説明
- 用户輸入多個“信息”
- 大語言模型將“信息”進行處理，轉成數組；（一維張量，向量）
- 通過餘弦相似度等相關算法，計算兩個向量是否相似
Ollama接口步驟
- 安裝 Ollama: https://ollama.ai/
- 下載模型: ollama pull nomic-embed-text
- Ollama 默認運行在 http://localhost:11434
推薦的嵌入模型:
- nomic-embed-text: 768維，效果好，速度快
- mxbai-embed-large: 1024維，效果更好
- bge-m3: 多語言支持
springboot中調用本地模型
@Test
@Disabled("需要本地運行 Ollama 服務")
public void testOllamaEmbedding() {
// Ollama API 地址
String apiUrl = "http://localhost:11434/api/embeddings";
String apiKey = ""; // Ollama 本地不需要 key
String model = "nomic-embed-text"; // 或 mxbai-embed-large
EmbeddingClient client = new EmbeddingClientImpl(apiUrl, apiKey);
// 水果庫
List<Fruit> fruits = Arrays.asList(new Fruit("紅富士蘋果", "紅色 甜 脆 蘋果 新鮮"), new Fruit("青蘋果", "綠色 酸 脆 蘋果 清爽"),
new Fruit("金帥蘋果", "黃色 甜 軟 蘋果"), new Fruit("香蕉", "黃色 甜 軟 香蕉 熱帶水果"), new Fruit("草莓", "紅色 甜 小 草莓 多汁 漿果"),
new Fruit("西瓜", "綠色外皮 紅色果肉 甜 大 西瓜 多汁 夏天"), new Fruit("葡萄", "紫色 甜 小 葡萄 多汁 成串"));
// 為每個水果生成嵌入向量
for (Fruit fruit : fruits) {
fruit.embedding = client.getEmbeddingVector(model, fruit.description);
}
// 用户搜索
String query = "紅色的甜水果";
double[] queryVector = client.getEmbeddingVector(model, query);
System.out.println("搜索: \"" + query + "\"");
System.out.println("向量維度: " + queryVector.length);
System.out.println();
// 按相似度排序
fruits.sort(Comparator.comparingDouble(f -> -cosineSimilarity(queryVector, f.embedding)));
// 輸出結果
System.out.println("搜索結果（按相似度排序）:");
for (Fruit f : fruits) {
double sim = cosineSimilarity(queryVector, f.embedding);
System.out.printf(" %s (%.4f): %s%n", f.name, sim, f.description);
}
}
/**
* 計算兩個向量的餘弦相似度
*/
public static double cosineSimilarity(double[] vectorA, double[] vectorB) {
if (vectorA.length != vectorB.length) {
throw new IllegalArgumentException("向量維度必須相同");
}
double dotProduct = 0;
double normA = 0;
double normB = 0;
for (int i = 0; i < vectorA.length; i++) {
dotProduct += vectorA[i] * vectorB[i];
normA += vectorA[i] * vectorA[i];
normB += vectorB[i] * vectorB[i];
}
if (normA == 0 || normB == 0) {
return 0;
}
return dotProduct / (Math.sqrt(normA) * Math.sqrt(normB));
}
核心方法
@Slf4j
public class EmbeddingClientImpl implements EmbeddingClient {
private final RestTemplate restTemplate;
private final String address;
private final String key;
public EmbeddingClientImpl(String address, String key) {
PoolingHttpClientConnectionManager connectionManager = new PoolingHttpClientConnectionManager();
connectionManager.setMaxTotal(100);
connectionManager.setDefaultMaxPerRoute(20);
// 設置請求配置
RequestConfig requestConfig = RequestConfig.custom()
.setConnectionRequestTimeout(Timeout.ofSeconds(30))
.setResponseTimeout(Timeout.ofSeconds(300)) // 5分鐘響應超時
.build();
// 使用 HttpClientBuilder 來構建 HttpClient
HttpClient httpClient = HttpClientBuilder.create()
.setConnectionManager(connectionManager)
.setDefaultRequestConfig(requestConfig)
.build();
// 創建 HttpComponentsClientHttpRequestFactory
HttpComponentsClientHttpRequestFactory requestFactory = new HttpComponentsClientHttpRequestFactory(httpClient);
requestFactory.setConnectTimeout(30000); // 30秒連接超時
requestFactory.setConnectionRequestTimeout(30000);
// 創建 RestTemplate，只使用 StringHttpMessageConverter 避免 Jackson 依賴問題
this.restTemplate = new RestTemplate(requestFactory);
// 清除默認的消息轉換器，只保留字符串轉換器
