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
C++ algoliadatastoreelasticsearchenterprise-searchfacetingfull-text-searchfuzzy-searchgeosearchin-memoryinstantsearchmerchandisingpineconesearchsearch-enginesemantic-searchsimilarity-searchsite-searchsynonymstypo-tolerancevector-search
Update queries and add support for query weights
ozanarmagan371dec626042
diff --git a/include/vector_query_ops.h b/include/vector_query_ops.hindex 21b17053..120c1d0b 100644--- a/include/vector_query_ops.h+++ b/include/vector_query_ops.h@@ -17,7 +17,8 @@ struct vector_query_t {bool query_doc_given = false;float alpha = 0.3;- std::vector<std::string> qs;+ std::vector<std::string> queries;+ std::vector<float> query_weights;void _reset() {// used for testing onlydiff --git a/src/collection.cpp b/src/collection.cppindex d4dc944f..3e5a8818 100644--- a/src/collection.cpp+++ b/src/collection.cpp@@ -1220,13 +1220,13 @@ Option<bool> Collection::validate_and_standardize_sort_fields(const std::vector<return Option<bool>(400, "Could not find a field named `" + sort_field_std.vector_query.query.field_name + "` in vector index.");}- if(!sort_field_std.vector_query.query.qs.empty()) {+ if(!sort_field_std.vector_query.query.queries.empty()) {if(embedding_fields.find(sort_field_std.vector_query.query.field_name) == embedding_fields.end()) {- return Option<bool>(400, "`qs` parameter is only supported for auto-embedding fields.");+ return Option<bool>(400, "`queries` parameter is only supported for auto-embedding fields.");}std::vector<std::vector<float>> embeddings;- for(const auto& q: sort_field_std.vector_query.query.qs) {+ for(const auto& q: sort_field_std.vector_query.query.queries) {EmbedderManager& embedder_manager = EmbedderManager::get_instance();auto embedder_op = embedder_manager.get_text_embedder(vector_field_it.value().embed[fields::model_config]);if(!embedder_op.ok()) {@@ -1263,22 +1263,28 @@ Option<bool> Collection::validate_and_standardize_sort_fields(const std::vector<embeddings.emplace_back(embedding_op.embedding);}- // get average of all embeddings- std::vector<float> avg_embedding(vector_field_it.value().num_dim, 0);- for(const auto& embedding: embeddings) {- for(size_t i = 0; i < embedding.size(); i++) {- avg_embedding[i] += embedding[i];+ if(sort_field_std.vector_query.query.query_weights.empty()) {+ // get average of all embeddings+ std::vector<float> avg_embedding(vector_field_it.value().num_dim, 0);+ for(const auto& embedding: embeddings) {+ for(size_t i = 0; i < embedding.size(); i++) {+ avg_embedding[i] += embedding[i];+ }+ }+ for(size_t i = 0; i < avg_embedding.size(); i++) {+ avg_embedding[i] /= embeddings.size();}- }- for(size_t i = 0; i < avg_embedding.size(); i++) {- avg_embedding[i] /= embeddings.size();- }- sort_field_std.vector_query.query.values = avg_embedding;- if(vector_field_it.value().vec_dist == cosine) {- std::vector<float> normalized_values(sort_field_std.vector_query.query.values.size());- hnsw_index_t::normalize_vector(sort_field_std.vector_query.query.values, normalized_values);- sort_field_std.vector_query.query.values = normalized_values;+ sort_field_std.vector_query.query.values = avg_embedding;+ } else {+ std::vector<float> weighted_embeddings(vector_field_it.value().num_dim, 0);+ for(size_t i = 0; i < embeddings.size(); i++) {+ for(size_t j = 0; j < embeddings[i].size(); j++) {+ weighted_embeddings[j] += embeddings[i][j] * sort_field_std.vector_query.query.query_weights[i];+ }+ }++ sort_field_std.vector_query.query.values = weighted_embeddings;}}@@ -6379,13 +6385,13 @@ Option<bool> Collection::parse_and_validate_vector_query(const