virtual-patient/virtual-patient-common/src/main/java/com/supervision/util/SimilarityUtil.java

package com.supervision.util;

import cn.hutool.core.collection.CollUtil;
import com.supervision.domain.QaSimilarityQuestionAnswer;
import lombok.extern.slf4j.Slf4j;
import org.springframework.ai.document.Document;
import org.springframework.ai.vectorstore.RedisVectorStore;
import org.springframework.ai.vectorstore.SearchRequest;
import org.springframework.ai.vectorstore.filter.FilterExpressionBuilder;

import java.util.*;

@Slf4j
public class SimilarityUtil {

    private static final RedisVectorStore redisVectorStore = SpringBeanUtil.getBean(RedisVectorStore.class);

    /**
     * 相似度比较,只找最高的
     *
     * @param question 问题
     * @return 最高的TOP
     */
    public static Optional<QaSimilarityQuestionAnswer> talkRedisVectorWithScoreByFirst(String question) {
        List<QaSimilarityQuestionAnswer> qaSimilarityQuestionAnswers = talkRedisVectorWithScore(question);
        return qaSimilarityQuestionAnswers.stream().findFirst();
    }

    /**
     * 相似度比较,找到病历里面配置答案的问题
     *
     * @param question              问题
     * @param libraryQuestionIdList 配置答案的问题ID
     * @return 匹配到的结果列表
     */
    public static List<QaSimilarityQuestionAnswer> talkRedisVectorWithScore(String question, Collection<String> libraryQuestionIdList) {
        log.info("开始调用talkQaSimilarity,问题:{}", question);
        try {

            SearchRequest searchRequest = SearchRequest
                    .query(question)
                    .withTopK(10);
            // 添加条件,只匹配对应标准问的问题
            if (CollUtil.isNotEmpty(libraryQuestionIdList)) {
                FilterExpressionBuilder b = new FilterExpressionBuilder();
                searchRequest.withFilterExpression(b.in("libraryQuestionId", new ArrayList<>(libraryQuestionIdList)).build());
            }
            // 走Redis向量库进行比较,找出最高的top10
            List<Document> documents = redisVectorStore.similaritySearch(searchRequest);
            return documents.stream().map(document -> {
                QaSimilarityQuestionAnswer qaSimilarityQuestionAnswer = new QaSimilarityQuestionAnswer();
                qaSimilarityQuestionAnswer.setMatchQuestion(document.getContent());
                qaSimilarityQuestionAnswer.setQuestionId(String.valueOf(document.getMetadata().get("questionId")));
                qaSimilarityQuestionAnswer.setDictId(String.valueOf(document.getMetadata().get("dictId")));
                qaSimilarityQuestionAnswer.setLibraryQuestionId(String.valueOf(document.getMetadata().get("libraryQuestionId")));
                // 计算相似度
                double score = computeScore(document);
                qaSimilarityQuestionAnswer.setMatchScore(score);
                return qaSimilarityQuestionAnswer;
                // 排序,降序,取最高的
            }).sorted(Comparator.comparing(QaSimilarityQuestionAnswer::getMatchScore).reversed()).toList();
        } catch (Exception e) {
            log.error("调用talkQaSimilarity error ", e);
            return new ArrayList<>();
        }
    }

    /**
     * 计算相似度得分
     *
     * @param document 文档
     * @return 得分
     */
    public static Double computeScore(Document document) {
        // 1- 可以使数据进行排序,相似度越高,数值越大(redis相似度给的数据是越小相似度越高)
        // -0.25目的是使数据趋近于中间,相似度不要太大(太大也不好调整),以使我们数据和张总之前提供的方法相似度差异稍小一点,但是不能小于0,如果小于0,取一个较大的值
        return Math.max(0, 1 - Double.parseDouble(String.valueOf(document.getMetadata().get("vector_score"))) - 0.25);
    }

    /**
     * 相似度比较,找出所有的
     *
     * @param question 问题
     * @return 相似度最高的top10
     */
    public static List<QaSimilarityQuestionAnswer> talkRedisVectorWithScore(String question) {
        return talkRedisVectorWithScore(question, null);
    }


}