feat(aihr): harden finance RAG retrieval and evidence gates

backend/ruoyi-modules/ruoyi-aihr/src/main/java/org/dromara/aihr/knowledge/domain/AihrKnowledgeQueryDto.java

@@ -57,22 +57,46 @@ public final class AihrKnowledgeQueryDto {
         String updatedAt,
         String mediaType,
         String detailRef,
-        LocatorSummary locatorSummary
+        LocatorSummary locatorSummary,
+        String sourceAuthority,
+        String sourceKind
     ) {
+        public Citation(String spaceCode, String sourceType, String docId, String title, String snippet,
+                        Long fragmentId, String domain, String status, String occurredAt, String updatedAt,
+                        String mediaType, String detailRef, LocatorSummary locatorSummary) {
+            this(spaceCode, sourceType, docId, title, snippet, fragmentId, domain, status, occurredAt,
+                updatedAt, mediaType, detailRef, locatorSummary, null, null);
+        }
+
         public Citation(String spaceCode, String sourceType, String docId, String title, String snippet,
                         Long fragmentId) {
             this(spaceCode, sourceType, docId, title, snippet, fragmentId,
                 "DOCUMENT".equals(sourceType) ? "ENTERPRISE" : sourceType, null, null, null,
-                null, null, null);
+                null, null, null, null, null);
         }
 
         public Citation(String spaceCode, String sourceType, String docId, String title, String snippet,
                         Long fragmentId, String domain, String status, String occurredAt, String updatedAt) {
             this(spaceCode, sourceType, docId, title, snippet, fragmentId, domain, status, occurredAt,
-                updatedAt, null, null, null);
+                updatedAt, null, null, null, null, null);
         }
     }
 
+    public record QueryEntity(Double amount, String currency, String comparator, String amountBasis) {}
+
+    public record QueryPlan(String originalQuery, String normalizedQuery, List<String> alternateQueries,
+                            String intent, String domain, QueryEntity entities,
+                            List<String> requiredAuthorities, boolean needsClarification,
+                            List<String> evidenceNeeds, List<String> clarificationReasons) {}
+
+    public record ChannelContribution(String channel, int channelRank, Double rawScore,
+                                      Double fusionScore, String queryVariant) {}
+
+    public record RetrievalCandidate(Long fragmentId, String title, String sourceAuthority, String sourceKind,
+                                     Integer candidateRank, Double fusionScore, Double rerankScore,
+                                     String rerankModel, boolean rerankApplied, boolean selectedForEvidence,
+                                     String decisionReason, List<ChannelContribution> contributions) {}
+
     public record LocatorSummary(Integer pageNumber, Integer slideNumber, Integer paragraphStart,
                                  Integer paragraphEnd, String sheetName, Integer rowStart, Integer rowEnd,
                                  Long startMs, Long endMs, Long frameMs) {}
@@ -127,8 +151,20 @@ public final class AihrKnowledgeQueryDto {
         String rewrittenQuery,
         List<Resource> resources,
         MemoryCandidateResponse memoryCandidate,
-        BroadcastContext broadcastContext
+        BroadcastContext broadcastContext,
+        QueryPlan queryPlan,
+        List<RetrievalCandidate> retrievalCandidates
     ) {
+        public QueryResponse(String requestId, String queryText, String answer, List<Citation> citations,
+                             List<String> usedSpaceCodes, boolean noEvidence, String promptVersion,
+                             SearchResponse legacy, Object data, String conversationId, Long contextVersion,
+                             String intent, String rewrittenQuery, List<Resource> resources,
+                             MemoryCandidateResponse memoryCandidate, BroadcastContext broadcastContext) {
+            this(requestId, queryText, answer, citations, usedSpaceCodes, noEvidence, promptVersion,
+                legacy, data, conversationId, contextVersion, intent, rewrittenQuery, resources,
+                memoryCandidate, broadcastContext, null, List.of());
+        }
+
         public QueryResponse(String requestId, String queryText, String answer, List<Citation> citations,
                              List<String> usedSpaceCodes, boolean noEvidence, String promptVersion,
                              SearchResponse legacy, Object data, String conversationId, Long contextVersion,
@@ -157,7 +193,7 @@ public final class AihrKnowledgeQueryDto {
         public QueryResponse withMemoryCandidate(MemoryCandidateResponse candidate) {
             return new QueryResponse(requestId, queryText, answer, citations, usedSpaceCodes, noEvidence,
                 promptVersion, legacy, data, conversationId, contextVersion, intent, rewrittenQuery,
-                resources, candidate, broadcastContext);
+                resources, candidate, broadcastContext, queryPlan, retrievalCandidates);
         }
     }
 }