std::string& vectreturn Option<bool>(400, "Field `" + vector_query.field_name + "` is marked as a non-indexed field in the schema.");}- if(!vector_query.qs.empty()) {+ if(!vector_query.queries.empty()) {if(embedding_fields.find(vector_query.field_name) == embedding_fields.end()) {- return Option<bool>(400, "`qs` parameter is only supported for auto-embedding fields.");+ return Option<bool>(400, "`queries` parameter is only supported for auto-embedding fields.");}std::vector<std::vector<float>> embeddings;- for(const auto& q: vector_query.qs) {+ for(const auto& q: vector_query.queries) {EmbedderManager& embedder_manager = EmbedderManager::get_instance();auto embedder_op = embedder_manager.get_text_embedder(vector_field_it.value().embed[fields::model_config]);if(!embedder_op.ok()) {@@ -6421,19 +6427,30 @@ Option<bool> Collection::parse_and_validate_vector_query(const std::string& vectembeddings.emplace_back(embedding_op.embedding);}++ if(vector_query.query_weights.empty()) {+ // get average of all embeddings+ std::vector<float> avg_embedding(vector_field_it.value().num_dim, 0);+ for(const auto& embedding: embeddings) {+ for(size_t i = 0; i < embedding.size(); i++) {+ avg_embedding[i] += embedding[i];+ }+ }+ for(size_t i = 0; i < avg_embedding.size(); i++) {+ avg_embedding[i] /= embeddings.size();+ }- // get average of all embeddings- std::vector<float> avg_embedding(vector_field_it.value().num_dim, 0);- for(const auto& embedding: embeddings) {- for(size_t i = 0; i < embedding.size(); i++) {- avg_embedding[i] += embedding[i];+ vector_query.values = avg_embedding;+ } else {+ std::vector<float> embeddings_with_weights(vector_field_it.value().num_dim, 0);+ for(size_t i = 0; i < embeddings.size(); i++) {+ for(size_t j = 0; j < embeddings[i].size(); j++) {+ embeddings_with_weights[j] += embeddings[i][j] * vector_query.query_weights[i];+ }}- }- for(size_t i = 0; i < avg_embedding.size(); i++) {- avg_embedding[i] /= embeddings.size();- }- vector_query.values = avg_embedding;+ vector_query.values = embeddings_with_weights;+ }}if(is_wildcard_query) {diff --git a/src/vector_query_ops.cpp b/src/vector_query_ops.cppindex c40fa07c..246d3bbd 100644--- a/src/vector_query_ops.cpp+++ b/src/vector_query_ops.cpp@@ -92,7 +92,7 @@ Option<bool> VectorQueryOps::parse_vector_query_str(const std::string& vector_qu}if(i < param_kvs.size() - 1 && StringUtils::get_occurence_count(param_kv_str, '[') != StringUtils::get_occurence_count(param_kv_str, ']')) {- // this is a hack to handle the case where the query string contains a comma in the `qs` parameter+ // this is a hack to handle the case where the query string contains a comma in the vector valuesparam_kvs[i+1] = param_kv_str + "," + param_kvs[i+1];continue;}@@ -174,10 +174,10 @@ Option<bool> VectorQueryOps::parse_vector_query_str(const std::string& vector_quvector_query.alpha = std::stof(param_kv[1]);}- if(param_kv[0] == "qs") {+ if(param_kv[0] == "queries") {if(param_kv[1].front() != '[' || param_kv[1].back() != ']') {return Option<bool>(400, "Malformed vector query string: "- "`qs` parameter must be a list of strings.");+ "`queries` parameter must be a list of strings.");}param_kv[1].erase(0, 1);@@ -187,9 +187,48 @@ Option<bool> VectorQueryOps::parse_vector_query_str(const std::string& vector_quStringUtils::split(param_kv[1], qs, ",");for(auto& q: qs) {StringUtils::trim(q);- vector_query.qs.push_back(q);+ vector_query.queries.push_back(q);}}++ if(param_kv[0] == "query_weights") {+ if(param_kv[1].front() != '[' || param_kv[1].back() != ']') {+ return Option<bool>(400, "Malformed vector query string: "+ "`query_weights` parameter must be a list