ziin/keyboard_backend
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(chat): 新增保存润色结果向量接口并重构向量类型

Browse Source 
- ChatController 新增 /save_embed 接口，接收 ChatSaveReq 保存用户选中润色句子的向量
- 统一向量参数由 List<Float> 改为 float[]，降低 GC 压力
- 向量搜索增加 ≥0.9 相似度过滤，仅返回高置信结果
- 精简 DemoController 测试接口，下线冗余的 testSaveEmbed/testSearch
- 调整 Embedding 模型为 qwen3-embedding-4b，降低资源占用
- 放开 /chat/save_embed 匿名访问，适配前端直调
This commit is contained in:
ziin
2025-12-08 20:45:15 +08:00
parent f72781d948
commit 39b19493e2
6 changed files with 62 additions and 57 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/main/java/com/yolo/keyborad/config/LLMConfig.java
View File

@@ -53,7 +53,7 @@ public class LLMConfig {
this.openAiApi(),
MetadataMode.EMBED,
OpenAiEmbeddingOptions.builder()
.model("qwen/qwen3-embedding-8b")
.model("qwen/qwen3-embedding-4b")
.dimensions(1536)
.user("user-6")
.build(),

3
src/main/java/com/yolo/keyborad/config/SaTokenConfigure.java
View File

@@ -84,7 +84,8 @@ public class SaTokenConfigure implements WebMvcConfigurer {
"/api/apple/validate-receipt",
"/character/list",
"/user/resetPassWord",
"/chat/talk"
"/chat/talk",
"/chat/save_embed"
};
}
@Bean

18
src/main/java/com/yolo/keyborad/controller/ChatController.java
View File

@@ -1,7 +1,11 @@
package com.yolo.keyborad.controller;
import cn.dev33.satoken.stp.StpUtil;
import cn.hutool.core.util.IdUtil;
import com.yolo.keyborad.common.BaseResponse;
import com.yolo.keyborad.common.ResultUtils;
import com.yolo.keyborad.model.dto.chat.ChatReq;
import com.yolo.keyborad.model.dto.chat.ChatSaveReq;
import com.yolo.keyborad.model.dto.chat.ChatStreamMessage;
import com.yolo.keyborad.model.entity.KeyboardCharacter;
import com.yolo.keyborad.service.KeyboardCharacterService;
@@ -12,6 +16,7 @@ import io.swagger.v3.oas.annotations.tags.Tag;
import jakarta.annotation.Resource;
import lombok.extern.slf4j.Slf4j;
import org.springframework.ai.chat.client.ChatClient;
import org.springframework.ai.embedding.EmbeddingResponse;
import org.springframework.ai.openai.OpenAiChatOptions;
import org.springframework.ai.openai.OpenAiEmbeddingModel;
import org.springframework.boot.context.properties.bind.DefaultValue;
@@ -47,15 +52,11 @@ public class ChatController {
@PostMapping("/talk")
@Operation(summary = "聊天润色接口", description = "聊天润色接口")
@Parameter(name = "userInput",required = true,description = "测试聊天接口",example = "talk to something")
public Flux<ServerSentEvent<ChatStreamMessage>> testTalk(@RequestBody ChatReq chatReq){
KeyboardCharacter character = keyboardCharacterService.getById(chatReq.getCharacterId());
// 1. LLM 流式输出
Flux<ChatStreamMessage> llmFlux = client
.prompt(character.getPrompt() +
"\nUser message: %s".formatted(chatReq.getMessage()))
.prompt(character.getPrompt())
.system("""
Format rules:
- Return EXACTLY 3 replies.
@@ -97,8 +98,9 @@ public class ChatController {
@PostMapping("/save_embed")
@Operation(summary = "保存润色后的句子", description = "保存润色后的句子")
@Parameter(name = "userInput",required = true,description = "测试聊天接口",example = "talk to something")
public Flux<String> testTalkWithVector(@RequestBody ChatReq chatReq) {
return null;
public BaseResponse<Boolean> testTalkWithVector(@RequestBody ChatSaveReq chatSaveReq) {
float[] embed = embeddingModel.embed(chatSaveReq.getUserInputMessage());
qdrantVectorService.upsertPoint(IdUtil.getSnowflakeNextId(), embed, chatSaveReq.getUserSelectMessage());
return ResultUtils.success(true);
}
}

30
src/main/java/com/yolo/keyborad/controller/DemoController.java
View File

@@ -97,23 +97,23 @@ public class DemoController {
}
@PostMapping("/testSaveEmbed")
@Operation(summary = "测试存储向量接口", description = "测试存储向量接口")
@Parameter(name = "userInput",required = true,description = "测试存储向量接口")
public BaseResponse<Boolean> testSaveEmbed(@RequestBody EmbedSaveReq embedSaveReq) {
qdrantVectorService.upsertPoint(embedSaveReq.getRecordItem().getId()
, embedSaveReq.getVector()
, JSONUtil.toJsonStr(embedSaveReq.getRecordItem()));
return ResultUtils.success(true);
}
// @PostMapping("/testSaveEmbed")
// @Operation(summary = "测试存储向量接口", description = "测试存储向量接口")
// @Parameter(name = "userInput",required = true,description = "测试存储向量接口")
// public BaseResponse<Boolean> testSaveEmbed(@RequestBody EmbedSaveReq embedSaveReq) {
// qdrantVectorService.upsertPoint(embedSaveReq.getRecordItem().getId()
// , embedSaveReq.getVector()
// , JSONUtil.toJsonStr(embedSaveReq.getRecordItem()));
// return ResultUtils.success(true);
// }
@PostMapping("/testSearch")
@Operation(summary = "测试搜索向量接口", description = "测试搜索向量接口")
@Parameter(name = "userInput",required = true,description = "测试搜索向量接口")
public BaseResponse<List<QdrantSearchItem>> testSearch(@RequestBody SearchEmbedReq searchEmbedReq) {
return ResultUtils.success(qdrantVectorService.searchPoint(searchEmbedReq.getUserInputEmbed(), 3));
}
// @PostMapping("/testSearch")
// @Operation(summary = "测试搜索向量接口", description = "测试搜索向量接口")
// @Parameter(name = "userInput",required = true,description = "测试搜索向量接口")
// public BaseResponse<List<QdrantSearchItem>> testSearch(@RequestBody SearchEmbedReq searchEmbedReq) {
// return ResultUtils.success(qdrantVectorService.searchPoint(searchEmbedReq.getUserInputEmbed(), 3));
// }
@PostMapping("/tsetSearchText")

