add new chat corrector

在助理最终执行物理SQL前，加入一步LLM优化性能功能

chat/server/src/main/java/com/tencent/supersonic/chat/server/executor/SqlExecutor.java

@@ -75,8 +75,12 @@ public class SqlExecutor implements ChatQueryExecutor {
             return null;
         }
 
-        QuerySqlReq sqlReq =
+        // 使用querySQL，它已经包含了所有修正（包括物理SQL修正）
-                QuerySqlReq.builder().sql(parseInfo.getSqlInfo().getCorrectedS2SQL()).build();
+        String finalSql = StringUtils.isNotBlank(parseInfo.getSqlInfo().getQuerySQL())
+                ? parseInfo.getSqlInfo().getQuerySQL()
+                : parseInfo.getSqlInfo().getCorrectedS2SQL();
+
+        QuerySqlReq sqlReq = QuerySqlReq.builder().sql(finalSql).build();
         sqlReq.setSqlInfo(parseInfo.getSqlInfo());
         sqlReq.setDataSetId(parseInfo.getDataSetId());
 
@@ -90,7 +94,7 @@ public class SqlExecutor implements ChatQueryExecutor {
         queryResult.setQueryTimeCost(System.currentTimeMillis() - startTime);
         if (queryResp != null) {
             queryResult.setQueryAuthorization(queryResp.getQueryAuthorization());
-            queryResult.setQuerySql(queryResp.getSql());
+            queryResult.setQuerySql(finalSql);
             queryResult.setQueryResults(queryResp.getResultList());
             queryResult.setQueryColumns(queryResp.getColumns());
             queryResult.setQueryState(QueryState.SUCCESS);

chat/server/src/main/java/com/tencent/supersonic/chat/server/service/impl/AgentServiceImpl.java

@@ -22,6 +22,7 @@ import lombok.extern.slf4j.Slf4j;
 import org.springframework.beans.BeanUtils;
 import org.springframework.beans.factory.annotation.Autowired;
 import org.springframework.beans.factory.annotation.Qualifier;
+import org.springframework.context.annotation.Lazy;
 import org.springframework.stereotype.Service;
 import org.springframework.util.CollectionUtils;
 
@@ -39,6 +40,7 @@ public class AgentServiceImpl extends ServiceImpl<AgentDOMapper, AgentDO> implem
     private MemoryService memoryService;
 
     @Autowired
+    @Lazy
     private ChatQueryService chatQueryService;
 
     @Autowired

chat/server/src/main/java/com/tencent/supersonic/chat/server/service/impl/ChatQueryServiceImpl.java

@@ -49,6 +49,7 @@ import net.sf.jsqlparser.schema.Column;
 import org.apache.commons.lang3.StringUtils;
 import org.apache.commons.lang3.tuple.Pair;
 import org.springframework.beans.factory.annotation.Autowired;
+import org.springframework.context.annotation.Lazy;
 import org.springframework.stereotype.Service;
 import org.springframework.util.CollectionUtils;
 
@@ -66,6 +67,7 @@ public class ChatQueryServiceImpl implements ChatQueryService {
     @Autowired
     private SemanticLayerService semanticLayerService;
     @Autowired
+    @Lazy
     private AgentService agentService;
 
     private final List<ChatQueryParser> chatQueryParsers = ComponentFactory.getChatParsers();

headless/api/src/main/java/com/tencent/supersonic/headless/api/pojo/SqlInfo.java

@@ -16,4 +16,7 @@ public class SqlInfo implements Serializable {
 
     // SQL to be executed finally
     private String querySQL;
+
+    // Physical SQL corrected by LLM for performance optimization
+    private String correctedQuerySQL;
 }

headless/api/src/main/java/com/tencent/supersonic/headless/api/pojo/enums/ChatWorkflowState.java