of floats.");+ }++ param_kv[1].erase(0, 1);+ param_kv[1].pop_back();++ std::vector<std::string> ws;+ StringUtils::split(param_kv[1], ws, ",");+ for(auto& w: ws) {+ StringUtils::trim(w);+ if(!StringUtils::is_float(w)) {+ return Option<bool>(400, "Malformed vector query string: "+ "`query_weights` parameter must be a list of floats.");+ }++ vector_query.query_weights.push_back(std::stof(w));+ }+ }+ }++ if(vector_query.queries.size() != vector_query.query_weights.size() && !vector_query.query_weights.empty()) {+ return Option<bool>(400, "Malformed vector query string: "+ "`queries` and `query_weights` must be of the same length.");+ }++ if(!vector_query.query_weights.empty()) {+ float sum = 0.0;+ for(auto& w: vector_query.query_weights) {+ sum += w;+ }++ if(sum != 1.0) {+ return Option<bool>(400, "Malformed vector query string: "+ "`query_weights` must sum to 1.0.");+ }}return Option<bool>(true);diff --git a/test/collection_sorting_test.cpp b/test/collection_sorting_test.cppindex 41fba04f..9d8d3bb9 100644--- a/test/collection_sorting_test.cpp+++ b/test/collection_sorting_test.cpp@@ -2623,7 +2623,7 @@ TEST_F(CollectionSortingTest, TestVectorQueryQsSorting) {ASSERT_EQ("0", results["hits"][1]["document"]["id"]);sort_fields = {- sort_by("_vector_query(embedding:([], qs: [powerpuff girls, cartoon]))", "asc"),+ sort_by("_vector_query(embedding:([], queries: [powerpuff girls, cartoon]))", "asc"),};results = coll->search("butter", {"name"}, "", {}, sort_fields, {0}, 10, 1, FREQUENCY, {true}, Index::DROP_TOKENS_THRESHOLD,diff --git a/test/collection_vector_search_test.cpp b/test/collection_vector_search_test.cppindex 1feb7dbc..f15f39f6 100644--- a/test/collection_vector_search_test.cpp+++ b/test/collection_vector_search_test.cpp@@ -3425,7 +3425,7 @@ TEST_F(CollectionVectorTest, TestVectorQueryQs) {"<mark>", "</mark>", {}, 1000, true, false, true, "", false, 6000 * 1000, 4, 7,fallback,4, {off}, 32767, 32767, 2,- false, true, "embedding:([], qs:[superhero, company])");+ false, true, "embedding:([], queries:[superhero, company])");ASSERT_TRUE(results.ok());ASSERT_EQ(results.get()["hits"].size(), 1);@@ -3463,7 +3463,7 @@ TEST_F(CollectionVectorTest, TestVectorQueryInvalidQs) {"<mark>", "</mark>", {}, 1000, true, false, true, "", false, 6000 * 1000, 4, 7,fallback,4, {off}, 32767, 32767, 2,- false, true, "embedding:([], qs:\"test\")");+ false, true, "embedding:([], queries:\"test\")");ASSERT_FALSE(results.ok());@@ -3477,7 +3477,7 @@ TEST_F(CollectionVectorTest, TestVectorQueryInvalidQs) {"<mark>", "</mark>", {}, 1000, true, false, true, "", false, 6000 * 1000, 4, 7,fallback,4, {off}, 32767, 32767, 2,- false, true, "embedding:([], qs:11)");+ false, true, "embedding:([], queries:11)");ASSERT_FALSE(results.ok());@@ -3518,8 +3518,170 @@ TEST_F(CollectionVectorTest, TestVectorQueryQsWithHybridSearch) {"<mark>", "</mark>", {}, 1000, true, false, true, "", false, 6000 * 1000, 4, 7,fallback,4, {off}, 32767, 32767, 2,- false, true, "embedding:([], qs:[superhero, company])");+ false, true, "embedding:([], queries:[superhero, company])");ASSERT_TRUE(results.ok());ASSERT_EQ(results.get()["hits"].size(), 1);+}++TEST_F(CollectionVectorTest, TestVectorQueryQsHybridSearchAlpha) {+ auto schema_json =+ R"({+ "name": "test",+ "fields": [+ {"name": "name", "type": "string"},+ {"name": "embedding", "type":"float[]", "embed":{"from": ["name"], "model_config": {"model_name": "ts/all-MiniLM-L12-v2"}}}+ ]+ })"_json;++ EmbedderManager::set_model_dir("/tmp/typesense_test/models");++ auto collection_create_op = collectionManager.create_collection(schema_json);++ ASSERT_TRUE(collection_create_op.ok());++ auto coll = collection_create_op.get();++ auto add_op = coll->add(R"({+ "name": "Apple iPhone"+ })"_json.dump());+ ASSERT_TRUE(add_op.ok());++ add_op = coll->add(R"({+ "name": "Samsung Galaxy"+ })"_json.dump());++ auto results = coll->search("apple", {"name"}, "", {}, {}, {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,+ false, true, "embedding:([], queries:[samsung, phone])");++ ASSERT_TRUE(results.ok());+ ASSERT_EQ(results.get()["hits"].size(), 2);+ ASSERT_EQ(results.get()["hits"][0]["document"]["name"], "Apple iPhone");+++ results = coll->search("apple", {"name"}, "", {}, {}, {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,+ false, true, "embedding:([], queries:[samsung, phone], alpha:0.9)");++ ASSERT_TRUE(results.ok());+ ASSERT_EQ(results.get()["hits"].size(), 2);+ ASSERT_EQ(results.get()["hits"][0]["document"]["name"], "Samsung Galaxy");+}++TEST_F(CollectionVectorTest, TestVectorQueryQsWeight) {+ auto schema_json =+ R"({+ "name": "test",+ "fields": [+ {"name": "name", "type": "string"},+ {"name": "embedding", "type":"float[]", "embed":{"from": ["name"], "model_config": {"model_name": "ts/all-MiniLM-L12-v2"}}}+ ]+ })"_json;++ EmbedderManager::set_model_dir("/tmp/typesense_test/models");++ auto collection_create_op = collectionManager.create_collection(schema_json);++ ASSERT_TRUE(collection_create_op.ok());++ auto coll = collection_create_op.get();++ auto add_op = coll->add(R"({+ "name": "Apple iPhone"+ })"_json.dump());+ ASSERT_TRUE(add_op.ok());++ add_op = coll->add(R"({+ "name": "Samsung Galaxy"+ })"_json.dump());++ auto results = coll->search("*", {}, "", {}, {}, {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,+ false, true, "embedding:([], queries:[samsung, apple], query_weights:[0.1, 0.9])");++ ASSERT_TRUE(results.ok());+ ASSERT_EQ(results.get()["hits"].size(), 2);+ ASSERT_EQ(results.get()["hits"][0]["document"]["name"], "Apple iPhone");+++ results = coll->search("*", {}, "", {}, {}, {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,+ false, true, "embedding:([], queries:[samsung, apple], query_weights:[0.9, 0.1])");++ ASSERT_TRUE(results.ok());+ ASSERT_EQ(results.get()["hits"].size(), 2);+ ASSERT_EQ(results.get()["hits"][0]["document"]["name"], "Samsung Galaxy");+}++TEST_F(CollectionVectorTest, TestVectorQueryQsWeightInvalid) {+ auto schema_json =+ R"({+ "name": "test",+ "fields": [+ {"name": "name", "type": "string"},+ {"name": "embedding", "type":"float[]", "embed":{"from": ["name"], "model_config": {"model_name": "ts/all-MiniLM-L12-v2"}}}+ ]+ })"_json;++ EmbedderManager::set_model_dir("/tmp/typesense_test/models");++ auto collection_create_op = collectionManager.create_collection(schema_json);++ ASSERT_TRUE(collection_create_op.ok());++ auto coll = collection_create_op.get();++ auto add_op = coll->add(R"({+ "name": "Apple iPhone"+ })"_json.dump());+ ASSERT_TRUE(add_op.ok());++ add_op = coll->add(R"({+ "name": "Samsung Galaxy"+ })"_json.dump());++ auto results = coll->search("*", {}, "", {}, {}, {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,+ false, true, "embedding:([], queries:[samsung, apple], query_weights:[0.1, 0.9, 0.1])");++ ASSERT_FALSE(results.ok());+ ASSERT_EQ(results.error(), "Malformed vector query string: `queries` and `query_weights` must be of the same length.");++ results = coll->search("*", {}, "", {}, {}, {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,+ false, true, "embedding:([], queries:[samsung, apple], query_weights:[0.4, 0.9])");++ ASSERT_FALSE(results.ok());+ ASSERT_EQ(results.error(), "Malformed vector query string: `query_weights` must sum to 1.0.");}\ No newline at end of file