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

@@ -7,6 +7,8 @@ import lombok.extern.slf4j.Slf4j;
 import org.dromara.aihr.knowledge.domain.AihrKnowledgeAppDto.AuthenticatedApp;
 import org.dromara.aihr.knowledge.domain.AihrKnowledgePrincipal;
 import org.dromara.aihr.domain.AihrSopDto.SnippetResponse;
+import org.dromara.aihr.knowledge.domain.AihrKnowledgeQueryDto.ChannelContribution;
+import org.dromara.aihr.knowledge.domain.AihrKnowledgeQueryDto.RetrievalCandidate;
 import org.springframework.dao.DataAccessException;
 import org.springframework.jdbc.core.JdbcTemplate;
 import org.springframework.stereotype.Service;
@@ -105,6 +107,37 @@ public class AihrKnowledgeQueryAuditService {
         recordEvidence(requestId, tenantId, snippets, fragmentIds);
     }
 
+    public void recordCandidates(String requestId, String tenantId, List<RetrievalCandidate> candidates) {
+        if (candidates == null || candidates.isEmpty()) return;
+        for (RetrievalCandidate candidate : candidates) {
+            if (candidate == null || candidate.fragmentId() == null || candidate.fragmentId() <= 0) continue;
+            List<ChannelContribution> contributions = candidate.contributions() == null
+                ? List.of() : candidate.contributions();
+            for (ChannelContribution contribution : contributions) {
+                try {
+                    jdbcTemplate.update("""
+                        insert into aihr_query_candidate_trace
+                          (tenant_id, request_id, fragment_id, candidate_rank, channel, channel_rank,
+                           raw_score, fusion_score, rerank_score, rerank_model, rerank_applied,
+                           source_authority, source_kind, query_variant, selected_for_evidence,
+                           decision_reason, create_time)
+                        values (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, now())
+                        """, tenantId, requestId, candidate.fragmentId(), candidate.candidateRank(),
+                        normalizedChannel(contribution.channel()), contribution.channelRank(),
+                        contribution.rawScore(), candidate.fusionScore(), candidate.rerankScore(),
+                        bounded(candidate.rerankModel(), 100), candidate.rerankApplied(),
+                        bounded(candidate.sourceAuthority(), 30), bounded(candidate.sourceKind(), 30),
+                        bounded(contribution.queryVariant(), 30), candidate.selectedForEvidence(),
+                        bounded(candidate.decisionReason(), 64));
+                } catch (DataAccessException ex) {
+                    log.warn("knowledge query candidate trace write failed for request {} fragment {}",
+                        requestId, candidate.fragmentId());
+                    break;
+                }
+            }
+        }
+    }
+
     private static String normalizedChannel(String value) {
         if (value == null || value.isBlank()) {
             return "UNKNOWN";
@@ -113,6 +146,12 @@ public class AihrKnowledgeQueryAuditService {
         return normalized.length() > 20 ? normalized.substring(0, 20) : normalized;
     }
 
+    private static String bounded(String value, int max) {
+        if (value == null || value.isBlank()) return null;
+        String normalized = value.trim();
+        return normalized.length() <= max ? normalized : normalized.substring(0, max);
+    }
+
     private String json(List<String> values) {
         try {
             return objectMapper.writeValueAsString(values == null ? List.of() : values);

