用 Spring AI + pgvector 构建 RAG 知识库 — 新加坡数学 AI 辅导 Phase 2 实战

{
  "content": "The ratio of boys to girls in a class is 3:5. If there are 24 boys, how many students are there altogether?",
  "metadata": {
    "grade": 5,
    "topic": "ratio.basic",
    "difficulty": "medium",
    "source": "PSLE Practice"
  }
}

spring:
  ai:
    ollama:
      embedding:
        model: nomic-embed-text    # 768维向量
    vectorstore:
      pgvector:
        initialize-schema: true    # 自动建 vector_store 表
        dimensions: 768            # 匹配 embedding 模型
        index-type: hnsw           # HNSW 索引

@Service
public class QuestionImportService {

    private final VectorStore vectorStore;
    private final JdbcTemplate jdbcTemplate;
    private final ObjectMapper objectMapper = new ObjectMapper();

    @EventListener(ApplicationReadyEvent.class)
    public void importQuestionsOnStartup() {
        Long count = jdbcTemplate.queryForObject(
                "SELECT count(*) FROM vector_store", Long.class);
        if (count != null && count > 0) {
            log.info("Vector store already has {} docs, skip", count);
            return;
        }

        log.info("Vector store is empty, importing...");
        importQuestions();
    }

    public void importQuestions() {
        var resource = new ClassPathResource("data/sg-math-questions.json");
        List<Map<String, Object>> questions = objectMapper.readValue(
                resource.getInputStream(), new TypeReference<>() {});

        List<Document> documents = questions.stream()
                .map(q -> new Document(
                    (String) q.get("content"),
                    (Map<String, Object>) q.get("metadata")))
                .toList();

        vectorStore.add(documents);  // 自动调用 Embedding API + 写入 PG
    }
}

SELECT count(*) FROM vector_store;
-- 40

SELECT content, metadata FROM vector_store LIMIT 1;
-- content: "Ali had 120 stickers..."
-- metadata: {"grade": 4, "topic": "fractions.of_remainder", ...}

private final ObjectMapper objectMapper = new ObjectMapper();

@Service
public class RagRetrievalService {

    private static final int TOP_K = 5;
    private static final double SIMILARITY_THRESHOLD = 0.5;

    private final VectorStore vectorStore;

    public List<Document> retrieveSimilarQuestions(
            String question, int grade) {
        var filterBuilder = new FilterExpressionBuilder();
        var filter = filterBuilder.lte("grade", grade).build();

        SearchRequest searchRequest = SearchRequest.builder()
                .query(question)
                .topK(TOP_K)
                .similarityThreshold(SIMILARITY_THRESHOLD)
                .filterExpression(filter)
                .build();

        return vectorStore.similaritySearch(searchRequest);
    }

    public String formatAsContext(List<Document> documents) {
        var sb = new StringBuilder(
            "=== Similar Questions from PSLE Question Bank ===\n\n");
        for (int i = 0; i < documents.size(); i++) {
            Document doc = documents.get(i);
            sb.append("Question %d: %s\n".formatted(
                i + 1, doc.getText()));
            sb.append("  Topic: %s\n".formatted(
                doc.getMetadata().get("topic")));
            sb.append("  Difficulty: %s\n".formatted(
                doc.getMetadata().get("difficulty")));
            sb.append("\n");
        }
        return sb.toString();
    }
}

private String runPlannerAgent(
        SolveRequest request, String ragContext) {
    String userMessage = """
        Grade: P%d
        Question: %s

        %s
        """.formatted(
            request.grade(),
            request.question(),
            ragContext);

    return chatClient.prompt()
            .system(PLANNER_SYSTEM_PROMPT)
            .user(userMessage)
            .call()
            .content();
}

@Service
public class SolveService {

    private final MathSolverOrchestrator orchestrator;

    @Cacheable(
        value = "solveResults",
        key = "#request.question().trim().toLowerCase()"
            + " + ':' + #request.grade()")
    public SolveResult solve(SolveRequest request) {
        log.info("Cache miss - running full agent pipeline");
        return orchestrator.solve(request);
    }
}

@Configuration
@EnableCaching
public class CacheConfig {

    @Bean
    public CacheManager cacheManager(
            RedisConnectionFactory connectionFactory) {
        var jsonSerializer =
            new Jackson2JsonRedisSerializer<>(Object.class);

        RedisCacheConfiguration config =
            RedisCacheConfiguration.defaultCacheConfig()
                .entryTtl(Duration.ofHours(24))
                .serializeValuesWith(
                    RedisSerializationContext.SerializationPair
                        .fromSerializer(jsonSerializer))
                .disableCachingNullValues();

        return RedisCacheManager.builder(connectionFactory)
                .cacheDefaults(config)
                .build();
    }
}