Add SqlGeneration abstraction and implementation, optimize LLMSqlParser skip. (#487)

chat/core/src/main/java/com/tencent/supersonic/chat/parser/EmbedLLMProxy.java

@@ -1,68 +1,42 @@
 package com.tencent.supersonic.chat.parser;
 
-import com.tencent.supersonic.chat.config.OptimizationConfig;
+import com.tencent.supersonic.chat.api.pojo.QueryContext;
 import com.tencent.supersonic.chat.parser.plugin.function.FunctionCallPromptGenerator;
-import com.tencent.supersonic.chat.parser.sql.llm.prompt.OutputFormat;
-import com.tencent.supersonic.chat.parser.sql.llm.prompt.SqlExampleLoader;
-import com.tencent.supersonic.chat.parser.sql.llm.prompt.SqlPromptGenerator;
 import com.tencent.supersonic.chat.parser.plugin.function.FunctionReq;
 import com.tencent.supersonic.chat.parser.plugin.function.FunctionResp;
+import com.tencent.supersonic.chat.parser.sql.llm.SqlGeneration;
+import com.tencent.supersonic.chat.parser.sql.llm.SqlGenerationFactory;
+import com.tencent.supersonic.chat.parser.sql.llm.prompt.OutputFormat;
 import com.tencent.supersonic.chat.query.llm.s2sql.LLMReq;
-import com.tencent.supersonic.chat.query.llm.s2sql.LLMReq.ElementValue;
 import com.tencent.supersonic.chat.query.llm.s2sql.LLMResp;
 import com.tencent.supersonic.common.util.ContextUtils;
-import com.tencent.supersonic.common.util.JsonUtil;
-import dev.langchain4j.data.message.AiMessage;
 import dev.langchain4j.model.chat.ChatLanguageModel;
-import dev.langchain4j.model.input.Prompt;
+import java.util.Objects;
-import dev.langchain4j.model.input.PromptTemplate;
-import dev.langchain4j.model.output.Response;
 import lombok.extern.slf4j.Slf4j;
 
-import java.util.HashMap;
-import java.util.List;
-import java.util.Map;
-
 @Slf4j
 public class EmbedLLMProxy implements LLMProxy {
 
+    @Override
+    public boolean isSkip(QueryContext queryContext) {
+        ChatLanguageModel chatLanguageModel = ContextUtils.getBean(ChatLanguageModel.class);
+        if (Objects.isNull(chatLanguageModel)) {
+            log.warn("chatLanguageModel is null, skip EmbedLLMProxy");
+            return true;
+        }
+        return false;
+    }
+
     public LLMResp query2sql(LLMReq llmReq, String modelClusterKey) {
 
-        ChatLanguageModel chatLanguageModel = ContextUtils.getBean(ChatLanguageModel.class);
+        SqlGeneration sqlGeneration = SqlGenerationFactory.get(llmReq.getSqlGenerationMode());
-
-        SqlExampleLoader sqlExampleLoader = ContextUtils.getBean(SqlExampleLoader.class);
-
-        OptimizationConfig config = ContextUtils.getBean(OptimizationConfig.class);
-
-        List<Map<String, String>> sqlExamples = sqlExampleLoader.retrieverSqlExamples(llmReq.getQueryText(),
-                config.getText2sqlCollectionName(), config.getText2sqlFewShotsNum());
-
-        String queryText = llmReq.getQueryText();
         String modelName = llmReq.getSchema().getModelName();
-        List<String> fieldNameList = llmReq.getSchema().getFieldNameList();
+        String sql = sqlGeneration.generation(llmReq, modelClusterKey);
-        List<ElementValue> linking = llmReq.getLinking();
-
-        SqlPromptGenerator sqlPromptGenerator = ContextUtils.getBean(SqlPromptGenerator.class);
-        String linkingPromptStr = sqlPromptGenerator.generateSchemaLinkingPrompt(queryText, modelName, fieldNameList,
-                linking, sqlExamples);
-
-        Prompt linkingPrompt = PromptTemplate.from(JsonUtil.toString(linkingPromptStr)).apply(new HashMap<>());
-        Response<AiMessage> linkingResult = chatLanguageModel.generate(linkingPrompt.toSystemMessage());
-
-        String schemaLinkStr = OutputFormat.schemaLinkParse(linkingResult.content().text());
-
-        String generateSqlPrompt = sqlPromptGenerator.generateSqlPrompt(queryText, modelName, schemaLinkStr,
-                llmReq.getCurrentDate(), sqlExamples);
-
-        Prompt sqlPrompt = PromptTemplate.from(JsonUtil.toString(generateSqlPrompt)).apply(new HashMap<>());
-        Response<AiMessage> sqlResult = chatLanguageModel.generate(sqlPrompt.toSystemMessage());
 
