Skip to content

Open Source alternative to Algolia + Pinecone and an Easier-to-Use alternative to ElasticSearch ⚡ 🔍 ✨ Fast, typo tolerant, in-memory fuzzy Search Engine for building delightful search experiences

Fix eval + hybrid search.

Kishore Nallan9a6ef3cbe482
diff --git a/src/index.cpp b/src/index.cpp
index 5d51487c..d7a5ead6 100644
--- a/src/index.cpp
+++ b/src/index.cpp
@@ -3367,11 +3367,22 @@ Option<bool> Index::search(std::vector<query_tokens_t>& field_query_tokens, cons
}
vec_results.emplace_back(seq_id, vec_dist_score);
}
-
+
+ // iteration needs to happen on sorted sequence ID but score wise sort needed for compute rank fusion
std::sort(vec_results.begin(), vec_results.end(), [](const auto& a, const auto& b) {
return a.second < b.second;
});
+ std::unordered_map<uint32_t, uint32_t> seq_id_to_rank;
+
+ for(size_t vec_index = 0; vec_index < vec_results.size(); vec_index++) {
+ seq_id_to_rank.emplace(vec_results[vec_index].first, vec_index);
+ }
+
+ std::sort(vec_results.begin(), vec_results.end(), [](const auto& a, const auto& b) {
+ return a.first < b.first;
+ });
+
std::vector<KV*> kvs;
if(group_limit != 0) {
for(auto& kv_map : topster->group_kv_map) {
@@ -3385,8 +3396,6 @@ Option<bool> Index::search(std::vector<query_tokens_t>& field_query_tokens, cons
topster->sort();
}
-
-
// Reciprocal rank fusion
// Score is sum of (1 / rank_of_document) * WEIGHT from each list (text match and vector search)
auto size = (group_limit != 0) ? kvs.size() : topster->size;
@@ -3444,7 +3453,7 @@ Option<bool> Index::search(std::vector<query_tokens_t>& field_query_tokens, cons
found_kv->text_match_score = found_kv->scores[found_kv->match_score_index];
int64_t match_score = float_to_int64_t(
(int64_t_to_float(found_kv->scores[found_kv->match_score_index])) +
- ((1.0 / (res_index + 1)) * VECTOR_SEARCH_WEIGHT));
+ ((1.0 / (seq_id_to_rank[seq_id] + 1)) * VECTOR_SEARCH_WEIGHT));
int64_t match_score_index = -1;
int64_t scores[3] = {0};
@@ -3466,7 +3475,7 @@ Option<bool> Index::search(std::vector<query_tokens_t>& field_query_tokens, cons
// Result has been found only in vector search: we have to add it to both KV and result_ids
// (1 / rank_of_document) * WEIGHT)
int64_t scores[3] = {0};
- int64_t match_score = float_to_int64_t((1.0 / (res_index + 1)) * VECTOR_SEARCH_WEIGHT);
+ int64_t match_score = float_to_int64_t((1.0 / (seq_id_to_rank[seq_id] + 1)) * VECTOR_SEARCH_WEIGHT);
int64_t match_score_index = -1;
auto compute_sort_scores_op = compute_sort_scores(sort_fields_std, sort_order, field_values,
diff --git a/test/collection_vector_search_test.cpp b/test/collection_vector_search_test.cpp
index 452e8a4a..42009f85 100644
--- a/test/collection_vector_search_test.cpp
+++ b/test/collection_vector_search_test.cpp
@@ -1423,6 +1423,54 @@ TEST_F(CollectionVectorTest, HybridSearchSortByGeopoint) {
ASSERT_EQ("butterfly", search_res["hits"][2]["document"]["name"].get<std::string>());
}
+TEST_F(CollectionVectorTest, HybridSearchWithEvalSort) {
+ nlohmann::json schema = R"({
+ "name": "coll1",
+ "fields": [
+ {"name": "name", "type": "string", "facet": true},
+ {"name": "category", "type": "string", "facet": true},
+ {"name": "vec", "type": "float[]", "embed":{"from": ["name"], "model_config": {"model_name": "ts/e5-small"}}}
+ ]
+ })"_json;
+
+ EmbedderManager::set_model_dir("/tmp/typesense_test/models");
+ Collection* coll1 = collectionManager.create_collection(schema).get();
+
+ nlohmann::json doc;
+ doc["id"] = "0";
+ doc["name"] = "Apple Fruit";
+ doc["category"] = "Fresh";
+ ASSERT_TRUE(coll1->add(doc.dump()).ok());
+
+ doc["id"] = "1";
+ doc["name"] = "Apple";
+ doc["category"] = "Phone";
+ ASSERT_TRUE(coll1->add(doc.dump()).ok());
+
+ doc["id"] = "2";
+ doc["name"] = "Apple Pie";
+ doc["category"] = "Notebook";
+ ASSERT_TRUE(coll1->add(doc.dump()).ok());
+
+ std::vector<sort_by> sort_fields;
+ CollectionManager::parse_sort_by_str("_eval([(category:Fresh):3,(category:Notebook):2,(category:Phone):1]):desc", sort_fields);
+
+ auto results_op = coll1->search("apple", {"name", "vec"}, "", {"name"}, sort_fields, {0}, 20, 1, FREQUENCY, {true},
+ Index::DROP_TOKENS_THRESHOLD,
+ spp::sparse_hash_set<std::string>(),
+ spp::sparse_hash_set<std::string>(), 10, "", 30, 5,
+ "", 10, {}, {}, {}, 0,
+ "<mark>", "</mark>", {}, 1000, true, false, true, "", false, 6000 * 1000, 4, 7,
+ fallback,
+ 4, {off}, 32767, 32767, 2);
+ ASSERT_EQ(true, results_op.ok());
+ ASSERT_EQ(3, results_op.get()["found"].get<size_t>());
+ ASSERT_EQ(3, results_op.get()["hits"].size());
+
+ ASSERT_EQ("0", results_op.get()["hits"][0]["document"]["id"].get<std::string>());
+ ASSERT_EQ("2", results_op.get()["hits"][1]["document"]["id"].get<std::string>());
+ ASSERT_EQ("1", results_op.get()["hits"][2]["document"]["id"].get<std::string>());
+}
TEST_F(CollectionVectorTest, EmbedFromOptionalNullField) {
nlohmann::json schema = R"({