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

@@ -22,6 +22,7 @@ import org.dromara.aihr.memory.AihrMemoryDto.ServiceMemoryRecall;
 import org.dromara.aihr.memory.AihrMemoryService;
 import org.dromara.aihr.service.AihrModelSeedService;
 import org.dromara.aihr.service.AihrSopSeedService;
+import org.dromara.aihr.service.AihrSopSeedService.SearchExecution;
 import org.dromara.common.core.constant.HttpStatus;
 import org.dromara.common.core.enums.UserType;
 import org.dromara.common.core.exception.ServiceException;
@@ -368,11 +369,17 @@ public class AihrKnowledgeQueryService {
                     principal == null ? Set.of() : principal.roles());
             SearchRequest searchRequest = new SearchRequest(
                 request.queryText(), request.category(), request.position(), retrievalSource, request.limit());
-            SearchResponse legacy = retrievalQuery.equals(request.queryText())
+            SearchExecution execution = sopService.searchAuthorizedDetailed(searchRequest,
-                ? sopService.searchAuthorized(searchRequest,
+                principal == null ? "" : principal.extPartyId(), spaceIds, retrievalQuery);
-                    principal == null ? "" : principal.extPartyId(), spaceIds)
+            if (execution == null) {
-                : sopService.searchAuthorized(searchRequest,
+                SearchResponse fallback = retrievalQuery.equals(request.queryText())
-                    principal == null ? "" : principal.extPartyId(), spaceIds, retrievalQuery);
+                    ? sopService.searchAuthorized(searchRequest,
+                        principal == null ? "" : principal.extPartyId(), spaceIds)
+                    : sopService.searchAuthorized(searchRequest,
+                        principal == null ? "" : principal.extPartyId(), spaceIds, retrievalQuery);
+                execution = new SearchExecution(fallback, null, List.of());
+            }
+            SearchResponse legacy = execution.response();
             List<Citation> citations = citations(
                 app.tenantId(), spaceIds, legacy.snippets(), formalPolicyOnly, request.queryText());
             boolean noEvidence = citations.isEmpty();
@@ -386,8 +393,10 @@ public class AihrKnowledgeQueryService {
                 citations.isEmpty() ? List.of() : List.of("DOCUMENT"), status, latency, legacy.promptVersion());
             auditService.recordEvidence(requestId, app.tenantId(), legacy.snippets(), citations.stream()
                 .map(Citation::fragmentId).filter(Objects::nonNull).toList());
+            auditService.recordCandidates(requestId, app.tenantId(), execution.candidates());
             return new QueryResponse(requestId, displayQuery, answer, citations, scopeCodes,
-                noEvidence, legacy.promptVersion(), legacy, null);
+                noEvidence, legacy.promptVersion(), legacy, null, null, null, "QA", request.queryText(),
+                List.of(), null, null, execution.queryPlan(), execution.candidates());
         } catch (RuntimeException ex) {
             auditService.record(requestId, principal, app, displayQuery, scopeCodes, List.of(),
                 ex instanceof ServiceException serviceEx && serviceEx.getCode() != null && serviceEx.getCode() < 500
@@ -419,7 +428,8 @@ public class AihrKnowledgeQueryService {
             documentResponse.usedSpaceCodes(), false, BROADCAST_PROMPT_VERSION, documentResponse.legacy(), null,
             documentResponse.conversationId(), documentResponse.contextVersion(), "QA",
             documentResponse.rewrittenQuery(), documentResponse.resources(), documentResponse.memoryCandidate(),
-            new BroadcastContext(broadcast.messageId(), broadcast.title(), broadcast.publishedAt())
+            new BroadcastContext(broadcast.messageId(), broadcast.title(), broadcast.publishedAt()),
+            documentResponse.queryPlan(), documentResponse.retrievalCandidates()
         );
     }
 
@@ -536,7 +546,8 @@ public class AihrKnowledgeQueryService {
         return new QueryResponse(
             response.requestId(), question, answer, citations, response.usedSpaceCodes(), noEvidence,
             response.promptVersion(), response.legacy(), response.data(), context.conversationId(), nextVersion,
-            context.intent(), context.rewrittenQuery(), resources, response.memoryCandidate(), response.broadcastContext()
+            context.intent(), context.rewrittenQuery(), resources, response.memoryCandidate(), response.broadcastContext(),
+            response.queryPlan(), response.retrievalCandidates()
         );
     }
 