this.restTemplate.setMessageConverters(
Collections.singletonList(new StringHttpMessageConverter(StandardCharsets.UTF_8)));
this.address = address;
this.key = key;
}
@Override
public String embedding(String model, String input) {
long start = System.currentTimeMillis();
String url = address;
HttpHeaders headers = new HttpHeaders();
headers.setContentType(MediaType.APPLICATION_JSON);
headers.setAcceptCharset(Collections.singletonList(StandardCharsets.UTF_8));
if (key != null && !key.isEmpty()) {
headers.add("Authorization", "Bearer " + key);
}
// 將 request 轉化為 body 字符串
JSONObject jsonObject = new JSONObject();
jsonObject.put("input", input);
jsonObject.put("model", model);
String body = jsonObject.toString();
log.debug("Embedding Request Body: {}", body);
// 請求
HttpEntity<String> req = new HttpEntity<>(body, headers);
ResponseEntity<String> result = restTemplate.postForEntity(url, req, String.class);
if (!result.getStatusCode().equals(HttpStatus.OK)) {
throw new RuntimeException("embeddings error, request: " + body + ", response: " + result.getBody());
}
log.info("embedding cost {} ms", System.currentTimeMillis() - start);
return result.getBody();
}
/**
* 獲取文本嵌入向量
* <p>
* 解析 OpenAI 格式的響應，提取 embedding 向量
*
* 響應格式示例： <pre>
* {
* "object": "list",
* "data": [{
* "object": "embedding",
* "index": 0,
* "embedding": [0.0023064255, -0.009327292, ...]
* }],
* "model": "text-embedding-ada-002",
* "usage": {"prompt_tokens": 8, "total_tokens": 8}
* }
* </pre>
* @param model 模型名稱
* @param input 輸入文本
* @return 嵌入向量
*/
@Override
public double[] getEmbeddingVector(String model, String input) {
String response = embedding(model, input);
return parseEmbeddingVector(response);
}
/**
* 解析嵌入向量響應
* @param response JSON響應字符串
* @return 向量數組
*/
private double[] parseEmbeddingVector(String response) {
try {
JSONObject jsonResponse = JSONObject.parseObject(response);
// OpenAI 格式
if (jsonResponse.containsKey("data")) {
JSONArray dataArray = jsonResponse.getJSONArray("data");
if (dataArray != null && !dataArray.isEmpty()) {
JSONObject firstData = dataArray.getJSONObject(0);
JSONArray embeddingArray = firstData.getJSONArray("embedding");
return jsonArrayToDoubleArray(embeddingArray);
}
}
// Ollama 格式 (直接返回 embedding 數組)
if (jsonResponse.containsKey("embedding")) {
JSONArray embeddingArray = jsonResponse.getJSONArray("embedding");
return jsonArrayToDoubleArray(embeddingArray);
}
// 阿里通義格式
if (jsonResponse.containsKey("output")) {
JSONObject output = jsonResponse.getJSONObject("output");
if (output.containsKey("embeddings")) {
JSONArray embeddings = output.getJSONArray("embeddings");
if (!embeddings.isEmpty()) {
JSONObject firstEmbedding = embeddings.getJSONObject(0);
JSONArray embeddingArray = firstEmbedding.getJSONArray("embedding");
return jsonArrayToDoubleArray(embeddingArray);
}
}
}
throw new RuntimeException("無法解析嵌入向量響應: " + response);
}
catch (Exception e) {
log.error("解析嵌入向量失敗: {}", response, e);
throw new RuntimeException("解析嵌入向量失敗", e);
}
}
/**
* 將 JSONArray 轉換為 double 數組
*/
private double[] jsonArrayToDoubleArray(JSONArray jsonArray) {
double[] result = new double[jsonArray.size()];
for (int i = 0; i < jsonArray.size(); i++) {
result[i] = jsonArray.getDoubleValue(i);
}
return result;
}
}
