|
| 1 | +package com.ai.lawyer.global.batch; |
| 2 | + |
| 3 | +/*@Slf4j |
| 4 | +@Configuration |
| 5 | +@RequiredArgsConstructor |
| 6 | +public class DataVectorizationJobConfig { |
| 7 | +
|
| 8 | + private final JobRepository jobRepository; |
| 9 | + private final PlatformTransactionManager transactionManager; |
| 10 | + private final EntityManagerFactory entityManagerFactory; |
| 11 | + private final VectorStore vectorStore; |
| 12 | +
|
| 13 | + private final JangRepository jangRepository; |
| 14 | + private final JoRepository joRepository; |
| 15 | + private final HangRepository hangRepository; |
| 16 | + private final HoRepository hoRepository; |
| 17 | +
|
| 18 | + private final TokenTextSplitter tokenSplitter = TokenTextSplitter.builder() |
| 19 | + .withChunkSize(800) |
| 20 | + .withMinChunkSizeChars(0) |
| 21 | + .withMinChunkLengthToEmbed(5) |
| 22 | + .withMaxNumChunks(10000) |
| 23 | + .withKeepSeparator(true) |
| 24 | + .build(); |
| 25 | +
|
| 26 | + private static final int CHUNK_SIZE = 10; // 배치 처리 시 한 번에 읽어올 데이터 수 |
| 27 | +
|
| 28 | + @Value("${batch.page.size.precedent}") |
| 29 | + private int precedentPageSize; // 하루에 처리할 판례 수 |
| 30 | +
|
| 31 | + @Value("${batch.page.size.law}") |
| 32 | + private int lawPageSize; // 하루에 처리할 법령 수 |
| 33 | +
|
| 34 | + @Bean |
| 35 | + public TaskExecutor taskExecutor() { |
| 36 | + ThreadPoolTaskExecutor executor = new ThreadPoolTaskExecutor(); |
| 37 | + executor.setCorePoolSize(10); |
| 38 | + executor.setMaxPoolSize(20); |
| 39 | + executor.setQueueCapacity(100); |
| 40 | + executor.setThreadNamePrefix("async-thread-"); |
| 41 | + executor.initialize(); |
| 42 | + return executor; |
| 43 | + } |
| 44 | +
|
| 45 | + // -------------- 전체 데이터 벡터화 정의 -------------- |
| 46 | + @Bean |
| 47 | + public Job dataVectorizationJob() { |
| 48 | + return new JobBuilder("dataVectorizationJob", jobRepository) |
| 49 | + .start(precedentVectorizationStep()) // 판례 벡터화 Step 실행 |
| 50 | + .next(lawVectorizationStep()) // 법령 벡터화 Step 실행 |
| 51 | + .build(); |
| 52 | + } |
| 53 | +
|
| 54 | + // -------------- 판례 벡터화 --------------- |
| 55 | + @Bean |
| 56 | + public Step precedentVectorizationStep() { |
| 57 | + log.info(">>>>>> 판례 벡터화 시작"); |
| 58 | + return new StepBuilder("precedentVectorizationStep", jobRepository) |
| 59 | + .<Precedent, List<Document>>chunk(CHUNK_SIZE, transactionManager) |
| 60 | + .reader(precedentItemReader()) |
| 61 | + .processor(precedentItemProcessor()) |
| 62 | + .writer(documentItemWriter()) |
| 63 | + .taskExecutor(taskExecutor()) |
| 64 | + .build(); |
| 65 | + } |
| 66 | +
|
| 67 | + @Bean |
| 68 | + public JpaPagingItemReader<Precedent> precedentItemReader() { |
| 69 | + return new JpaPagingItemReaderBuilder<Precedent>() |
| 70 | + .name("precedentItemReader") |
| 71 | + .entityManagerFactory(entityManagerFactory) |
| 72 | + .pageSize(CHUNK_SIZE) |
| 73 | + .maxItemCount(precedentPageSize) |
| 74 | + .queryString("SELECT p FROM Precedent p ORDER BY p.id ASC") |
| 75 | + .build(); |
| 76 | + } |
| 77 | +
|
| 78 | + @Bean |
| 79 | + public ItemProcessor<Precedent, List<Document>> precedentItemProcessor() { |
| 80 | +
|
| 81 | + return precedent -> { |
| 82 | + String content = precedent.getPrecedentContent(); |
| 83 | + if (content == null || content.isBlank()) return null; |
| 84 | +
|
| 85 | + Document originalDoc = new Document(content, Map.of( |
| 86 | + "type", "판례", |
| 87 | + "caseNumber", precedent.getCaseNumber(), |
| 88 | + "court", precedent.getCourtName(), |
| 89 | + "caseName", precedent.getCaseName() |
| 90 | + )); |
| 91 | +
|
| 92 | + List<Document> chunkDocs = tokenSplitter.split(originalDoc); |
| 93 | + List<Document> finalChunks = new ArrayList<>(); |
| 94 | +
|
| 95 | + // 청크별로 메타데이터에 인덱스 추가 -> 구분 용도 |
| 96 | + for (int i = 0; i < chunkDocs.size(); i++) { |
| 97 | + Document chunk = chunkDocs.get(i); |
| 98 | + Map<String, Object> newMetadata = new HashMap<>(chunk.getMetadata()); |
| 99 | + newMetadata.put("chunkIndex", i); |
| 100 | + finalChunks.add(new Document(chunk.getText(), newMetadata)); |
| 101 | + } |
| 102 | + return finalChunks; |
| 103 | + }; |
| 104 | + } |
| 105 | +
|
| 106 | + // -------------- 법령 백터화 --------------- |
| 107 | + @Bean |
| 108 | + public Step lawVectorizationStep() { |
| 109 | + log.info(">>>>>> 법령 벡터화 시작"); |
| 110 | + return new StepBuilder("lawVectorizationStep", jobRepository) |
| 111 | + .<Law, List<Document>>chunk(CHUNK_SIZE, transactionManager) // 법령은 한 번에 10개씩 처리 |
| 112 | + .reader(lawItemReader()) |
| 113 | + .processor(lawItemProcessor()) |
| 114 | + .writer(documentItemWriter()) |
| 115 | + .taskExecutor(taskExecutor()) |
| 116 | + .build(); |
| 117 | + } |
| 118 | +
|
| 119 | + @Bean |
| 120 | + public JpaPagingItemReader<Law> lawItemReader() { |
| 121 | + return new JpaPagingItemReaderBuilder<Law>() |
| 122 | + .name("lawItemReader") |
| 123 | + .entityManagerFactory(entityManagerFactory) |
| 124 | + .pageSize(CHUNK_SIZE) |
| 125 | + .maxItemCount(lawPageSize) |
| 126 | + .queryString("SELECT l FROM Law l ORDER BY l.id ASC") |
| 127 | + .build(); |
| 128 | + } |
| 129 | +
|
| 130 | + @Bean |
| 131 | + public ItemProcessor<Law, List<Document>> lawItemProcessor() { |
| 132 | + return law -> { |
| 133 | + List<Document> finalChunks = new ArrayList<>(); |
| 134 | +
|
| 135 | + List<Jang> jangs = jangRepository.findByLaw(law); |
| 136 | +
|
| 137 | + for (Jang jang : jangs) { |
| 138 | +
|
| 139 | + StringBuilder contentBuilder = new StringBuilder(); |
| 140 | +
|
| 141 | + contentBuilder.append(law.getLawName()).append("\n"); |
| 142 | +
|
| 143 | + if (jang.getContent() != null && !jang.getContent().isBlank()) { |
| 144 | + contentBuilder.append(jang.getContent()).append("\n"); |
| 145 | + } |
| 146 | +
|
| 147 | + List<Jo> jos = joRepository.findByJang(jang); |
| 148 | + for (Jo jo : jos) { |
| 149 | +
|
| 150 | + if (jo.getContent() != null && !jo.getContent().isBlank()) { |
| 151 | + contentBuilder.append(jo.getContent()).append("\n"); |
| 152 | + } |
| 153 | +
|
| 154 | + List<Hang> hangs = hangRepository.findByJo(jo); |
| 155 | + for (Hang hang : hangs) { |
| 156 | + if (hang.getContent() != null && !hang.getContent().isBlank()) { |
| 157 | + contentBuilder.append(hang.getContent()).append("\n"); |
| 158 | + } |
| 159 | +
|
| 160 | + List<Ho> hos = hoRepository.findByHang(hang); |
| 161 | + for (Ho ho : hos) { |
| 162 | + if (ho.getContent() != null && !ho.getContent().isBlank()) { |
| 163 | + contentBuilder.append(ho.getContent()).append("\n"); |
| 164 | + } |
| 165 | + } |
| 166 | + } |
| 167 | + } |
| 168 | +
|
| 169 | + // === Jang 단위로 문서화 === |
| 170 | + String finalContent = contentBuilder.toString(); |
| 171 | +
|
| 172 | + if (!finalContent.isBlank()) { |
| 173 | + Map<String, Object> metadata = new HashMap<>(); |
| 174 | + metadata.put("type", "법령"); |
| 175 | + metadata.put("lawName", law.getLawName()); |
| 176 | + metadata.put("jangId", jang.getId()); |
| 177 | +
|
| 178 | + Document originalDoc = new Document(finalContent, metadata); |
| 179 | +
|
| 180 | + List<Document> chunkDocs = tokenSplitter.split(originalDoc); |
| 181 | +
|
| 182 | + for (int i = 0; i < chunkDocs.size(); i++) { |
| 183 | + Document chunk = chunkDocs.get(i); |
| 184 | + Map<String, Object> newMetadata = new HashMap<>(chunk.getMetadata()); |
| 185 | + newMetadata.put("chunkIndex", i); |
| 186 | + finalChunks.add(new Document(chunk.getText(), newMetadata)); |
| 187 | + } |
| 188 | + } |
| 189 | + } |
| 190 | +
|
| 191 | + return finalChunks.isEmpty() ? null : finalChunks; |
| 192 | + }; |
| 193 | + } |
| 194 | +
|
| 195 | + @Bean |
| 196 | + public ItemWriter<List<Document>> documentItemWriter() { |
| 197 | + return chunk -> { |
| 198 | + List<Document> totalDocuments = chunk.getItems().stream() |
| 199 | + .flatMap(List::stream) |
| 200 | + .collect(Collectors.toList()); |
| 201 | +
|
| 202 | + if (!totalDocuments.isEmpty()) { |
| 203 | + vectorStore.add(totalDocuments); |
| 204 | + log.info(">>>>>> {}개의 Document 청크를 벡터 저장소에 저장했습니다.", totalDocuments.size()); |
| 205 | + } |
| 206 | + }; |
| 207 | + } |
| 208 | +}*/ |
| 209 | + |
0 commit comments