsHit> topK(uint32_t k, Vector vector, uint32_t search_k) override;
+    std::vector<NnsHit> topKfilter(uint32_t k, Vector vector, uint32_t search_k, const BitVector &bitvector) override;
 
     V getVector(uint32_t docid) const { return _dva.get(docid); }
     double uniformRnd() { return _rndGen.nextUniform(); } 
@@ -196,6 +197,45 @@ public:
 };
 
 std::vector<NnsHit>
+RpLshNns::topKfilter(uint32_t k, Vector vector, uint32_t search_k, const BitVector &blacklist)
+{
+    std::vector<NnsHit> result;
+    result.reserve(k);
+
+    std::vector<float> tmp(_numDims);
+    vespalib::ArrayRef<float> tmpArr(tmp);
+
+    LsMaskHash query_hash = mask_hash_from_pv(vector, _transformationMatrix);
+    LshHitHeap heap(std::max(k, search_k));
+    int limit_hash_dist = 99999;
+    int skipCnt = 0;
+    int fullCnt = 0;
+    int whdcCnt = 0;
+    size_t docidLimit = _generated_doc_hashes.size();
+    for (uint32_t docid = 0; docid < docidLimit; ++docid) {
+        if (blacklist.isSet(docid)) continue;
+        int hd = hash_dist(query_hash, _generated_doc_hashes[docid]);
+        if (hd <= limit_hash_dist) {
+            ++fullCnt;
+            double dist = l2distCalc.l2sq_dist(vector, _dva.get(docid), tmpArr);
+            LshHit h(docid, dist, hd);
+            if (heap.maybe_use(h)) {
+                ++whdcCnt;
+                limit_hash_dist = heap.limitHashDistance();
+            }
+        } else {
+            ++skipCnt;
+        }
+    }
+    std::vector<LshHit> best = heap.bestLshHits();
+    size_t numHits = std::min((size_t)k, best.size());
+    for (size_t i = 0; i < numHits; ++i) {
+        result.emplace_back(best[i].docid, SqDist(best[i].distance));
+    }
+    return result;
+}
+
+std::vector<NnsHit>
 RpLshNns::topK(uint32_t k, Vector vector, uint32_t search_k)
 {
     std::vector<NnsHit> result;