@@ -8,5 +8,6 @@ public enum ChatWorkflowState {
     VALIDATING,
     SQL_CORRECTING,
     PROCESSING,
+    PHYSICAL_SQL_CORRECTING,
     FINISHED
 }

headless/chat/src/main/java/com/tencent/supersonic/headless/chat/corrector/LLMPhysicalSqlCorrector.java

@@ -0,0 +1,97 @@
+package com.tencent.supersonic.headless.chat.corrector;
+
+import com.tencent.supersonic.common.pojo.ChatApp;
+import com.tencent.supersonic.common.pojo.enums.AppModule;
+import com.tencent.supersonic.common.util.ChatAppManager;
+import com.tencent.supersonic.headless.api.pojo.SemanticParseInfo;
+import com.tencent.supersonic.headless.chat.ChatQueryContext;
+import dev.langchain4j.model.chat.ChatLanguageModel;
+import dev.langchain4j.model.input.Prompt;
+import dev.langchain4j.model.input.PromptTemplate;
+import dev.langchain4j.model.output.structured.Description;
+import dev.langchain4j.provider.ModelProvider;
+import dev.langchain4j.service.AiServices;
+import lombok.Data;
+import lombok.ToString;
+import lombok.extern.slf4j.Slf4j;
+import org.apache.commons.lang3.StringUtils;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+import java.util.HashMap;
+import java.util.Map;
+import java.util.Objects;
+
+/**
+ * 物理SQL修正器 - 使用LLM优化物理SQL性能
+ */
+@Slf4j
+public class LLMPhysicalSqlCorrector extends BaseSemanticCorrector {
+
+    private static final Logger keyPipelineLog = LoggerFactory.getLogger("keyPipeline");
+
+    public static final String APP_KEY = "PHYSICAL_SQL_CORRECTOR";
+    private static final String INSTRUCTION = ""
+            + "#Role: You are a senior database performance optimization expert experienced in SQL tuning."
+            + "\n\n#Task: You will be provided with a user question and the corresponding physical SQL query,"
+            + " please analyze and optimize this SQL to improve query performance." + "\n\n#Rules:"
+            + "\n1. ALWAYS add appropriate index hints if the database supports them."
+            + "\n2. Optimize JOIN order by placing smaller tables first."
+            + "\n3. Add reasonable query limits to prevent large result sets if no LIMIT exists."
+            + "\n4. Optimize WHERE condition order by placing high-selectivity conditions first."
+            + "\n5. Ensure the optimized SQL is syntactically correct and logically equivalent."
+            + "\n6. If the SQL is already well-optimized, return 'positive'."
+            + "\n\n#Question: {{question}}" + "\n\n#OriginalSQL: {{sql}}" + "\n\n#Response:";
+
+    public LLMPhysicalSqlCorrector() {
+        ChatAppManager.register(APP_KEY, ChatApp.builder().prompt(INSTRUCTION).name("物理SQL修正")
+                .appModule(AppModule.CHAT).description("通过大模型对物理SQL做性能优化").enable(false).build());
+    }
+
+    @Data
+    @ToString
+    static class PhysicalSql {
+        @Description("either positive or negative")
+        private String opinion;
+
+        @Description("optimized sql if negative")
+        private String sql;
+    }
+
+    interface PhysicalSqlExtractor {
+        PhysicalSql generatePhysicalSql(String text);
+    }
+
+    @Override
+    public void doCorrect(ChatQueryContext chatQueryContext, SemanticParseInfo semanticParseInfo) {
+        ChatApp chatApp = chatQueryContext.getRequest().getChatAppConfig().get(APP_KEY);
+        if (!chatQueryContext.getRequest().getText2SQLType().enableLLM() || Objects.isNull(chatApp)
+                || !chatApp.isEnable()) {
+            return;
+        }
+
+        ChatLanguageModel chatLanguageModel =
+                ModelProvider.getChatModel(chatApp.getChatModelConfig());
+        PhysicalSqlExtractor extractor =
+                AiServices.create(PhysicalSqlExtractor.class, chatLanguageModel);
+        Prompt prompt = generatePrompt(chatQueryContext.getRequest().getQueryText(),
+                semanticParseInfo, chatApp.getPrompt());
+        PhysicalSql physicalSql =
+                extractor.generatePhysicalSql(prompt.toUserMessage().singleText());
+        keyPipelineLog.info("LLMPhysicalSqlCorrector modelReq:\n{} \nmodelResp:\n{}", prompt.text(),
+                physicalSql);
+        if ("NEGATIVE".equalsIgnoreCase(physicalSql.getOpinion())
+                && StringUtils.isNotBlank(physicalSql.getSql())) {
+            semanticParseInfo.getSqlInfo().setCorrectedQuerySQL(physicalSql.getSql());
+        }
+    }
+
+    private Prompt generatePrompt(String queryText, SemanticParseInfo semanticParseInfo,
+            String promptTemplate) {
+        Map<String, Object> variable = new HashMap<>();
+        variable.put("question", queryText);
+        variable.put("sql", semanticParseInfo.getSqlInfo().getQuerySQL());
+
+        return PromptTemplate.from(promptTemplate).apply(variable);
+    }
+}

