背景
公司想要做本地私有化的知识库，一年前OpenAI出来的时候用 Faiss加上 调用第三方的API（向量啥的） 做了一版了，后面裁员，这玩意就废弃了，又要重新搞。
实现步骤
1、【上传文档】 → 【解析为文本】 → 【调用向量模型向量化】 → 【将向量数据存储到ES】
2、【用户提问】 → 【把问题调用向量模型向量化】 → 【去ES里做向量检索】 → 【将查询的相似数据做语义相关度排序】 → 【把排序后的数据加上预制提示词调用 LLM 总结并返回】
前提
首先在huggingface上下载模型
模型下载地址：https://huggingface.co/
1、使用Ollama本地部署deepseek-r1:32b
2、将 m3e-large 构建为 Api 服务
3、将 bce-reranker-base_v1 构建为 Api 服务
演示效果
源码获取方式，关注公众号，回复【知识库源码】代码包含 前端，springboot后端，python端
步骤一：上传并解析文档
简单实现了一下pdf，word解析，用了OpenCV，OCR解析文档中的图片，图片解析效果一般
public void uploadFile(MultipartFile file) {
try {
String projectPath = System.getProperty("user.dir");
Path tempDirPath = Paths.get(projectPath, TEMP_DIR);
if (!Files.exists(tempDirPath)) {
Files.createDirectories(tempDirPath);
}
// 获取文件名和扩展名
String originalFilename = file.getOriginalFilename();
String fileExtension = getFileExtension(originalFilename);
// 生成临时文件路径
String timestamp = LocalDateTime.now().format(DateTimeFormatter.ofPattern("yyyyMMddHHmmss"));
String tempFileName = timestamp + "_" + originalFilename;
Path tempFilePath = Paths.get(projectPath, TEMP_DIR, tempFileName);
// 保存上传的文件
file.transferTo(tempFilePath.toFile());
// 解析文件内容
Map<String, Object> parsedContent = parseFile(tempFilePath.toString(), fileExtension);
// 保存到 Elasticsearch
String text = parsedContent.get("text").toString();
KnowsIndex knowsIndex = new KnowsIndex();
knowsIndex.setId(String.valueOf(IdUtil.getSnowflakeNextId()));
knowsIndex.setContent(text);
knowsIndex.setContent_vec(EmbedClient.getEmbedding(embeddingUri, embeddingApiKey, text));
esDocumentService.indexSellList(Arrays.asList(knowsIndex));
// 清理临时文件
Files.deleteIfExists(tempFilePath);
} catch (Exception e) {
e.printStackTrace();
}
}
步骤二：调用 m3e-large 向量数据
public static double[] getEmbedding(String uri, String apiKey, String inputText) throws IOException {
OkHttpClient client = new OkHttpClient();
// 创建请求体
JSONObject requestBody = new JSONObject();
requestBody.put("input", Collections.singletonList(inputText));
// 创建请求
MediaType mediaType = MediaType.parse("application/json; charset=utf-8");
RequestBody body = RequestBody.Companion.create(requestBody.toJSONString(), mediaType);
Request request = new Request.Builder()
.url(uri)
.addHeader("Authorization", "Bearer " + apiKey)
.addHeader("Content-Type", "application/json")
.post(body)
.build();
// 发送请求
Response response = client.newCall(request).execute();
if (!response.isSuccessful()) {
throw new IOException("Unexpected code " + response);
}
// 解析JSON响应
String responseBody = response.body().string();
EmbeddingResponse embeddingResponse = JSON.parseObject(responseBody, EmbeddingResponse.class);
// 返回嵌入向量
return embeddingResponse.getData().get(0).getEmbedding();
}
步骤三：保存到ES
public void indexSellList(List<KnowsIndex> knowsIndexList) throws IOException {
for (KnowsIndex knowsIndex : knowsIndexList) {
knowsIndex.setContent_vec(EmbedClient.getEmbedding(embeddingUri, embeddingApiKey, knowsIndex.getContent()));
IndexResponse response = client.index(i -> i
.index(INDEX_NAME)
.id(knowsIndex.getId())
.document(knowsIndex)
);
log.info("Sell indexed: {}", response.id());
}
}
步骤四：用户搜索
用户搜索问题向量化后，去ES检索
public List<SearchResult> searchVector(double[] queryVector) throws IOException {
// 创建向量相似度查询
ScriptScoreQuery scriptScoreQuery = ScriptScoreQuery.of(q -> q
.query(QueryBuilders.matchAll().build()._toQuery())
.script(Script.of(s -> s.inline(i -> i
.source("double score = cosineSimilarity(params.query_vector, 'content_vec'); " +