         LLMResp result = new LLMResp();
-        result.setQuery(queryText);
+        result.setQuery(llmReq.getQueryText());
-        result.setSchemaLinkingOutput(linkingPromptStr);
-        result.setSchemaLinkStr(schemaLinkStr);
         result.setModelName(modelName);
-        result.setSqlOutput(sqlResult.content().text());
+        result.setSqlOutput(sql);
         return result;
     }
 

chat/core/src/main/java/com/tencent/supersonic/chat/parser/LLMProxy.java

@@ -1,5 +1,6 @@
 package com.tencent.supersonic.chat.parser;
 
+import com.tencent.supersonic.chat.api.pojo.QueryContext;
 import com.tencent.supersonic.chat.parser.plugin.function.FunctionReq;
 import com.tencent.supersonic.chat.parser.plugin.function.FunctionResp;
 import com.tencent.supersonic.chat.query.llm.s2sql.LLMReq;
@@ -12,6 +13,8 @@ import com.tencent.supersonic.chat.query.llm.s2sql.LLMResp;
  */
 public interface LLMProxy {
 
+    boolean isSkip(QueryContext queryContext);
+
     LLMResp query2sql(LLMReq llmReq, String modelClusterKey);
 
     FunctionResp requestFunction(FunctionReq functionReq);

chat/core/src/main/java/com/tencent/supersonic/chat/parser/PythonLLMProxy.java

@@ -1,6 +1,7 @@
 package com.tencent.supersonic.chat.parser;
 
 import com.alibaba.fastjson.JSON;
+import com.tencent.supersonic.chat.api.pojo.QueryContext;
 import com.tencent.supersonic.chat.config.LLMParserConfig;
 import com.tencent.supersonic.chat.parser.plugin.function.FunctionCallConfig;
 import com.tencent.supersonic.chat.parser.plugin.function.FunctionReq;
@@ -12,6 +13,7 @@ import com.tencent.supersonic.common.util.JsonUtil;
 import java.net.URI;
 import java.net.URL;
 import lombok.extern.slf4j.Slf4j;
+import org.apache.commons.lang3.StringUtils;
 import org.springframework.http.HttpEntity;
 import org.springframework.http.HttpHeaders;
 import org.springframework.http.HttpMethod;
@@ -26,6 +28,16 @@ import org.springframework.web.util.UriComponentsBuilder;
 @Slf4j
 public class PythonLLMProxy implements LLMProxy {
 
+    @Override
+    public boolean isSkip(QueryContext queryContext) {
+        LLMParserConfig llmParserConfig = ContextUtils.getBean(LLMParserConfig.class);
+        if (StringUtils.isEmpty(llmParserConfig.getUrl())) {
+            log.warn("llmParserUrl is empty, skip PythonLLMProxy, config:{}", llmParserConfig);
+            return true;
+        }
+        return false;
+    }
+
     public LLMResp query2sql(LLMReq llmReq, String modelClusterKey) {
 
         long startTime = System.currentTimeMillis();

chat/core/src/main/java/com/tencent/supersonic/chat/parser/sql/llm/LLMRequestService.java

@@ -60,13 +60,11 @@ public class LLMRequestService {
     private OptimizationConfig optimizationConfig;
 
     public boolean isSkip(QueryContext queryCtx) {
-        QueryReq request = queryCtx.getRequest();
+        if (llmProxy.isSkip(queryCtx)) {
-        if (StringUtils.isEmpty(llmParserConfig.getUrl())) {
-            log.info("llm parser url is empty, skip {} , llmParserConfig:{}", LLMSqlParser.class, llmParserConfig);
             return true;
         }
         if (SatisfactionChecker.isSkip(queryCtx)) {
-            log.info("skip {}, queryText:{}", LLMSqlParser.class, request.getQueryText());
+            log.info("skip {}, queryText:{}", LLMSqlParser.class, queryCtx.getRequest().getQueryText());
             return true;
         }
         return false;
@@ -104,7 +102,7 @@ public class LLMRequestService {
     }
 
     public LLMReq getLlmReq(QueryContext queryCtx, SemanticSchema semanticSchema,
-                            ModelCluster modelCluster, List<ElementValue> linkingValues) {
+            ModelCluster modelCluster, List<ElementValue> linkingValues) {
         Map<Long, String> modelIdToName = semanticSchema.getModelIdToName();
         String queryText = queryCtx.getRequest().getQueryText();
 