headless/chat/src/main/java/com/tencent/supersonic/headless/chat/parser/llm/PromptHelper.java

@@ -52,7 +52,8 @@ public class PromptHelper {
         for (int i = 0; i < selfConsistencyNumber; i++) {
             List<Text2SQLExemplar> shuffledList = new ArrayList<>(exemplars);
             // only shuffle the exemplars from config
-            List<Text2SQLExemplar> subList=shuffledList.subList(llmReq.getDynamicExemplars().size(),shuffledList.size());
+            List<Text2SQLExemplar> subList =
+                    shuffledList.subList(llmReq.getDynamicExemplars().size(), shuffledList.size());
             Collections.shuffle(subList);
             results.add(shuffledList.subList(0, Math.min(shuffledList.size(), fewShotNumber)));
         }

headless/server/src/main/java/com/tencent/supersonic/headless/server/utils/ChatWorkflowEngine.java

@@ -8,6 +8,7 @@ import com.tencent.supersonic.headless.api.pojo.request.SemanticQueryReq;
 import com.tencent.supersonic.headless.api.pojo.response.ParseResp;
 import com.tencent.supersonic.headless.api.pojo.response.SemanticTranslateResp;
 import com.tencent.supersonic.headless.chat.ChatQueryContext;
+import com.tencent.supersonic.headless.chat.corrector.LLMPhysicalSqlCorrector;
 import com.tencent.supersonic.headless.chat.corrector.SemanticCorrector;
 import com.tencent.supersonic.headless.chat.mapper.SchemaMapper;
 import com.tencent.supersonic.headless.chat.parser.SemanticParser;
@@ -76,6 +77,10 @@ public class ChatWorkflowEngine {
                     long start = System.currentTimeMillis();
                     performTranslating(queryCtx, parseResult);
                     parseResult.getParseTimeCost().setSqlTime(System.currentTimeMillis() - start);
+                    queryCtx.setChatWorkflowState(ChatWorkflowState.PHYSICAL_SQL_CORRECTING);
+                    break;
+                case PHYSICAL_SQL_CORRECTING:
+                    performPhysicalSqlCorrecting(queryCtx);
                     queryCtx.setChatWorkflowState(ChatWorkflowState.FINISHED);
                     break;
                 default:
@@ -162,4 +167,25 @@ public class ChatWorkflowEngine {
             parseResult.setErrorMsg(String.join("\n", errorMsg));
         }
     }
+
+    private void performPhysicalSqlCorrecting(ChatQueryContext queryCtx) {
+        List<SemanticQuery> candidateQueries = queryCtx.getCandidateQueries();
+        if (CollectionUtils.isNotEmpty(candidateQueries)) {
+            for (SemanticQuery semanticQuery : candidateQueries) {
+                for (SemanticCorrector corrector : semanticCorrectors) {
+                    if (corrector instanceof LLMPhysicalSqlCorrector) {
+                        corrector.correct(queryCtx, semanticQuery.getParseInfo());
+                        // 如果物理SQL被修正了，更新querySQL为修正后的版本
+                        SemanticParseInfo parseInfo = semanticQuery.getParseInfo();
+                        if (StringUtils.isNotBlank(parseInfo.getSqlInfo().getCorrectedQuerySQL())) {
+                            parseInfo.getSqlInfo().setQuerySQL(parseInfo.getSqlInfo().getCorrectedQuerySQL());
+                            log.info("Physical SQL corrected and updated querySQL: {}", 
+                                    parseInfo.getSqlInfo().getQuerySQL());
+                        }
+                        break;
+                    }
+                }
+            }
+        }
+    }
 }

