(improvement)(chat) 优化提示工程、重试机制

tencentmusic · Sep 12, 2024 · 40de8cf · 40de8cf
1 parent 2fa3bfe
commit 40de8cf
Show file tree

Hide file tree

Showing 2 changed files with 27 additions and 15 deletions.
diff --git a/...less/chat/src/main/java/com/tencent/supersonic/headless/chat/parser/llm/LLMSqlParser.java b/...less/chat/src/main/java/com/tencent/supersonic/headless/chat/parser/llm/LLMSqlParser.java
@@ -74,6 +74,11 @@ private void tryParse(ChatQueryContext queryCtx, Long dataSetId) {
             } catch (Exception e) {
                 log.error("currentRetryRound:{}, runText2SQL failed", currentRetry, e);
             }
+            Double temperature = llmReq.getModelConfig().getTemperature();
+            if(temperature == 0){
+                //报错时增加随机性，减少无效重试
+                llmReq.getModelConfig().setTemperature(0.5);
+            }
             currentRetry++;
         }
         if (MapUtils.isEmpty(sqlRespMap)) {

diff --git a/...rc/main/java/com/tencent/supersonic/headless/chat/parser/llm/OnePassSCSqlGenStrategy.java b/...rc/main/java/com/tencent/supersonic/headless/chat/parser/llm/OnePassSCSqlGenStrategy.java
@@ -5,11 +5,12 @@
 import com.tencent.supersonic.common.pojo.Text2SQLExemplar;
 import com.tencent.supersonic.headless.chat.query.llm.s2sql.LLMReq;
 import com.tencent.supersonic.headless.chat.query.llm.s2sql.LLMResp;
-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 dev.langchain4j.model.output.structured.Description;
+import dev.langchain4j.service.AiServices;
+import lombok.Data;
 import lombok.extern.slf4j.Slf4j;
 import org.apache.commons.lang3.StringUtils;
 import org.apache.commons.lang3.tuple.Pair;
@@ -24,6 +25,18 @@
 @Slf4j
 public class OnePassSCSqlGenStrategy extends SqlGenStrategy {
 
+    @Data
+    static class SemanticSql {
+        @Description("thought or remarks to tell users about the sql, make it short.")
+        private String thought;
+        @Description("sql to generate")
+        private String sql;
+    }
+
+    interface SemanticSqlExtractor {
+        SemanticSql generateSemanticSql(String text);
+    }
+
     private static final String INSTRUCTION =
             ""
                     + "\n#Role: You are a data analyst experienced in SQL languages."
@@ -36,9 +49,9 @@ public class OnePassSCSqlGenStrategy extends SqlGenStrategy {
                     + "3.DO NOT include date filter in the where clause if not explicitly expressed in the `Question`."
                     + "4.DO NOT calculate date range using functions."
                     + "5.DO NOT miss the AGGREGATE operator of metrics, always add it as needed."
-                    + "6.ONLY respond with the converted SQL statement."
                     + "\n#Exemplars:\n{{exemplar}}"
-                    + "Question:{{question}},Schema:{{schema}},SideInfo:{{information}},SQL:";
+                    + "\n#Question:"
+                    + "Question:{{question}},Schema:{{schema}},SideInfo:{{information}}";
 
     @Override
     public LLMResp generate(LLMReq llmReq) {
@@ -55,6 +68,8 @@ public LLMResp generate(LLMReq llmReq) {
             Prompt prompt = generatePrompt(llmReq, llmResp);
             prompt2Exemplar.put(prompt, exemplars);
         }
+        ChatLanguageModel chatLanguageModel = getChatLanguageModel(llmReq.getModelConfig());
+        SemanticSqlExtractor extractor = AiServices.create(SemanticSqlExtractor.class, chatLanguageModel);
 
         // 3.perform multiple self-consistency inferences parallelly
         Map<String, Prompt> output2Prompt = new ConcurrentHashMap<>();
@@ -66,18 +81,10 @@ public LLMResp generate(LLMReq llmReq) {
                             keyPipelineLog.info(
                                     "OnePassSCSqlGenStrategy reqPrompt:\n{}",
                                     prompt.toUserMessage());
-                            ChatLanguageModel chatLanguageModel =
-                                    getChatLanguageModel(llmReq.getModelConfig());
-                            Response<AiMessage> response =
-                                    chatLanguageModel.generate(prompt.toUserMessage());
-                            String sqlOutput =
-                                    StringUtils.normalizeSpace(response.content().text());
-                            // replace ```
-                            String sqlOutputFormat =
-                                    sqlOutput.replaceAll("(?s)```sql\\s*(.*?)\\s*```", "$1").trim();
-                            output2Prompt.put(sqlOutputFormat, prompt);
+                            SemanticSql s2Sql = extractor.generateSemanticSql(prompt.toUserMessage().singleText());
+                            output2Prompt.put(s2Sql.getSql(), prompt);
                             keyPipelineLog.info(
-                                    "OnePassSCSqlGenStrategy modelResp:\n{}", sqlOutputFormat);
+                                    "OnePassSCSqlGenStrategy modelResp:\n{}", s2Sql.getSql());
                         });
 
         // 4.format response.