feat(aihr): keyframe caption and rerank in retrieval pipeline

视频关键帧 vision prompt 从纯 OCR 扩展为「文字+一句画面描述」,纯视觉
信息(动作/场景演示)进入文本向量可检索;检索 RRF 融合后接
category=rerank 模型(bge-reranker-v2-m3)语义重排,未配置或失败保持
RRF 顺序(TechSpec 12.3 最后一环)。BACKLOG 记入 B7 视觉向量检索
(Doubao-embedding-vision,二期)。端到端验证:重传视频片段带画面描述;
检索「首次反馈时限」视频转写片段重排后命中第一。

backend/ruoyi-modules/ruoyi-aihr/src/main/java/org/dromara/aihr/service/AihrSopSeedService.java

@@ -720,7 +720,17 @@ public class AihrSopSeedService {
         }
     }
 
+    private static final String OCR_PROMPT =
+        "请识别图片中的所有文字内容，保持原有格式和段落结构，直接输出识别到的文字，不要添加任何解释说明。如果图片不含可识别文字，返回空字符串。";
+    /** 视频关键帧：文字 + 一句画面描述，让纯视觉信息（动作/场景演示）也能进入文本向量检索。 */
+    private static final String FRAME_PROMPT =
+        "这是一段物业培训/工作视频的画面截图。请输出：1）画面中的可见文字（保持原文）；2）另起一行，以「(画面：…)」格式用一句话描述画面正在发生的内容（场景、人物动作、演示要点）。不要添加其他解释。";
+
     private String callVisionOcr(ChatRuntime runtime, byte[] imageBytes, String mimeType) throws Exception {
+        return callVisionOcr(runtime, imageBytes, mimeType, OCR_PROMPT);
+    }
+
+    private String callVisionOcr(ChatRuntime runtime, byte[] imageBytes, String mimeType, String prompt) throws Exception {
         String dataUrl = "data:" + mimeType + ";base64," + Base64.getEncoder().encodeToString(imageBytes);
 
         ObjectNode body = objectMapper.createObjectNode();
@@ -735,7 +745,7 @@ public class AihrSopSeedService {
 
         ObjectNode textPart = content.addObject();
         textPart.put("type", "text");
-        textPart.put("text", "请识别图片中的所有文字内容，保持原有格式和段落结构，直接输出识别到的文字，不要添加任何解释说明。如果图片不含可识别文字，返回空字符串。");
+        textPart.put("text", prompt);
 
         ObjectNode imagePart = content.addObject();
         imagePart.put("type", "image_url");
@@ -1466,12 +1476,99 @@ public class AihrSopSeedService {
         vector[index] += (hash & 1) == 0 ? 1 : -1;
     }
 
