sparse: disable refinement by default

Because of the removal of forward index, the current refine searching is slower than before. To prevent performance degradation when using a non-zero value of drop_ratio_search, disable refinement by default. Signed-off-by: Shawn Wang <[email protected]>
zilliztech · Jan 16, 2025 · 2b9ee37 · 2b9ee37
1 parent 45d757c
commit 2b9ee37
Show file tree

Hide file tree

Showing 2 changed files with 14 additions and 5 deletions.
diff --git a/src/index/sparse/sparse_index_node.cc b/src/index/sparse/sparse_index_node.cc
@@ -104,7 +104,7 @@ class SparseInvertedIndexNode : public IndexNode {
         auto nq = dataset->GetRows();
         auto queries = static_cast<const sparse::SparseRow<T>*>(dataset->GetTensor());
         auto k = cfg.k.value();
-        auto refine_factor = cfg.refine_factor.value_or(10);
+        auto refine_factor = cfg.refine_factor.value_or(1);
         auto drop_ratio_search = cfg.drop_ratio_search.value_or(0.0f);
 
         auto p_id = std::make_unique<sparse::label_t[]>(nq * k);
@@ -185,7 +185,8 @@ class SparseInvertedIndexNode : public IndexNode {
         auto computer = computer_or.value();
         auto drop_ratio_search = cfg.drop_ratio_search.value_or(0.0f);
 
-        const bool approximated = drop_ratio_search > 0;
+        // TODO: set approximated to false for now since the refinement is too slow after forward index is removed.
+        const bool approximated = false;
 
         auto vec = std::vector<std::shared_ptr<IndexNode::iterator>>(nq, nullptr);
         try {

diff --git a/src/index/sparse/sparse_inverted_index_config.h b/src/index/sparse/sparse_inverted_index_config.h
@@ -38,11 +38,19 @@ class SparseInvertedIndexConfig : public BaseConfig {
             .for_search()
             .for_range_search()
             .for_iterator();
+        /**
+         * refine_factor is used for approximate search.
+         * refine_factor == 1 means no refinement, and is the default value.
+         * refine_factor > 1 means refinement. The larger the value, the more
+         * accurate the approximate result will be, but the slower the
+         * performance.
+         * Be aware that if you opt to use a large drop_ratio_search, it is
+         * necessary for you to manually modify this value.
+         */
         KNOWHERE_CONFIG_DECLARE_FIELD(refine_factor)
             .description("refine factor")
-            .set_default(10)
-            .for_search()
-            .for_range_search();
+            .set_default(1)
+            .for_search();
         /**
          * The term frequency part of score of BM25 is:
          * tf * (k1 + 1) / (tf + k1 * (1 - b + b * (doc_len / avgdl)))