"score = Math.min(1.0, Math.max(0.0, score)); " + // 确保评分在[0, 1]之间
"if (score < params.threshold) { return 0; } else { return score; }")
.params(Map.of(
"query_vector", JsonData.of(queryVector),
"threshold", JsonData.of(SIMILARITY_THRESHOLD) // 将阈值作为参数传递给脚本
))))));
// 创建bool查询，向量相似度查询作为should子句
Query boolQuery = QueryBuilders.bool(b -> b
.should(scriptScoreQuery._toQuery())
);
Query functionScoreQuery = QueryBuilders.functionScore(fs -> fs
.query(boolQuery)
.scoreMode(FunctionScoreMode.Max)
.boostMode(FunctionBoostMode.Replace)
.minScore((double) SIMILARITY_THRESHOLD)
);
// 执行合并后的查询
SearchResponse<KnowsIndex> combinedSearchResponse = client.search(s -> s
.index(INDEX_NAME)
.query(functionScoreQuery),
KnowsIndex.class);
// 处理查询的结果
return combinedSearchResponse.hits().hits().stream()
.map(hit -> {
double finalScore = Objects.nonNull(hit.score()) ? hit.score() : 0.0;
return finalScore >= SIMILARITY_THRESHOLD ? new SearchResult(hit.source(), finalScore) : null;
})
.filter(Objects::nonNull)
.sorted(Comparator.comparingDouble(SearchResult::getScore).reversed())
.collect(Collectors.toList());
}
步骤五：检索后的文档内容 调用deepseek总结
可以换成阿里千问或者是其他大语言模型
public class OllamaClient {
private static Map<String, Object> PARAMS = new HashMap<>();
private static Map<String, Object> OPTIONS = new HashMap<>();
static {
OPTIONS.put("temperature", 0.3); // # 控制随机性（0-1，值越大越随机）
OPTIONS.put("top_p", 0.5); // # 采样策略（0-1，值越小越集中）
OPTIONS.put("max_tokens", 1024); // # 生成的最大 token 数
PARAMS.put("model", "deepseek-r1:32b");
PARAMS.put("stream", true);
PARAMS.put("options", OPTIONS);
}
public static String PROMPT = "你是一个知识库，必须严格按照知识库检索的内容做最精简的回答，只回答关键信息，坚决杜绝胡编乱造，注意字数。" +
"当所有知识库内容都与产品问题无关时，或者知识库检索到任何相关信息时，你的回答必须是“没有找到”这句话。" +
" 以下是知识库：\n" +
" { %content% }\n" +
" 以上是知识库。 \n 以下是提问：";
public static void sendMsg(HttpServletResponse response, String uri, String query, String content) {
try {
// 设置SSE必要的响应头
response.setContentType("text/event-stream");
response.setCharacterEncoding("UTF-8");
response.setHeader("Cache-Control", "no-cache");
response.setHeader("Connection", "keep-alive");
URL url = new URL(uri);
HttpURLConnection conn = (HttpURLConnection) url.openConnection();
conn.setRequestMethod("POST");
conn.setRequestProperty("Accept", "text/event-stream");
conn.setRequestProperty("Content-Type", "application/json");
conn.setDoOutput(true);
PARAMS.put("prompt", PROMPT.replace("%content%", content) + query);
String json = JSON.toJSONString(PARAMS);
try (OutputStream os = conn.getOutputStream()) {
os.write(json.getBytes(StandardCharsets.UTF_8));
}
int responseCode = conn.getResponseCode();
if (responseCode >= HttpURLConnection.HTTP_OK && responseCode < HttpURLConnection.HTTP_USE_PROXY) {
try (BufferedReader br = new BufferedReader(new InputStreamReader(conn.getInputStream(), StandardCharsets.UTF_8));
PrintWriter writer = response.getWriter()) {
String line;
while ((line = br.readLine()) != null) {
if (!line.trim().isEmpty()) {
// 构造SSE消息格式
writer.write("data: " + line + "\n\n");
writer.flush();
}
}
}
} else {
throw new RuntimeException("Failed : HTTP error code : " + responseCode);
}
} catch (Exception e) {
try {
response.setStatus(HttpServletResponse.SC_INTERNAL_SERVER_ERROR);
PrintWriter writer = response.getWriter();
writer.write("data: {\"error\": \"" + e.getMessage() + "\"}\n\n");
writer.flush();
} catch (IOException ioe) {
e.printStackTrace();
}
}
}
}
前端代码
流式回复，Cursor生成，自己简单调了一下
END
源码获取方式，关注公众号：程序员嗨皮，回复【知识库源码】
代码包含前端，springboot后端，python端，不包含模型，模型太大了，自己去下载就行了。有不明白的可以加我微信联系我