+    private Optional<ChatRuntime> rerankRuntime() {
+        try {
+            List<ChatRuntime> rows = jdbcTemplate.query("""
+                select c.model_name,
+                       coalesce(nullif(c.api_host, ''), nullif(p.api_host, '')) as resolved_api_host,
+                       coalesce(nullif(c.api_key, ''), nullif(p.api_key, '')) as resolved_api_key
+                  from aihr_model_config c
+                  left join aihr_model_provider p
+                    on p.tenant_id = c.tenant_id and p.provider_code = c.provider_code
+                 where c.tenant_id = ?
+                   and c.category = 'rerank'
+                   and c.enabled = 1
+                   and (p.status is null or p.status = '0')
+                 order by case when c.model_show = 'Y' then 0 else 1 end, c.id asc
+                 limit 1
+                """, (rs, rowNum) -> new ChatRuntime(
+                rs.getString("model_name"),
+                rs.getString("resolved_api_host"),
+                rs.getString("resolved_api_key")
+            ), TENANT_ID);
+            return rows.stream()
+                .filter(runtime -> !isBlank(runtime.modelName()) && !isBlank(runtime.baseUrl()) && !isBlank(runtime.apiKey()))
+                .findFirst();
+        } catch (DataAccessException e) {
+            return Optional.empty();
+        }
+    }
+
+    /**
+     * RRF 融合后的候选交给 rerank 模型按语义相关性重排（TechSpec 12.3）。
+     * 未配置 rerank 模型或调用失败时保持 RRF 顺序，检索不受影响。
+     */
+    private List<KnowledgeHit> rerankHits(String queryText, List<KnowledgeHit> hits) {
+        if (hits.size() < 2) {
+            return hits;
+        }
+        Optional<ChatRuntime> runtime = rerankRuntime();
+        if (runtime.isEmpty()) {
+            return hits;
+        }
+        try {
+            ObjectNode body = objectMapper.createObjectNode();
+            body.put("model", runtime.get().modelName());
+            body.put("query", queryText);
+            body.put("top_n", hits.size());
+            body.put("return_documents", false);
+            ArrayNode documents = body.putArray("documents");
+            hits.forEach(hit -> documents.add(truncate(hit.content(), 2000)));
+
+            HttpRequest httpRequest = HttpRequest.newBuilder()
+                .uri(URI.create(normalizeBaseUrl(runtime.get().baseUrl()) + "/rerank"))
+                .timeout(Duration.ofSeconds(30))
+                .header("Content-Type", "application/json")
+                .header("Authorization", "Bearer " + runtime.get().apiKey())
+                .POST(HttpRequest.BodyPublishers.ofString(objectMapper.writeValueAsString(body)))
+                .build();
+            HttpResponse<String> response = HttpClient.newBuilder()
+                .connectTimeout(Duration.ofSeconds(15))
+                .build()
+                .send(httpRequest, HttpResponse.BodyHandlers.ofString());
+            if (!ok(response.statusCode())) {
+                throw new IllegalStateException("rerank HTTP " + response.statusCode());
+            }
+            JsonNode results = objectMapper.readTree(response.body()).path("results");
+            if (!results.isArray() || results.isEmpty()) {
+                return hits;
+            }
+            List<KnowledgeHit> reranked = new ArrayList<>();
+            for (JsonNode item : results) {
+                int index = item.path("index").asInt(-1);
+                if (index >= 0 && index < hits.size()) {
+                    reranked.add(hits.get(index));
+                }
+            }
+            // 模型漏返回的候选按原序补到末尾，保证不丢结果
+            for (KnowledgeHit hit : hits) {
+                if (!reranked.contains(hit)) {
+                    reranked.add(hit);
+                }
+            }
+            return reranked;
+        } catch (Exception e) {
+            log.warn("rerank failed, keep rrf order: {}", e.getMessage());
+            return hits;
+        }
+    }
+
     private SearchResponse dbSearch(String category, String queryText, Integer limit) {
         try {
             int limitValue = limitValue(limit);
             List<KnowledgeHit> keywordHits = likeHits(category, queryText, limitValue);
             List<KnowledgeHit> textHits = rrf(keywordHits, fulltextHits(category, queryText, limitValue), limitValue);
-            List<KnowledgeHit> hits = rrf(textHits, vectorHits(category, queryText, limitValue), limitValue);
+            List<KnowledgeHit> hits = rerankHits(queryText, rrf(textHits, vectorHits(category, queryText, limitValue), limitValue));
             if (hits.isEmpty()) {
                 return null;
             }
@@ -2032,7 +2129,7 @@ public class AihrSopSeedService {
                     return Optional.empty();
                 }
                 try {
-                    return Optional.of(normalizeExtractedText(callVisionOcr(runtime.get(), bytes, mimeType)));
+                    return Optional.of(normalizeExtractedText(callVisionOcr(runtime.get(), bytes, mimeType, FRAME_PROMPT)));
                 } catch (Exception e) {
                     log.warn("video frame ocr failed for {}: {}", fileName, e.getMessage());
                     return Optional.empty();