16
src/main/java/com/yolo/keyborad/model/dto/chat/ChatSaveReq.java Normal file
View File

@@ -0,0 +1,16 @@
package com.yolo.keyborad.model.dto.chat;
import lombok.Data;
/*
* @author: ziin
* @date: 2025/12/8 19:26
*/
@Data
public class ChatSaveReq {
private String userInputMessage;
private String userSelectMessage;
}

50
src/main/java/com/yolo/keyborad/service/impl/QdrantVectorService.java
View File

@@ -13,6 +13,7 @@ import org.springframework.ai.embedding.EmbeddingModel;
import org.springframework.ai.embedding.EmbeddingResponse;
import org.springframework.stereotype.Service;
import java.util.Arrays;
import java.util.List;
import java.util.Map;
import java.util.concurrent.ExecutionException;
@@ -43,7 +44,7 @@ public class QdrantVectorService {
* @param vector 向量（和 collection 中定义的 size 一致）
* @param payload 额外信息，例如原文、标题、userId 等
*/
public void upsertPoint(long id, List<Float> vector,String payload){
public void upsertPoint(long id, float[] vector,String payload){
try {
qdrantClient.upsertAsync(
@@ -90,12 +91,12 @@ public class QdrantVectorService {
// }
// }
public List<QdrantSearchItem> searchPoint(List<Float> userVector, int limit) {
public List<QdrantSearchItem> searchPoint(float[] userVector, int limit) {
try {
Points.QueryPoints query = Points.QueryPoints.newBuilder()
.setCollectionName(COLLECTION_NAME) // ★ 必须
.setQuery(nearest(userVector)) // ★ 语义向量
.setLimit(limit) // TopK
.setLimit(limit) // 限制返回数量
.setWithPayload(enable(true)) // ★ 带上 payload
.build();
@@ -107,37 +108,31 @@ public class QdrantVectorService {
// 3. 把 Protobuf 的 ScoredPoint 转成你的 DTO
return batchResult.getResultList().stream()
.filter(p -> p.getScore() >= 0.9) // ★ 只要相似度 ≥ 90%
.map(p -> {
QdrantSearchItem item = new QdrantSearchItem();
// id：你插入时用的是 setId(id(id))，所以这里取 num
if (p.getId().hasNum()) {
item.setId(p.getId().getNum());
}
// score
item.setScore(p.getScore());
// payload：你之前是 putAllPayload(Map.of("payload", value(payload)))
// 这里从 Struct 里拿 "payload" 字段
var fieldsMap = p.getPayloadMap();
var payloadValue = fieldsMap.get("payload");
if (payloadValue != null && payloadValue.hasStringValue()) {
item.setPayload(payloadValue.getStringValue());
}
// vector：你插入时用的是 vectors(vector)，即 unnamed 单向量
// proto 结构一般是 Vectors.vector.data[]
if (p.getVectors().hasVector()) {
List<Float> vec = p.getVectors().getVector().getDataList();
item.setVector(vec);
item.setVector(p.getVectors().getVector().getDataList());
}
return item;
})
.toList();
} catch (InterruptedException | ExecutionException e) {
log.error("search point 失败", e);
throw new BusinessException(ErrorCode.OPERATION_ERROR);
@@ -149,38 +144,29 @@ public class QdrantVectorService {
/**
* 把一段文本做 embedding 然后写入 Qdrant
*
* @param id 业务 ID（比如业务表主键）
* @param text 用来做向量的文本（一般是内容）
*/
public void indexText(long id, String text) {
// 1. 文本 → 向量
List<Float> vector = embedTextToVector(text);
// public void indexText(long id, String text) {
// // 1. 文本 → 向量
// embedTextToVector(text);
//
// // 2. 存到 Qdrant，payload 里顺便存原文
// upsertPoint(id, vector, text);
// }
// 2. 存到 Qdrant，payload 里顺便存原文
upsertPoint(id, vector, text);
}
private float[] embedTextToVector(String text) {
private List<Float> embedTextToVector(String text) {
EmbeddingResponse response = embeddingModel.embedForResponse(List.of(text));
Embedding embedding = response.getResult(); // 就一条
// Spring AI 里一般是 List<Double>
float[] output = embedding.getOutput();
// 转成 Qdrant 需要的 List<Float>
return Floats.asList(output);
return embeddingModel.embed(text);
}
public List<QdrantSearchItem> searchText(String userInput) {
long t0 = System.currentTimeMillis();
List<Float> floats = this.embedTextToVector(userInput);
float[] floats = this.embedTextToVector(userInput);
long t1 = System.currentTimeMillis();
List<QdrantSearchItem> qdrantSearchItems = this.searchPoint(floats, 3);
List<QdrantSearchItem> qdrantSearchItems = this.searchPoint(floats, 1);
long t2 = System.currentTimeMillis();
log.info("embedding = {} ms, qdrant = {} ms, total = {} ms",