package com.custom.webui.service.impl;
|
|
import com.custom.webui.config.props.RagProperties;
|
import com.custom.webui.dto.resp.OcrResultDTO;
|
import com.custom.webui.dto.resp.RagChunk;
|
import com.custom.webui.exception.BusinessException;
|
import com.custom.webui.service.RagProxyService;
|
import com.fasterxml.jackson.databind.JsonNode;
|
import com.fasterxml.jackson.databind.ObjectMapper;
|
import com.fasterxml.jackson.databind.node.ObjectNode;
|
import lombok.extern.slf4j.Slf4j;
|
import org.springframework.beans.factory.annotation.Qualifier;
|
import org.springframework.core.io.FileSystemResource;
|
import org.springframework.http.HttpEntity;
|
import org.springframework.http.HttpHeaders;
|
import org.springframework.http.MediaType;
|
import org.springframework.http.ResponseEntity;
|
import org.springframework.stereotype.Service;
|
import org.springframework.util.LinkedMultiValueMap;
|
import org.springframework.util.MultiValueMap;
|
import org.springframework.http.client.SimpleClientHttpRequestFactory;
|
import org.springframework.web.client.RestTemplate;
|
|
import java.nio.file.Files;
|
import java.nio.file.Path;
|
import java.util.ArrayList;
|
import java.util.Collections;
|
import java.util.List;
|
|
/**
|
* RAG 微服务代理实现:通过 HTTP 调用 Python FastAPI 服务。
|
*
|
* <p>接口对应关系:
|
* <pre>
|
* POST /api/rag/ingest (multipart) -> Python POST /api/rag/ingest
|
* POST /api/rag/search (json) -> Python POST /api/rag/search
|
* DELETE /api/rag/collection -> Python DELETE /api/rag/collection/{name}
|
* </pre>
|
*
|
* <p><b>降级策略</b>:检索失败返回空列表而非抛异常。
|
* 理由:知识库是「增强」能力,它挂掉不应该让用户连话都聊不了。
|
* 而入库失败必须抛出 —— 用户明确点了「向量化」,静默失败会让人以为成功了。
|
*/
|
@Slf4j
|
@Service
|
public class RagProxyServiceImpl implements RagProxyService {
|
|
private final RagProperties ragProperties;
|
private final RestTemplate restTemplate;
|
private final ObjectMapper objectMapper;
|
|
public RagProxyServiceImpl(RagProperties ragProperties,
|
@Qualifier("ragRestTemplate") RestTemplate restTemplate,
|
ObjectMapper objectMapper) {
|
this.ragProperties = ragProperties;
|
this.restTemplate = restTemplate;
|
this.objectMapper = objectMapper;
|
log.info("RAG 代理初始化 enabled={} baseUrl={}",
|
ragProperties.isEnabled(), ragProperties.getBaseUrl());
|
}
|
|
@Override
|
public IngestResult ingest(Path filePath, String collectionName, Long userId) {
|
if (!ragProperties.isEnabled()) {
|
throw BusinessException.badRequest("RAG 功能未启用");
|
}
|
if (filePath == null || !Files.exists(filePath)) {
|
throw BusinessException.badRequest("待入库文件不存在: " + filePath);
|
}
|
|
// 用 multipart/form-data 传文件本体 + 集合名 + 用户 id
|
MultiValueMap<String, Object> body = new LinkedMultiValueMap<>();
|
body.add("file", new FileSystemResource(filePath));
|
body.add("collection_name", collectionName);
|
body.add("user_id", String.valueOf(userId));
|
|
HttpHeaders headers = new HttpHeaders();
|
headers.setContentType(MediaType.MULTIPART_FORM_DATA);
|
|
String url = baseUrl() + "/api/rag/ingest";
|
try {
|
ResponseEntity<String> response = restTemplate.postForEntity(
|
url, new HttpEntity<>(body, headers), String.class);
|
|
JsonNode root = objectMapper.readTree(response.getBody());
|
JsonNode data = unwrapData(root);
|
|
String returnedCollection = text(data, "collection_name", collectionName);
|
int chunkCount = data.path("chunk_count").asInt(0);
|
|
log.info("RAG 入库完成 collection={} chunks={}", returnedCollection, chunkCount);
|
return new IngestResult(returnedCollection, chunkCount, "入库成功,共 " + chunkCount + " 个片段");
|
|
} catch (BusinessException e) {
|
throw e;
|
} catch (Exception e) {
|
log.error("调用 RAG 入库接口失败 url={}", url, e);
|
throw BusinessException.serverError(
|
"调用 RAG 服务失败(" + url + ")。请确认 Python 服务已启动:"
|
+ "uvicorn main:app --host 0.0.0.0 --port 8000。原因: " + rootMessage(e));
|
}
|
}
|
|
@Override
|
public List<RagChunk> search(String query, List<String> collectionNames, int topK) {
|
if (!ragProperties.isEnabled() || collectionNames == null || collectionNames.isEmpty()) {