launchers/headless/src/main/resources/META-INF/spring.factories

@@ -14,7 +14,8 @@ com.tencent.supersonic.headless.chat.parser.SemanticParser=\
 
 com.tencent.supersonic.headless.chat.corrector.SemanticCorrector=\
     com.tencent.supersonic.headless.chat.corrector.RuleSqlCorrector,\
-    com.tencent.supersonic.headless.chat.corrector.LLMSqlCorrector
+    com.tencent.supersonic.headless.chat.corrector.LLMSqlCorrector,\
+    com.tencent.supersonic.headless.chat.corrector.LLMPhysicalSqlCorrector
 
 com.tencent.supersonic.headless.chat.knowledge.file.FileHandler=\
     com.tencent.supersonic.headless.chat.knowledge.file.FileHandlerImpl

launchers/standalone/src/main/resources/META-INF/spring.factories

@@ -15,7 +15,8 @@ com.tencent.supersonic.headless.chat.parser.SemanticParser=\
 
 com.tencent.supersonic.headless.chat.corrector.SemanticCorrector=\
     com.tencent.supersonic.headless.chat.corrector.RuleSqlCorrector,\
-    com.tencent.supersonic.headless.chat.corrector.LLMSqlCorrector
+    com.tencent.supersonic.headless.chat.corrector.LLMSqlCorrector,\
+    com.tencent.supersonic.headless.chat.corrector.LLMPhysicalSqlCorrector
 
 com.tencent.supersonic.headless.chat.knowledge.file.FileHandler=\
     com.tencent.supersonic.headless.chat.knowledge.file.FileHandlerImpl

pom.xml

@@ -10,9 +10,9 @@
     <version>${revision}</version>
 
     <modules>
+        <module>common</module>
         <module>auth</module>
         <module>chat</module>
-        <module>common</module>
         <module>launchers</module>
         <module>headless</module>
     </modules>
@@ -31,6 +31,7 @@
         <java.target.version>21</java.target.version>
         <maven.compiler.source>21</maven.compiler.source>
         <maven.compiler.target>21</maven.compiler.target>
+        <maven.compiler.release>21</maven.compiler.release>
         <file.encoding>UTF-8</file.encoding>
         <jsqlparser.version>4.9</jsqlparser.version>
         <pagehelper.version>6.1.0</pagehelper.version>
@@ -254,6 +255,13 @@
                     <source>${java.source.version}</source>
                     <target>${java.target.version}</target>
                     <encoding>${file.encoding}</encoding>
+                    <annotationProcessorPaths>
+                        <path>
+                            <groupId>org.projectlombok</groupId>
+                            <artifactId>lombok</artifactId>
+                            <version>${lombok.version}</version>
+                        </path>
+                    </annotationProcessorPaths>
                 </configuration>
             </plugin>
             <plugin>

webapp/package.json

@@ -0,0 +1,5 @@
+{
+	"dependencies": {
+		"supersonic-chat-sdk": "link:packages/chat-sdk"
+	}
+}