@@ -564,7 +575,8 @@ public class AihrKnowledgeQueryService {
             String ref = citationDetailService.issue(citation, tenantId);
             return new Citation(citation.spaceCode(), citation.sourceType(), citation.docId(), citation.title(),
                 citation.snippet(), citation.fragmentId(), citation.domain(), citation.status(), citation.occurredAt(),
-                citation.updatedAt(), location.sourceKind(), ref, location.summary());
+                citation.updatedAt(), location.sourceKind(), ref, location.summary(),
+                citation.sourceAuthority(), citation.sourceKind());
         }).toList();
     }
 
@@ -611,7 +623,7 @@ public class AihrKnowledgeQueryService {
             response.requestId(), response.queryText(), answer, List.copyOf(citations), List.copyOf(usedScopes),
             false, response.promptVersion(), response.legacy(), response.data(), response.conversationId(),
             response.contextVersion(), response.intent(), response.rewrittenQuery(), response.resources(),
-            response.memoryCandidate()
+            response.memoryCandidate(), response.broadcastContext(), response.queryPlan(), response.retrievalCandidates()
         );
     }
 
@@ -720,26 +732,7 @@ public class AihrKnowledgeQueryService {
         args.add(tenantId);
         args.addAll(allowedSpaceIds);
         args.addAll(fragmentIds);
-        String formalSourceClause = formalPolicyOnly ? """
+        String formalSourceClause = "";
-               and exists (
-                   select 1
-                     from aihr_knowledge_attach governed_attach
-                     join aihr_knowledge_source_governance governance
-                       on governance.tenant_id = governed_attach.tenant_id
-                      and governance.attachment_id = governed_attach.id
-                    where governed_attach.tenant_id = f.tenant_id
-                      and governed_attach.knowledge_id = f.knowledge_id
-                      and governed_attach.doc_id = f.doc_id
-                      and governed_attach.status = 2
-                      and governance.source_doc_id = f.doc_id
-                      and governance.authority_type = 'FORMAL_POLICY'
-                      and governance.lifecycle_status = 'APPROVED'
-                      and governance.effective_date <= current_date()
-                      and (governance.expires_date is null or governance.expires_date >= current_date())
-                      and char_length(trim(governance.source_version)) > 0
-                      and governance.content_sha256 regexp '^[0-9a-f]{64}$'
-               )
-            """ : "";
         List<Citation> rows = jdbcTemplate.query("""
             select k.code as space_code, f.doc_id,
                    coalesce((
@@ -756,39 +749,39 @@ public class AihrKnowledgeQueryService {
                           and governed_title_chunk.published_fragment_id = f.id
                         limit 1
                    ), a.name, k.name) as title,
-                   f.content, f.id as fragment_id
+                   f.content, f.id as fragment_id, governed_asset.source_authority,
-              from aihr_knowledge_fragment f
+                   governed_asset.source_type, governed_asset.source_name
-              join aihr_knowledge_info k on k.id = f.knowledge_id and k.tenant_id = f.tenant_id
+               from aihr_knowledge_fragment f
-              left join aihr_knowledge_attach a
+               join aihr_knowledge_info k on k.id = f.knowledge_id and k.tenant_id = f.tenant_id
-                on a.tenant_id = f.tenant_id and a.knowledge_id = f.knowledge_id and a.doc_id = f.doc_id
+               left join aihr_knowledge_attach a
-             where f.tenant_id = ?
+                 on a.tenant_id = f.tenant_id and a.knowledge_id = f.knowledge_id and a.doc_id = f.doc_id
-               and f.knowledge_id in (%s)
+               join aihr_chunk_revision governed_chunk
-               and f.id in (%s)
+                 on governed_chunk.tenant_id = f.tenant_id and governed_chunk.published_fragment_id = f.id
-               and exists (
+               join aihr_data_asset governed_asset
-                   select 1
+                 on governed_asset.tenant_id = governed_chunk.tenant_id