@@ -146,7 +144,7 @@ public class LLMRequestService {
     }
 
     protected List<String> getFieldNameList(QueryContext queryCtx, ModelCluster modelCluster,
-                                            LLMParserConfig llmParserConfig) {
+            LLMParserConfig llmParserConfig) {
 
         Set<String> results = getTopNFieldNames(modelCluster, llmParserConfig);
 

chat/core/src/main/java/com/tencent/supersonic/chat/parser/sql/llm/SqlGeneration.java

@@ -0,0 +1,10 @@
+package com.tencent.supersonic.chat.parser.sql.llm;
+
+
+import com.tencent.supersonic.chat.query.llm.s2sql.LLMReq;
+
+public interface SqlGeneration {
+
+    String generation(LLMReq llmReq, String modelClusterKey);
+
+}

chat/core/src/main/java/com/tencent/supersonic/chat/parser/sql/llm/SqlGenerationFactory.java

@@ -0,0 +1,22 @@
+package com.tencent.supersonic.chat.parser.sql.llm;
+
+
+import com.tencent.supersonic.chat.query.llm.s2sql.LLMReq.SqlGenerationMode;
+import java.util.Map;
+import java.util.concurrent.ConcurrentHashMap;
+
+/**
+ * Sql generation factory
+ */
+public class SqlGenerationFactory {
+
+    private static Map<SqlGenerationMode, SqlGeneration> sqlGenerationMap = new ConcurrentHashMap<>();
+
+    public static SqlGeneration get(SqlGenerationMode strategyType) {
+        return sqlGenerationMap.get(strategyType);
+    }
+
+    public static void addSqlGenerationForFactory(SqlGenerationMode strategy, SqlGeneration sqlGeneration) {
+        sqlGenerationMap.put(strategy, sqlGeneration);
+    }
+}

chat/core/src/main/java/com/tencent/supersonic/chat/parser/sql/llm/TwoStepsSqlGeneration.java

