AgenticRAG自适应检索深度实战 - 十、Spring AI 完整实现
本文是 AgenticRAG自适应检索深度实战 的分章阅读版。
十、Spring AI 完整实现
10.1 整合 Agentic RAG
@Service
public class AgenticRAGFacade {
@Autowired private QueryClassifier classifier;
@Autowired private SelfRAGService selfRAG;
@Autowired private CRAGService crag;
@Autowired private RAGFusionService fusion;
@Autowired private MultiHopRAGService multiHop;
@Autowired private AgenticGraphRAGService graphRAG;
public RAGResponse ask(RAGRequest req) {
String query = req.getQuery();
// 1. 分类
QueryProfile profile = analyze(query);
// 2. 路由
String answer = switch (profile.getStrategy()) {
case NO_RAG -> llm.chat(query);
case SINGLE_STEP -> crag.ask(query);
case MULTI_QUERY -> fusion.ask(query);
case MULTI_HOP -> multiHop.ask(query);
case GRAPH -> graphRAG.ask(query);
case SELF_RAG -> selfRAG.ask(query, 3).getAnswer();
};
return RAGResponse.builder()
.answer(answer)
.strategy(profile.getStrategy())
.citations(extractCitations(answer))
.build();
}
private QueryProfile analyze(String query) {
QueryComplexity complexity = classifier.classifyByRule(query);
Strategy strategy = switch (complexity) {
case SIMPLE -> Strategy.NO_RAG;
case MEDIUM -> containsRelations(query) ? Strategy.GRAPH : Strategy.SINGLE_STEP;
case COMPLEX -> isMultiHop(query) ? Strategy.MULTI_HOP : Strategy.MULTI_QUERY;
};
return new QueryProfile(complexity, strategy);
}
}
10.2 Spring AI 原生 RAG
@Configuration
public class SpringAIRAGConfig {
@Bean
public ChatClient chatClient(ChatClient.Builder builder) {
return builder
.defaultSystem("你是知识库助手,基于提供的文档回答,不要编造。")
.build();
}
@Bean
public VectorStore vectorStore(JedisConnectionFactory cf) {
return RedisVectorStore.builder(cf)
.indexName("wiki-index")
.prefix("wiki:")
.vector(1024)
.build();
}
}
@Service
public class SpringAINaiveRAG {
@Autowired private ChatClient chatClient;
@Autowired private VectorStore vectorStore;
public String ask(String query) {
// Spring AI 原生 API
List<Document> docs = vectorStore.similaritySearch(
SearchRequest.query(query).withTopK(5)
);
return chatClient.prompt()
.user(query)
.documents(docs) // 自动拼接到 prompt
.call()
.content();
}
}
10.3 自定义 Agentic Advisor
@Component
public class AgenticRAGAdvisor implements BaseAdvisor {
@Override
public AdvisedResponse adviseRequest(ChatClientRequest request,
AdvisorChain chain) {
String query = request.userText();
// 1. 检索
List<Document> docs = adaptiveRetrieve(query);
// 2. 评估
docs = filterRelevant(query, docs);
// 3. 注入到 system
String context = docs.stream()
.map(Document::getText)
.collect(Collectors.joining("\n---\n"));
request.systemParams().put("context", context);
return chain.next(request);
}
private List<Document> adaptiveRetrieve(String query) {
// 根据 query 复杂度选择策略
// ...
}
}
// 使用
String answer = chatClient.prompt()
.user(query)
.advisors(new AgenticRAGAdvisor())
.call()
.content();
相关页面