|
return Collections.emptyList();
|
}
|
if (query == null || query.isBlank()) {
|
return Collections.emptyList();
|
}
|
|
ObjectNode payload = objectMapper.createObjectNode();
|
payload.put("query", query);
|
payload.put("top_k", topK > 0 ? topK : ragProperties.getDefaultTopK());
|
var array = payload.putArray("collection_names");
|
collectionNames.forEach(array::add);
|
|
String url = baseUrl() + "/api/rag/search";
|
try {
|
HttpHeaders headers = new HttpHeaders();
|
headers.setContentType(MediaType.APPLICATION_JSON);
|
|
ResponseEntity<String> response = restTemplate.postForEntity(
|
url, new HttpEntity<>(objectMapper.writeValueAsString(payload), headers), String.class);
|
|
JsonNode root = objectMapper.readTree(response.getBody());
|
JsonNode data = unwrapData(root);
|
JsonNode chunksNode = data.path("chunks");
|
|
List<RagChunk> chunks = new ArrayList<>();
|
if (chunksNode.isArray()) {
|
for (JsonNode node : chunksNode) {
|
RagChunk chunk = new RagChunk();
|
chunk.setText(text(node, "text", ""));
|
chunk.setFilename(text(node, "filename", "未知文档"));
|
// 章节路径由 Python 分块时标注并存入向量元数据,这里必须一并取出,
|
// 否则引用信息会退化成只有文件名,无法定位到文档的哪一节
|
chunk.setSection(text(node, "section", null));
|
chunk.setCollectionName(text(node, "collection_name", null));
|
chunk.setChunkIndex(node.path("chunk_index").isMissingNode()
|
? null : node.path("chunk_index").asInt());
|
if (node.hasNonNull("score")) {
|
chunk.setScore(node.get("score").asDouble());
|
}
|
// 过滤掉空片段,避免把无意义的空白拼进 prompt
|
if (!chunk.getText().isBlank()) {
|
chunks.add(chunk);
|
}
|
}
|
}
|
|
log.debug("RAG 检索完成 query={} collections={} 命中 {} 条",
|
abbreviate(query, 50), collectionNames, chunks.size());
|
return chunks;
|
|
} catch (Exception e) {
|
// 降级:检索失败不阻断对话
|
log.warn("调用 RAG 检索接口失败,本次降级为纯对话 url={} err={}", url, rootMessage(e));
|
return Collections.emptyList();
|
}
|
}
|
|
@Override
|
public String buildRagContext(List<RagChunk> chunks) {
|
if (chunks == null || chunks.isEmpty()) {
|
return null;
|
}
|
|
StringBuilder sb = new StringBuilder();
|
int index = 1;
|
int maxChars = ragProperties.getMaxContextChars();
|
|
for (RagChunk chunk : chunks) {
|
// 用与前端完全一致的序号做前缀,让模型能引用 [1][2],
|
// 用户在前端看到的「引用来源」列表也是同一套编号,两边可以对齐。
|
// 来源信息里带上章节(如果有),模型据此可以写出更精确的引用。
|
String source = chunk.getFilename() == null ? "未知文档" : chunk.getFilename();
|
if (chunk.getSection() != null && !chunk.getSection().isBlank()) {
|
source = source + " · " + chunk.getSection();
|
}
|
String header = "[" + index + "] 来源:" + source + "\n";
|
String body = chunk.getText();
|
|
// 总量控制:超出上限就停止追加,防止撑爆模型上下文窗口
|
if (sb.length() + header.length() + body.length() > maxChars) {
|
int remaining = maxChars - sb.length() - header.length();
|
if (remaining > 200) {
|
sb.append(header).append(body, 0, remaining).append("...\n\n");
|
index++;
|
}
|
log.debug("RAG 上下文达到字符上限 {},已截断", maxChars);
|
break;
|
}
|
|
sb.append(header).append(body).append("\n\n");
|
index++;
|
}
|
|
return sb.isEmpty() ? null : sb.toString().trim();
|
}
|
|
@Override
|
public boolean deleteCollection(String collectionName) {
|
if (!ragProperties.isEnabled() || collectionName == null || collectionName.isBlank()) {
|
return false;
|
}
|
String url = baseUrl() + "/api/rag/collection/" + collectionName;
|
try {
|
restTemplate.delete(url);
|
log.info("已删除向量集合 {}", collectionName);
|
return true;
|
} catch (Exception e) {
|
log.warn("删除向量集合失败 collection={} err={}", collectionName, rootMessage(e));
|
return false;
|
}
|
}
|
|
@Override
|
public OcrResultDTO recognizeImage(Path imagePath) {
|
if (!ragProperties.isEnabled()) {
|
throw BusinessException.badRequest("RAG 功能未启用,无法使用图片识别");
|
}
|
if (imagePath == null || !Files.exists(imagePath)) {
|
throw BusinessException.badRequest("图片文件不存在: " + imagePath);
|
}
|
|
MultiValueMap<String, Object> body = new LinkedMultiValueMap<>();
|
body.add("file", new FileSystemResource(imagePath));
|
|
HttpHeaders headers = new HttpHeaders();
|
headers.setContentType(MediaType.MULTIPART_FORM_DATA);
|
|
String url = baseUrl() + "/api/ocr/recognize";
|
try {
|
// OCR 超时单独设置:RagRestTemplate 的读超时是按「向量检索」调的,
|
// 首次 OCR 要加载 onnx 模型,可能明显更久
|
RestTemplate client = ocrRestTemplate();
|
ResponseEntity<String> response = client.postForEntity(
|
url, new HttpEntity<>(body, headers), String.class);
|
|
JsonNode root = objectMapper.readTree(response.getBody());
|
JsonNode data = unwrapData(root);
|
|
OcrResultDTO dto = new OcrResultDTO();
|
dto.setFilename(text(data, "filename", imagePath.getFileName().toString()));
|
dto.setText(text(data, "text", ""));
|
dto.setEngine(text(data, "engine", null));
|
dto.setImageWidth(data.path("image_width").isMissingNode() ? null : data.path("image_width").asInt());
|
dto.setImageHeight(data.path("image_height").isMissingNode() ? null : data.path("image_height").asInt());
|
dto.setCharCount(data.path("char_count").asInt(dto.getText() == null ? 0 : dto.getText().length()));
|
dto.setElapsedMs(data.path("elapsed_ms").isMissingNode() ? null : data.path("elapsed_ms").asLong());
|
|
JsonNode blocks = data.path("blocks");
|
if (blocks.isArray()) {
|
for (JsonNode node : blocks) {
|
OcrResultDTO.Block b = new OcrResultDTO.Block();
|
b.setText(text(node, "text", ""));
|
if (node.hasNonNull("confidence")) {
|
b.setConfidence(node.get("confidence").asDouble());
|
}
|
JsonNode box = node.path("box");
|
if (box.isArray()) {
|
List<Double> coords = new ArrayList<>(box.size());
|
box.forEach(v -> coords.add(v.asDouble()));
|
b.setBox(coords);
|
}
|
dto.getBlocks().add(b);
|
}
|
}
|
|
log.info("OCR 完成 file={} engine={} 字符={} 耗时={}ms",
|
dto.getFilename(), dto.getEngine(), dto.getCharCount(), dto.getElapsedMs());
|
return dto;
|
|
} catch (Exception e) {
|
log.error("调用 OCR 接口失败 url={}", url, e);
|
throw BusinessException.serverError(
|
"图片识别失败(" + url + ")。请确认 RAG 服务已启动且 OCR 引擎可用:"
|
+ "macOS 执行 bash rag-service/tools/build_vision_ocr.sh,"
|
+ "或 pip install rapidocr-onnxruntime。原因: " + rootMessage(e));
|
}
|
}
|
|
/** OCR 专用客户端:读超时按 OCR 配置,避免被检索的短超时误杀 */
|
private RestTemplate ocrRestTemplate() {
|
SimpleClientHttpRequestFactory factory = new SimpleClientHttpRequestFactory();
|
factory.setConnectTimeout(5000);
|
factory.setReadTimeout(ragProperties.getOcrTimeoutMillis());
|
return new RestTemplate(factory);
|
}
|
|
@Override
|
public boolean isHealthy() {
|
if (!ragProperties.isEnabled()) {
|
return false;
|
}
|
try {
|
ResponseEntity<String> response =
|
restTemplate.getForEntity(baseUrl() + "/health", String.class);
|
return response.getStatusCode().is2xxSuccessful();
|
} catch (Exception e) {
|
return false;
|
}
|
}
|
|
// ==================== 内部辅助 ====================
|
|
/** 去掉末尾斜杠,避免拼出 //api/rag/... */
|
private String baseUrl() {
|
String url = ragProperties.getBaseUrl();
|
return url != null && url.endsWith("/") ? url.substring(0, url.length() - 1) : url;
|
}
|
|
/**
|
* 兼容两种响应格式:{@code {code,msg,data}} 与直接的对象。
|
* Python 侧返回统一结构时取 data,否则把根节点当作数据。
|
*/
|
private JsonNode unwrapData(JsonNode root) {
|
if (root != null && root.has("data") && !root.get("data").isNull()) {
|
return root.get("data");
|
}
|
return root == null ? objectMapper.createObjectNode() : root;
|
}
|
|
private String text(JsonNode node, String field, String defaultValue) {
|
JsonNode value = node.get(field);
|
return value == null || value.isNull() ? defaultValue : value.asText();
|
}
|
|
/** 取异常链最底层的信息,RestTemplate 的异常包装很深,直接 getMessage 可读性差 */
|
private String rootMessage(Throwable t) {
|
Throwable cursor = t;
|
while (cursor.getCause() != null && cursor.getCause() != cursor) {
|
cursor = cursor.getCause();
|
}
|
String msg = cursor.getMessage();
|
return msg == null ? cursor.getClass().getSimpleName() : abbreviate(msg, 300);
|
}
|
|
private String abbreviate(String s, int max) {
|
if (s == null) {
|
return "";
|
}
|
return s.length() <= max ? s : s.substring(0, max) + "...";
|
}
|
}
|