@@ -0,0 +1,73 @@
+package com.tencent.supersonic.chat.parser.sql.llm;
+
+
+import com.tencent.supersonic.chat.config.OptimizationConfig;
+import com.tencent.supersonic.chat.parser.sql.llm.prompt.OutputFormat;
+import com.tencent.supersonic.chat.parser.sql.llm.prompt.SqlExampleLoader;
+import com.tencent.supersonic.chat.parser.sql.llm.prompt.SqlPromptGenerator;
+import com.tencent.supersonic.chat.query.llm.s2sql.LLMReq;
+import com.tencent.supersonic.chat.query.llm.s2sql.LLMReq.ElementValue;
+import com.tencent.supersonic.chat.query.llm.s2sql.LLMReq.SqlGenerationMode;
+import com.tencent.supersonic.common.util.ContextUtils;
+import com.tencent.supersonic.common.util.JsonUtil;
+import dev.langchain4j.data.message.AiMessage;
+import dev.langchain4j.model.chat.ChatLanguageModel;
+import dev.langchain4j.model.input.Prompt;
+import dev.langchain4j.model.input.PromptTemplate;
+import dev.langchain4j.model.output.Response;
+import java.util.HashMap;
+import java.util.List;
+import java.util.Map;
+import lombok.extern.slf4j.Slf4j;
+import org.springframework.beans.factory.InitializingBean;
+import org.springframework.beans.factory.annotation.Autowired;
+import org.springframework.stereotype.Service;
+
+@Service
+@Slf4j
+public class TwoStepsSqlGeneration implements SqlGeneration, InitializingBean {
+
+    @Autowired
+    private ChatLanguageModel chatLanguageModel;
+
+    @Autowired
+    private SqlExampleLoader sqlExampleLoader;
+
+    @Autowired
+    private OptimizationConfig optimizationConfig;
+
+    @Override
+    public String generation(LLMReq llmReq, String modelClusterKey) {
+        String text2sqlCollectionName = optimizationConfig.getText2sqlCollectionName();
+        int text2sqlFewShotsNum = optimizationConfig.getText2sqlFewShotsNum();
+        String queryText = llmReq.getQueryText();
+
+        List<Map<String, String>> sqlExamples = sqlExampleLoader.retrieverSqlExamples(queryText, text2sqlCollectionName,
+                text2sqlFewShotsNum);
+
+        String modelName = llmReq.getSchema().getModelName();
+        List<String> fieldNameList = llmReq.getSchema().getFieldNameList();
+        List<ElementValue> linking = llmReq.getLinking();
+
+        SqlPromptGenerator sqlPromptGenerator = ContextUtils.getBean(SqlPromptGenerator.class);
+        String linkingPromptStr = sqlPromptGenerator.generateSchemaLinkingPrompt(queryText, modelName, fieldNameList,
+                linking, sqlExamples);
+
+        Prompt linkingPrompt = PromptTemplate.from(JsonUtil.toString(linkingPromptStr)).apply(new HashMap<>());
+        Response<AiMessage> linkingResult = chatLanguageModel.generate(linkingPrompt.toSystemMessage());
+
+        String schemaLinkStr = OutputFormat.schemaLinkParse(linkingResult.content().text());
+
+        String generateSqlPrompt = sqlPromptGenerator.generateSqlPrompt(queryText, modelName, schemaLinkStr,
+                llmReq.getCurrentDate(), sqlExamples);
+
+        Prompt sqlPrompt = PromptTemplate.from(JsonUtil.toString(generateSqlPrompt)).apply(new HashMap<>());
+        Response<AiMessage> sqlResult = chatLanguageModel.generate(sqlPrompt.toSystemMessage());
+        return sqlResult.content().text();
+    }
+
+    @Override
+    public void afterPropertiesSet() {
+        SqlGenerationFactory.addSqlGenerationForFactory(SqlGenerationMode.TWO_STEPS, this);
+    }
+}

chat/core/src/main/java/com/tencent/supersonic/chat/query/llm/s2sql/LLMReq.java

@@ -18,6 +18,8 @@ public class LLMReq {
 
     private String priorExts;
 
+    private SqlGenerationMode sqlGenerationMode = SqlGenerationMode.TWO_STEPS;
+
     @Data
     public static class ElementValue {
 
@@ -43,4 +45,25 @@ public class LLMReq {
 
         private String tableName;
     }
+
+    public enum SqlGenerationMode {
+
+        ONE_STEP("ONE_STEP"),
+
+        TWO_STEPS("TWO_STEPS"),
+
+        TWO_STEPS_WITH_CS("TWO_STEPS_WITH_CS");
+
+
+        private String name;
+
+        SqlGenerationMode(String name) {
+            this.name = name;
+        }
+
+        public String getName() {
+            return name;
+        }
+
+    }
 }

chat/knowledge/src/main/java/com/tencent/supersonic/knowledge/dictionary/builder/WordBuilderFactory.java

@@ -23,4 +23,4 @@ public class WordBuilderFactory {
     public static BaseWordBuilder get(DictWordType strategyType) {
         return wordNatures.get(strategyType);
     }
-}
+}

launchers/standalone/src/main/resources/application-local.yaml

@@ -53,20 +53,23 @@ langchain4j:
       temperature: 0.0
       timeout: PT60S
   #2.embedding-model
+  #2.1 in_memory
   embedding-model:
     provider: in_memory
-#  embedding-model:
+  #2.2 open_ai
-#    hugging-face:
-#      access-token: hg_access_token
-#      model-id: sentence-transformers/all-MiniLM-L6-v2
-#      timeout: 1h
-
 #  embedding-model:
 #    provider: open_ai
 #    openai:
 #      api-key: api_key
 #      modelName: all-minilm-l6-v2.onnx
 
+  #2.2 hugging_face
+#  embedding-model:
+#    provider: hugging_face
+#    hugging-face:
+#      access-token: hg_access_token
+#      model-id: sentence-transformers/all-MiniLM-L6-v2
+#      timeout: 1h
 
 #langchain4j log
 logging: