(improvement)(chat) Reduce frequent loading of embedding models to improve loading performance. (#1478)

common/src/main/java/com/tencent/supersonic/common/config/EmbeddingModelParameterConfig.java

@@ -4,9 +4,9 @@ import com.google.common.collect.ImmutableMap;
 import com.google.common.collect.Lists;
 import com.tencent.supersonic.common.pojo.EmbeddingModelConfig;
 import com.tencent.supersonic.common.pojo.Parameter;
-import dev.langchain4j.inmemory.spring.InMemoryAutoConfig;
 import dev.langchain4j.provider.AzureModelFactory;
 import dev.langchain4j.provider.DashscopeModelFactory;
+import dev.langchain4j.provider.EmbeddingModelConstant;
 import dev.langchain4j.provider.InMemoryModelFactory;
 import dev.langchain4j.provider.OllamaModelFactory;
 import dev.langchain4j.provider.OpenAiModelFactory;
@@ -76,7 +76,7 @@ public class EmbeddingModelParameterConfig extends ParameterConfig {
 
 
     public static final Parameter EMBEDDING_MODEL_NAME =
-            new Parameter("s2.embedding.model.name", InMemoryAutoConfig.BGE_SMALL_ZH,
+            new Parameter("s2.embedding.model.name", EmbeddingModelConstant.BGE_SMALL_ZH,
                     "ModelName", "",
                     "string", "向量模型配置", null,
                     getDependency(EMBEDDING_MODEL_PROVIDER.getName(),
@@ -90,7 +90,7 @@ public class EmbeddingModelParameterConfig extends ParameterConfig {
                             ZhipuModelFactory.PROVIDER
                     ),
                     ImmutableMap.of(
-                            InMemoryModelFactory.PROVIDER, InMemoryAutoConfig.BGE_SMALL_ZH,
+                            InMemoryModelFactory.PROVIDER, EmbeddingModelConstant.BGE_SMALL_ZH,
                             OpenAiModelFactory.PROVIDER, "text-embedding-ada-002",
                             OllamaModelFactory.PROVIDER, "all-minilm",
                             AzureModelFactory.PROVIDER, "text-embedding-ada-002",

common/src/main/java/dev/langchain4j/inmemory/spring/InMemoryAutoConfig.java

@@ -5,6 +5,7 @@ import dev.langchain4j.model.embedding.AllMiniLmL6V2QuantizedEmbeddingModel;
 import dev.langchain4j.model.embedding.BgeSmallZhEmbeddingModel;
 import dev.langchain4j.model.embedding.EmbeddingModel;
 import dev.langchain4j.model.embedding.S2OnnxEmbeddingModel;
+import dev.langchain4j.provider.EmbeddingModelConstant;
 import dev.langchain4j.store.embedding.EmbeddingStoreFactory;
 import org.apache.commons.lang3.StringUtils;
 import org.springframework.boot.autoconfigure.condition.ConditionalOnProperty;
@@ -17,10 +18,6 @@ import static dev.langchain4j.inmemory.spring.Properties.PREFIX;
 @Configuration
 @EnableConfigurationProperties(Properties.class)
 public class InMemoryAutoConfig {
-
-    public static final String BGE_SMALL_ZH = "bge-small-zh";
-    public static final String ALL_MINILM_L6_V2 = "all-minilm-l6-v2-q";
-
     @Bean
     @ConditionalOnProperty(PREFIX + ".embedding-store.persist-path")
     EmbeddingStoreFactory inMemoryChatModel(Properties properties) {
@@ -37,10 +34,10 @@ public class InMemoryAutoConfig {
             return new S2OnnxEmbeddingModel(modelPath, vocabularyPath);
         }
         String modelName = embeddingModelProperties.getModelName();
-        if (BGE_SMALL_ZH.equalsIgnoreCase(modelName)) {
+        if (EmbeddingModelConstant.BGE_SMALL_ZH.equalsIgnoreCase(modelName)) {
             return new BgeSmallZhEmbeddingModel();
         }
-        if (ALL_MINILM_L6_V2.equalsIgnoreCase(modelName)) {
+        if (EmbeddingModelConstant.ALL_MINILM_L6_V2.equalsIgnoreCase(modelName)) {
             return new AllMiniLmL6V2QuantizedEmbeddingModel();
         }
         return new BgeSmallZhEmbeddingModel();

common/src/main/java/dev/langchain4j/model/embedding/S2OnnxEmbeddingModel.java

@@ -8,6 +8,7 @@ import java.net.URL;
 import java.nio.file.Files;
 import java.nio.file.Path;
 import java.nio.file.Paths;
+import java.util.Objects;
 
 /**
  * An embedding model that runs within your Java application's process.
@@ -18,13 +19,14 @@ import java.nio.file.Paths;
  * Copy from dev.langchain4j.model.embedding.OnnxEmbeddingModel.
  */
 public class S2OnnxEmbeddingModel extends AbstractInProcessEmbeddingModel {
+    private static volatile OnnxBertBiEncoder cachedModel;
+    private static volatile String cachedModelPath;
+    private static volatile String cachedVocabularyPath;
 
