diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/client/AppearancePatentLlmClient.java b/backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/client/AppearancePatentLlmClient.java index 457825493..26ac304be 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/client/AppearancePatentLlmClient.java +++ b/backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/client/AppearancePatentLlmClient.java @@ -226,12 +226,13 @@ public class AppearancePatentLlmClient { List chunk, String apiKey) { try { - return new BatchOutcome<>(invokeTitleBatch(rows, chunk, apiKey), null); + return BatchOutcome.ok(invokeTitleBatch(rows, chunk, apiKey)); } catch (Exception ex) { String message = failureMessage(ex); - log.warn("[appearance-patent] title batch failed rows={} firstAsin={} err={}", - chunk.size(), rows.get(chunk.get(0)).getAsin(), message); - return new BatchOutcome<>(Map.of(), message); + log.warn("[appearance-patent] title batch failed rows={} firstAsin={} content={} err={}", + chunk.size(), rows.get(chunk.get(0)).getAsin(), + ex instanceof BatchContentException, message); + return BatchOutcome.failed(ex, message); } } @@ -240,18 +241,21 @@ public class AppearancePatentLlmClient { String prompt, String apiKey) { try { - return new BatchOutcome<>(invokeAppearanceBatch(rows, chunk, prompt, apiKey), null); + return BatchOutcome.ok(invokeAppearanceBatch(rows, chunk, prompt, apiKey)); } catch (Exception ex) { String message = failureMessage(ex); - log.warn("[appearance-patent] appearance batch failed rows={} firstAsin={} err={}", - chunk.size(), rows.get(chunk.get(0)).getAsin(), message); - return new BatchOutcome<>(Map.of(), message); + log.warn("[appearance-patent] appearance batch failed rows={} firstAsin={} content={} err={}", + chunk.size(), rows.get(chunk.get(0)).getAsin(), + ex instanceof BatchContentException, message); + return BatchOutcome.failed(ex, message); } } /** * 阶段调度:切批并发发请求 → 漏行/失真行单发补检一次 → 仍缺失的行记失败。 * 传输失败按整批记失败,绝不逐行重发(一次请求带走一批,劣化窗口下不会放大成 N 倍); + * 正文读不出来(contentFailure)按失真处理——整批行进 missing 走同一条补检路, + * 补检仍失败才落回退,代价是每行一次单图请求,换回的是行上有结论而不是「外观识别异常」。 * 补检请求不再套校验门,既保留今天对模型自身过度提取的容忍,也杜绝无限重做。 */ private void pass(List indexes, int batchSize, BatchInvoker invoker, @@ -266,11 +270,13 @@ public class AppearancePatentLlmClient { for (int c = 0; c < chunks.size(); c++) { List chunk = chunks.get(c); BatchOutcome outcome = outcomes.get(c); - if (outcome.transportError() != null) { + if (outcome.failure() != null) { for (int index : chunk) { - errors.put(index, outcome.transportError()); + errors.put(index, outcome.failure()); + } + if (!outcome.contentFailure()) { + continue; } - continue; } for (int index : chunk) { T value = outcome.byIndex().get(index); @@ -298,7 +304,34 @@ public class AppearancePatentLlmClient { BatchOutcome invoke(List chunk); } - private record BatchOutcome(Map byIndex, String transportError) { + /** + * 一批请求的产物。failure 非空=这批没拿到可用结论,contentFailure 决定后续怎么走: + * + *
    + *
  • contentFailure=true:上游回了 200、正文也拿到了,只是按批量协议读不出来。 + * 这是模型失真(实测到它跳过 items 信封直接回数组),换单行请求重问往往就读得出来, + * 所以整批行要进补检,而不是按失败收口——2026-10-05 报表「有行没结论」就是这么来的。
  • + *
  • contentFailure=false:传输层失败(超时/5xx/响应体不是 JSON)。劣化窗口里逐行重发 + * 只会把一次抖动放大成 N 次请求,维持整批记失败。
  • + *
+ */ + private record BatchOutcome(Map byIndex, String failure, boolean contentFailure) { + + static BatchOutcome ok(Map byIndex) { + return new BatchOutcome<>(byIndex, null, false); + } + + /** 失败分类由异常类型决定,不允许调用方各自猜文案。 */ + static BatchOutcome failed(Exception ex, String message) { + return new BatchOutcome<>(Map.of(), message, ex instanceof BatchContentException); + } + } + + /** 正文到手但按批量协议读不出来:交给单行补检,不按传输失败整批记死。 */ + private static final class BatchContentException extends RuntimeException { + BatchContentException(String message, Throwable cause) { + super(message, cause); + } } private static List> partition(List indexes, int batchSize) { @@ -326,7 +359,7 @@ public class AppearancePatentLlmClient { } } catch (InterruptedException interruptedException) { Thread.currentThread().interrupt(); - return new BatchOutcome(Map.of(), "LLM 批次等待被中断"); + return new BatchOutcome(Map.of(), "LLM 批次等待被中断", false); } }; futures.add(CompletableFuture.supplyAsync(SecretUsageContext.wrap(task, usageContext), executor)); @@ -336,9 +369,10 @@ public class AppearancePatentLlmClient { try { outcomes.add(futures.get(i).join()); } catch (Exception ex) { + // 调度层炸了(invoker 本该自己 catch 干净):按传输失败收口,宁可整批记失败也不逐行放大 log.warn("[appearance-patent] llm batch crashed rows={} err={}", chunks.get(i).size(), failureMessage(ex)); - outcomes.add(new BatchOutcome(Map.of(), failureMessage(ex))); + outcomes.add(new BatchOutcome(Map.of(), failureMessage(ex), false)); } } return outcomes; @@ -495,12 +529,24 @@ public class AppearancePatentLlmClient { && !normalize(judgement.reason()).isBlank(); } - /** 批量响应解包:序号 → 条目,按序号对账而不是按数组位置(模型可能漏项或乱序)。 */ + /** + * 批量响应解包:序号 → 条目,按序号对账而不是按数组位置(模型可能漏项或乱序)。 + * 顶层直接回数组同样收下——条目本来就带 i,items 信封只是提示词要求的包装 + * (生产实测 2026-10-01 22:12、10-05 21:09 各一次,整批完好的结论因缺信封被丢弃)。 + * + *

本方法抛出的一律是 {@link BatchContentException}:正文已经到手,只是按协议读不出来, + * 与传输失败区分对待才有单行补检。 + */ private Map parseIndexedItems(String content) { - JsonNode root = readBatchRoot(content); - JsonNode items = root.path("items"); + JsonNode root; + try { + root = readBatchRoot(content); + } catch (RuntimeException ex) { + throw new BatchContentException(failureMessage(ex), ex); + } + JsonNode items = root.isArray() ? root : root.path("items"); if (!items.isArray()) { - throw new IllegalStateException("LLM 批量响应缺少 items 数组"); + throw new BatchContentException("LLM 批量响应缺少 items 数组", null); } Map byIndex = new LinkedHashMap<>(); for (JsonNode item : items) { diff --git a/backend-java/src/test/java/com/nanri/aiimage/modules/appearancepatent/client/AppearancePatentLlmClientHttpTest.java b/backend-java/src/test/java/com/nanri/aiimage/modules/appearancepatent/client/AppearancePatentLlmClientHttpTest.java index 230d8269b..84aeef316 100644 --- a/backend-java/src/test/java/com/nanri/aiimage/modules/appearancepatent/client/AppearancePatentLlmClientHttpTest.java +++ b/backend-java/src/test/java/com/nanri/aiimage/modules/appearancepatent/client/AppearancePatentLlmClientHttpTest.java @@ -43,6 +43,16 @@ class AppearancePatentLlmClientHttpTest { private volatile boolean crossAssignFirstTitleItem = false; private volatile boolean reverseTitleItems = false; private volatile boolean dropLastAppearanceItem = false; + /** + * 正文形状钩子:模拟「HTTP 200 但正文按批量协议读不出来」的上游形状。 + * garbled_batch 只污染多行请求(补检的单行请求照常成功),garbled_all 连单行也污染, + * top_level_array 是 2026-10-05 21:09 生产实测到的形状——条目自带 i,只是没包在 items 里。 + */ + private volatile String appearanceBodyShape = "normal"; + private volatile String titleBodyShape = "normal"; + + /** 没有任何 JSON 结构的正文:readBatchRoot 两次解析都会抛,等同生产那 23 次失败。 */ + private static final String GARBLED_BODY = "外观判定结论:第一件商品无侵权,第二件商品无侵权。以上为纯文本描述。"; private AppearancePatentLlmClient client; @@ -155,6 +165,9 @@ class AppearancePatentLlmClientHttpTest { List texts = new ArrayList<>(); parseIndexedLines(userTextOf(request), indexes, texts); boolean batch = indexes.size() > 1; + if (isGarbledBody(titleBodyShape, indexes.size())) { + return GARBLED_BODY; + } List> items = new ArrayList<>(); for (int i = 0; i < indexes.size(); i++) { if (batch && dropLastTitleItem && i == indexes.size() - 1) { @@ -180,6 +193,9 @@ class AppearancePatentLlmClientHttpTest { List indexes = new ArrayList<>(); List texts = new ArrayList<>(); parseIndexedLines(appearanceUserTextOf(request), indexes, texts); + if (isGarbledBody(appearanceBodyShape, indexes.size())) { + return GARBLED_BODY; + } List> items = new ArrayList<>(); for (int i = 0; i < indexes.size(); i++) { if (indexes.size() > 1 && dropLastAppearanceItem && i == indexes.size() - 1) { @@ -189,11 +205,19 @@ class AppearancePatentLlmClientHttpTest { "appearance_status", texts.get(i).contains("原创个性杯") ? "侵权" : "无侵权", "appearance_reason", appearanceReasonFor(texts.get(i)))); } + if ("top_level_array".equals(appearanceBodyShape)) { + return objectMapper.writeValueAsString(items); + } String json = objectMapper.writeValueAsString(Map.of("items", items)); boolean dirtySingle = indexes.size() <= 1 && !texts.isEmpty() && texts.get(0).contains("原创个性杯"); return dirtySingle ? "```json\n" + json + "\n```" : json; } + /** 污染范围按请求行数决定:只污染批量才能验证「整批读不出来时改为逐行补检」。 */ + private static boolean isGarbledBody(String shape, int requestedRows) { + return "garbled_all".equals(shape) || ("garbled_batch".equals(shape) && requestedRows > 1); + } + private String appearanceReasonFor(String text) { return text.contains("原创个性杯") ? "【视觉拆解】:特殊造型\n【判定依据】:高度相似知名设计" @@ -555,6 +579,80 @@ class AppearancePatentLlmClientHttpTest { } } + /** + * 外观批量正文读不出来(HTTP 200)=和漏行同一类失真:必须逐行补检把结论拿回来, + * 不允许一次上游抖动把整批行直接打成「外观识别异常」——那正是报表里「有行没结论」的来源, + * 而且外层会把这批行标成已检测,用户重发任务也捞不回来。 + */ + @Test + void appearanceBatchGarbledBodyRedoesRowsIndividually() { + appearanceBodyShape = "garbled_batch"; + List rows = List.of( + row("1", "F001", "普通直筒杯A", "", "https://img.example.com/f1.jpg"), + row("2", "F002", "普通直筒杯B", "", "https://img.example.com/f2.jpg"), + row("3", "F003", "原创个性杯", "", "https://img.example.com/f3.jpg")); + + List result = client.inspectRows(rows, null, "test-key"); + + assertThat(imageRequestCount.get()).isEqualTo(4); + assertThat(result.get(0).getAppearanceRisk()).isEqualTo("无侵权"); + assertThat(result.get(1).getAppearanceRisk()).isEqualTo("无侵权"); + assertThat(result.get(2).getAppearanceRisk()).isEqualTo("侵权"); + } + + /** 补检有上界:单行补检也读不出来时,每行只做一次补检就落回退,绝不套娃重发。 */ + @Test + void appearanceBatchGarbledBodyRedoIsBoundedToOnePass() { + appearanceBodyShape = "garbled_all"; + List rows = new ArrayList<>(); + for (int i = 1; i <= 3; i++) { + rows.add(row(String.valueOf(i), "G00" + i, "普通直筒杯" + i, "", "https://img.example.com/g" + i + ".jpg")); + } + + List result = client.inspectRows(rows, null, "test-key"); + + assertThat(imageRequestCount.get()).isEqualTo(4); + for (AppearancePatentResultRowDto item : result) { + assertThat(item.getAppearanceRisk()).isEqualTo("外观识别异常"); + } + } + + /** + * 顶层数组正文(生产实测 2026-10-05 21:09):条目自带 i,只是没包在 items 里, + * 按既有整串解包口径(parseResults)同样收下,不许白扔一整批。 + */ + @Test + void appearanceBatchTopLevelArrayBodyIsRead() { + appearanceBodyShape = "top_level_array"; + List rows = List.of( + row("1", "H001", "普通直筒杯A", "", "https://img.example.com/h1.jpg"), + row("2", "H002", "原创个性杯", "", "https://img.example.com/h2.jpg")); + + List result = client.inspectRows(rows, null, "test-key"); + + assertThat(imageRequestCount.get()).isEqualTo(1); + assertThat(result.get(0).getAppearanceRisk()).isEqualTo("无侵权"); + assertThat(result.get(1).getAppearanceRisk()).isEqualTo("侵权"); + assertThat(result.get(1).getConclusion()).isEqualTo("侵权"); + } + + /** 商标提取那一路共用同一套批次调度:整批正文读不出来时同样逐行补检。 */ + @Test + void titleBatchGarbledBodyRedoesRowsIndividually() { + titleBodyShape = "garbled_batch"; + List rows = List.of( + row("1", "I001", "Apple Magic Case", "", "https://img.example.com/i1.jpg"), + row("2", "I002", "Nike Air 鞋", "", "https://img.example.com/i2.jpg")); + + List result = client.inspectRows(rows, null, "test-key"); + + assertThat(textRequestCount.get()).isEqualTo(3); + assertThat(result.get(0).getTitleReason()).isEqualTo("Apple"); + assertThat(result.get(1).getTitleReason()).isEqualTo("Nike"); + assertThat(result.get(0).getAppearanceRisk()).isEqualTo("无侵权"); + assertThat(result.get(1).getAppearanceRisk()).isEqualTo("无侵权"); + } + private static int countOccurrences(String haystack, String needle) { int count = 0; for (int i = haystack.indexOf(needle); i >= 0; i = haystack.indexOf(needle, i + needle.length())) {