-                     from aihr_chunk_revision governed_chunk
+                and governed_asset.id = governed_chunk.asset_id
-                     join aihr_data_asset governed_asset
+                and coalesce(governed_asset.published_version_id, governed_asset.current_version_id) = governed_chunk.version_id
-                       on governed_asset.tenant_id = governed_chunk.tenant_id
+               join aihr_dataset_membership governed_dataset
-                      and governed_asset.id = governed_chunk.asset_id
+                 on governed_dataset.tenant_id = governed_chunk.tenant_id
-                      and coalesce(governed_asset.published_version_id, governed_asset.current_version_id) = governed_chunk.version_id
+                and governed_dataset.version_id = governed_chunk.version_id
-                     join aihr_dataset_membership governed_dataset
+                and governed_dataset.dataset_code = 'production'
-                       on governed_dataset.tenant_id = governed_chunk.tenant_id
+                and governed_dataset.status = 'ACTIVE'
-                      and governed_dataset.version_id = governed_chunk.version_id
+              where f.tenant_id = ?
-                      and governed_dataset.dataset_code = 'production'
+                and f.knowledge_id in (%s)
-                      and governed_dataset.status = 'ACTIVE'
+                and f.id in (%s)
-                    where governed_chunk.tenant_id = f.tenant_id
+                and governed_asset.lifecycle_status = 'PUBLISHED'
-                      and governed_chunk.published_fragment_id = f.id
+                and governed_asset.trust_level = 'HUMAN_VERIFIED'
-                      and governed_asset.lifecycle_status = 'PUBLISHED'
+                and (governed_asset.effective_from is null or governed_asset.effective_from <= current_date())
-                      and governed_asset.trust_level = 'HUMAN_VERIFIED'
+                and (governed_asset.effective_to is null or governed_asset.effective_to >= current_date())
-                      and (governed_asset.effective_from is null or governed_asset.effective_from <= current_date())
+                %s
-                      and (governed_asset.effective_to is null or governed_asset.effective_to >= current_date())
-               )
-               %s
             """.formatted(placeholders(allowedSpaceIds.size()), placeholders(fragmentIds.size()),
                 formalSourceClause),
             (rs, rowNum) -> new Citation(
                 rs.getString("space_code"), "DOCUMENT", rs.getString("doc_id"), rs.getString("title"),
-                rs.getString("content"), rs.getLong("fragment_id")
+                rs.getString("content"), rs.getLong("fragment_id"), "ENTERPRISE", null, null, null,
+                null, null, null, rs.getString("source_authority"),
+                AihrRagQueryPlanner.sourceKind(rs.getString("source_authority"), rs.getString("source_type"),
+                    rs.getString("source_name"))
             ), args.toArray());
         Map<Long, Citation> byId = new LinkedHashMap<>();
         rows.forEach(row -> byId.put(row.fragmentId(), row));
@@ -796,13 +789,15 @@ public class AihrKnowledgeQueryService {
             .map(byId::get)
             .filter(java.util.Objects::nonNull)
             .filter(row -> !formalPolicyOnly
-                || AihrFormalPolicyClassifier.isRelevantFormalEvidence(queryText, row.snippet()))
+                || (Set.of("FORMAL_POLICY", "OPERATING_MANUAL", "APPROVED_SOP").contains(row.sourceKind())
+                    && AihrFormalPolicyClassifier.isRelevantFormalEvidence(queryText, row.snippet())))
             .map(row -> new Citation(
                 row.spaceCode(), row.sourceType(), row.docId(), row.title(),
                 formalPolicyOnly
                     ? AihrFormalPolicyClassifier.relevantEvidenceSnippet(queryText, row.snippet(), 500)
                     : truncate(row.snippet(), 500),
-                row.fragmentId()
+                row.fragmentId(), row.domain(), row.status(), row.occurredAt(), row.updatedAt(),
+                row.mediaType(), row.detailRef(), row.locatorSummary(), row.sourceAuthority(), row.sourceKind()
             ))
             .toList();
     }