-    private static OnnxBertBiEncoder model = null;
-
-    /**
-     * @param pathToModel The path to the .onnx model file (e.g., "/home/me/model.onnx").
-     */
     public S2OnnxEmbeddingModel(String pathToModel, String vocabularyPath) {
+        if (shouldReloadModel(pathToModel, vocabularyPath)) {
+            synchronized (S2OnnxEmbeddingModel.class) {
+                if (shouldReloadModel(pathToModel, vocabularyPath)) {
                     URL resource = AbstractInProcessEmbeddingModel.class.getResource("/bert-vocabulary-en.txt");
                     if (StringUtils.isNotBlank(vocabularyPath)) {
                         try {
@@ -33,21 +35,26 @@ public class S2OnnxEmbeddingModel extends AbstractInProcessEmbeddingModel {
                             throw new RuntimeException(e);
                         }
                     }
-        if (model == null) {
+                    cachedModel = loadFromFileSystem(Paths.get(pathToModel), resource);
-            model = loadFromFileSystem(Paths.get(pathToModel), resource);
+                    cachedModelPath = pathToModel;
+                    cachedVocabularyPath = vocabularyPath;
+                }
+            }
         }
     }
 
-    /**
-     * @param pathToModel The path to the .onnx model file (e.g., "/home/me/model.onnx").
-     */
     public S2OnnxEmbeddingModel(String pathToModel) {
         this(pathToModel, null);
     }
 
     @Override
     protected OnnxBertBiEncoder model() {
-        return model;
+        return cachedModel;
+    }
+
+    private static boolean shouldReloadModel(String pathToModel, String vocabularyPath) {
+        return cachedModel == null || !Objects.equals(cachedModelPath, pathToModel)
+                || !Objects.equals(cachedVocabularyPath, vocabularyPath);
     }
 
     static OnnxBertBiEncoder loadFromFileSystem(Path pathToModel, URL vocabularyFile) {

common/src/main/java/dev/langchain4j/provider/EmbeddingModelConstant.java

@@ -0,0 +1,16 @@
+package dev.langchain4j.provider;
+
+import dev.langchain4j.model.embedding.AllMiniLmL6V2QuantizedEmbeddingModel;
+import dev.langchain4j.model.embedding.BgeSmallZhEmbeddingModel;
+import dev.langchain4j.model.embedding.EmbeddingModel;
+import org.springframework.stereotype.Service;
+
+@Service
+public class EmbeddingModelConstant {
+
+    public static final String BGE_SMALL_ZH = "bge-small-zh";
+    public static final String ALL_MINILM_L6_V2 = "all-minilm-l6-v2-q";
+    public static final EmbeddingModel BGE_SMALL_ZH_MODEL = new BgeSmallZhEmbeddingModel();
+    public static final EmbeddingModel ALL_MINI_LM_L6_V2_MODEL = new AllMiniLmL6V2QuantizedEmbeddingModel();
+
+}

common/src/main/java/dev/langchain4j/provider/InMemoryModelFactory.java

@@ -3,17 +3,12 @@ package dev.langchain4j.provider;
 import com.tencent.supersonic.common.pojo.ChatModelConfig;
 import com.tencent.supersonic.common.pojo.EmbeddingModelConfig;
 import dev.langchain4j.model.chat.ChatLanguageModel;
-import dev.langchain4j.model.embedding.AllMiniLmL6V2QuantizedEmbeddingModel;
-import dev.langchain4j.model.embedding.BgeSmallZhEmbeddingModel;
 import dev.langchain4j.model.embedding.EmbeddingModel;
 import dev.langchain4j.model.embedding.S2OnnxEmbeddingModel;
 import org.apache.commons.lang3.StringUtils;
 import org.springframework.beans.factory.InitializingBean;
 import org.springframework.stereotype.Service;
 
-import static dev.langchain4j.inmemory.spring.InMemoryAutoConfig.ALL_MINILM_L6_V2;
-import static dev.langchain4j.inmemory.spring.InMemoryAutoConfig.BGE_SMALL_ZH;
-
 @Service
 public class InMemoryModelFactory implements ModelFactory, InitializingBean {
     public static final String PROVIDER = "IN_MEMORY";
@@ -31,13 +26,13 @@ public class InMemoryModelFactory implements ModelFactory, InitializingBean {
             return new S2OnnxEmbeddingModel(modelPath, vocabularyPath);
         }
         String modelName = embeddingModel.getModelName();
-        if (BGE_SMALL_ZH.equalsIgnoreCase(modelName)) {
+        if (EmbeddingModelConstant.BGE_SMALL_ZH.equalsIgnoreCase(modelName)) {
-            return new BgeSmallZhEmbeddingModel();
+            return EmbeddingModelConstant.BGE_SMALL_ZH_MODEL;
         }
-        if (ALL_MINILM_L6_V2.equalsIgnoreCase(modelName)) {
+        if (EmbeddingModelConstant.ALL_MINILM_L6_V2.equalsIgnoreCase(modelName)) {
-            return new AllMiniLmL6V2QuantizedEmbeddingModel();
+            return EmbeddingModelConstant.ALL_MINI_LM_L6_V2_MODEL;
         }
-        return new BgeSmallZhEmbeddingModel();
+        return EmbeddingModelConstant.BGE_SMALL_ZH_MODEL;
     }
 
     @Override

headless/server/src/main/java/com/tencent/supersonic/headless/server/facade/service/impl/S2SemanticLayerService.java

@@ -150,8 +150,8 @@ public class S2SemanticLayerService implements SemanticLayerService {
             //2.query from cache
 
             String cacheKey = queryCache.getCacheKey(queryReq);
-            log.debug("cacheKey:{}", cacheKey);
             Object query = queryCache.query(queryReq, cacheKey);
+            log.info("cacheKey:{},query:{}", cacheKey, query);
             if (Objects.nonNull(query)) {
                 SemanticQueryResp queryResp = (SemanticQueryResp) query;
                 queryResp.setUseCache(true);