webapp/packages/chat-sdk/src/common/type.ts

@@ -79,6 +79,7 @@ export type SqlInfoType = {
   parsedS2SQL: string;
   correctedS2SQL: string;
   querySQL: string;
+  correctedQuerySQL?: string;
 };
 
 export type ChatContextType = {

webapp/packages/chat-sdk/src/components/ChatItem/SqlItem.tsx

@@ -58,32 +58,28 @@ const SqlItem: React.FC<Props> = ({
   const getSchemaMapText = () => {
     return `
 Schema映射
-${schema?.fieldNameList?.length > 0 ? `名称：${schema.fieldNameList.join('、')}` : ''}${
+${schema?.fieldNameList?.length > 0 ? `名称：${schema.fieldNameList.join('、')}` : ''}${schema?.values?.length > 0
-      schema?.values?.length > 0
         ? `
 取值：${schema.values
-            .map((item: any) => {
+          .map((item: any) => {
-              return `${item.fieldName}: ${item.fieldValue}`;
+            return `${item.fieldName}: ${item.fieldValue}`;
-            })
+          })
-            .join('、')}`
+          .join('、')}`
         : ''
-    }${
+      }${priorExts
-      priorExts
         ? `
 附加：${priorExts}`
         : ''
-    }${
+      }${terms?.length > 0
-      terms?.length > 0
         ? `
 术语：${terms
-            .map((item: any) => {
+          .map((item: any) => {
-              return `${item.name}${item.alias?.length > 0 ? `(${item.alias.join(',')})` : ''}: ${
+            return `${item.name}${item.alias?.length > 0 ? `(${item.alias.join(',')})` : ''}: ${item.description
-                item.description
               }`;
-            })
+          })
-            .join('、')}`
+          .join('、')}`
         : ''
-    }
+      }
 
 `;
   };
@@ -91,16 +87,16 @@ ${schema?.fieldNameList?.length > 0 ? `名称：${schema.fieldNameList.join('、
   const getFewShotText = () => {
     return `
 Few-shot示例${fewShots
-      .map((item: any, index: number) => {
+        .map((item: any, index: number) => {
-        return `
+          return `
 
 示例${index + 1}：
 问题：${item.question}
 SQL：
 ${format(item.sql)}
 `;
-      })
+        })
-      .join('')}
+        .join('')}
 `;
   };
 
@@ -120,6 +116,14 @@ ${format(sqlInfo.correctedS2SQL)}
 `;
   };
 
+  const getCorrectedQuerySQLText = () => {
+    return `
+物理SQL修正
+
+${format(sqlInfo.correctedQuerySQL || '')}
+`;
+  };
+
   const getQuerySQLText = () => {
     return `
 最终执行SQL
@@ -155,6 +159,9 @@ ${executeErrorMsg}
     if (sqlInfo.correctedS2SQL) {
       text += getCorrectedS2SQLText();
     }
+    if (sqlInfo.correctedQuerySQL) {
+      text += getCorrectedQuerySQLText();
+    }
     if (sqlInfo.querySQL) {
       text += getQuerySQLText();
     }
@@ -183,9 +190,8 @@ ${executeErrorMsg}
         <div className={`${tipPrefixCls}-content-options`}>
           {llmReq && (
             <div
-              className={`${tipPrefixCls}-content-option ${
+              className={`${tipPrefixCls}-content-option ${sqlType === 'schemaMap' ? `${tipPrefixCls}-content-option-active` : ''
-                sqlType === 'schemaMap' ? `${tipPrefixCls}-content-option-active` : ''
+                }`}
-              }`}
               onClick={() => {
                 setSqlType(sqlType === 'schemaMap' ? '' : 'schemaMap');
               }}
@@ -195,9 +201,8 @@ ${executeErrorMsg}
           )}
           {fewShots.length > 0 && (
             <div
-              className={`${tipPrefixCls}-content-option ${
+              className={`${tipPrefixCls}-content-option ${sqlType === 'fewShots' ? `${tipPrefixCls}-content-option-active` : ''
-                sqlType === 'fewShots' ? `${tipPrefixCls}-content-option-active` : ''
+                }`}
-              }`}
               onClick={() => {
                 setSqlType(sqlType === 'fewShots' ? '' : 'fewShots');
               }}
@@ -207,9 +212,8 @@ ${executeErrorMsg}
           )}
           {sqlInfo.parsedS2SQL && (
             <div
-              className={`${tipPrefixCls}-content-option ${
+              className={`${tipPrefixCls}-content-option ${sqlType === 'parsedS2SQL' ? `${tipPrefixCls}-content-option-active` : ''
-                sqlType === 'parsedS2SQL' ? `${tipPrefixCls}-content-option-active` : ''
+                }`}
-              }`}
               onClick={() => {
                 setSqlType(sqlType === 'parsedS2SQL' ? '' : 'parsedS2SQL');
               }}
@@ -219,9 +223,8 @@ ${executeErrorMsg}
           )}
           {sqlInfo.correctedS2SQL && (
             <div
-              className={`${tipPrefixCls}-content-option ${
+              className={`${tipPrefixCls}-content-option ${sqlType === 'correctedS2SQL' ? `${tipPrefixCls}-content-option-active` : ''
-                sqlType === 'correctedS2SQL' ? `${tipPrefixCls}-content-option-active` : ''
+                }`}
-              }`}
               onClick={() => {
                 setSqlType(sqlType === 'correctedS2SQL' ? '' : 'correctedS2SQL');
               }}
@@ -229,16 +232,26 @@ ${executeErrorMsg}
               修正S2SQL
             </div>
           )}
+          {sqlInfo.correctedQuerySQL && (
+            <div
+              className={`${tipPrefixCls}-content-option ${sqlType === 'correctedQuerySQL' ? `${tipPrefixCls}-content-option-active` : ''
+                }`}
+              onClick={() => {
+                setSqlType(sqlType === 'correctedQuerySQL' ? '' : 'correctedQuerySQL');
+              }}
+            >
+              物理SQL修正
+            </div>
+          )}
           {sqlInfo.querySQL && (
             <div
-              className={`${tipPrefixCls}-content-option ${
+              className={`${tipPrefixCls}-content-option ${sqlType === 'querySQL' ? `${tipPrefixCls}-content-option-active` : ''
-                sqlType === 'querySQL' ? `${tipPrefixCls}-content-option-active` : ''
+                }`}
-              }`}
               onClick={() => {
                 setSqlType(sqlType === 'querySQL' ? '' : 'querySQL');
               }}
             >
-              最终执行SQL
+              {sqlInfo.correctedQuerySQL ? '最终执行SQL' : '最终执行SQL'}
             </div>
           )}
           <Button className={`${prefixCls}-export-log`} size="small" onClick={onExportLog}>
@@ -248,13 +261,12 @@ ${executeErrorMsg}
         </div>
       </div>
       <div
-        className={`${prefixCls} ${
+        className={`${prefixCls} ${!window.location.pathname.includes('/chat') &&
-          !window.location.pathname.includes('/chat') &&
           integrateSystem &&
           integrateSystem !== 'wiki'
-            ? `${prefixCls}-copilot`
+          ? `${prefixCls}-copilot`
-            : ''
+          : ''
-        }`}
+          }`}
       >
         {sqlType === 'schemaMap' && (
           <div className={`${prefixCls}-code`}>
@@ -290,9 +302,8 @@ ${executeErrorMsg}
                 <div className={`${prefixCls}-schema-content`}>
                   {terms
                     .map((item: any) => {
-                      return `${item.name}${
+                      return `${item.name}${item.alias?.length > 0 ? `(${item.alias.join(',')})` : ''
-                        item.alias?.length > 0 ? `(${item.alias.join(',')})` : ''
+                        }: ${item.description}`;
-                      }: ${item.description}`;
                     })
                     .join('、')}
                 </div>