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

@@ -0,0 +1,236 @@
+package org.dromara.aihr.knowledge.service;
+
+import org.dromara.aihr.knowledge.domain.AihrKnowledgeQueryDto.QueryEntity;
+import org.dromara.aihr.knowledge.domain.AihrKnowledgeQueryDto.QueryPlan;
+
+import java.util.ArrayList;
+import java.util.LinkedHashSet;
+import java.util.List;
+import java.util.Locale;
+import java.util.Optional;
+import java.util.Set;
+import java.util.regex.Matcher;
+import java.util.regex.Pattern;
+
+/** Deterministic, authorization-free planning for bounded enterprise knowledge retrieval. */
+public final class AihrRagQueryPlanner {
+
+    public static final String FINANCE_REIMBURSEMENT = "FINANCE_REIMBURSEMENT";
+    public static final String FINANCE_POLICY = "FINANCE_POLICY";
+    public static final String GENERAL_KNOWLEDGE = "GENERAL_KNOWLEDGE";
+    private static final List<String> FORMAL_AUTHORITIES =
+        List.of("FORMAL_POLICY", "OPERATING_MANUAL", "APPROVED_SOP");
+    private static final Pattern ARABIC_AMOUNT = Pattern.compile(
+        "(?<![0-9A-Za-z])(\\d+(?:\\.\\d+)?)\\s*(?:元|块钱|块|人民币)");
+    private static final Pattern CHINESE_HUNDREDS = Pattern.compile("([一二三四五六七八九])百(?:\\s*(?:元|块钱|块))?");
+    private static final Pattern HAN_RUN = Pattern.compile("[\\p{IsHan}]{2,16}");
+    private static final Pattern LATIN_TERM = Pattern.compile("[A-Za-z][A-Za-z0-9_-]{1,31}");
+    private static final List<String> BUSINESS_TERMS = List.of(
+        "费用报销", "报销流程", "发票登记", "财务复审", "零星采购", "小额采购", "个人借款",
+        "采购审批", "报销单", "借款", "报销", "发票", "采购", "自采", "财务", "报备", "审批",
+        "材料", "流程", "额度", "月累计", "单笔"
+    );
+    private static final List<String> STOP_PHRASES = List.of(
+        "我有个", "告诉我", "请告诉", "请问", "帮我", "怎么走", "怎么", "如何", "一下", "这个",
+        "那个", "有没有", "能不能", "可以吗", "是什么", "的", "以下", "以内", "以上", "不超过"
+    );
+    private static final Set<String> STOP_TERMS = Set.of(
+        "我有", "有个", "个报", "告诉", "诉我", "怎么", "么走", "以下", "以内", "以上", "块钱",
+        "这个", "那个", "一下", "请问", "能不", "不能", "可以", "以吗"
+    );
+
+    private AihrRagQueryPlanner() {}
+
+    public static QueryPlan plan(String rawQuery) {
+        String original = clean(rawQuery);
+        QueryEntity entities = entities(original);
+        String domain = domain(original);
+        boolean finance = isFormalFinanceDomain(domain);
+        String normalized = normalize(original, domain, entities);
+        List<String> alternates = alternates(original, domain);
+        List<String> evidenceNeeds = evidenceNeeds(original, domain, entities);
+        List<String> reasons = new ArrayList<>();
+        if (FINANCE_REIMBURSEMENT.equals(domain) && entities.amount() != null
+            && "UNKNOWN".equals(entities.amountBasis())) {
+            reasons.add("金额口径不明确：请确认是单笔金额还是月累计额度");
+        }
+        return new QueryPlan(
+            original,
+            normalized,
+            alternates,
+            "KNOWLEDGE_QA",
+            domain,
+            entities,
+            finance ? FORMAL_AUTHORITIES : List.of(),
+            !reasons.isEmpty(),
+            evidenceNeeds,
+            List.copyOf(reasons)
+        );
+    }
+
+    public static List<String> sparseTerms(String queryText) {
+        String original = clean(queryText).toLowerCase(Locale.ROOT);
+        if (original.isBlank()) return List.of();
+        LinkedHashSet<String> terms = new LinkedHashSet<>();
+        for (String businessTerm : BUSINESS_TERMS) {
+            if (original.contains(businessTerm)) terms.add(businessTerm);
+        }
+        Matcher amountMatcher = ARABIC_AMOUNT.matcher(original);
+        StringBuffer withoutAmounts = new StringBuffer();
+        while (amountMatcher.find()) {
+            terms.add(amountMatcher.group(1) + "元");
+            amountMatcher.appendReplacement(withoutAmounts, " ");
+        }
+        amountMatcher.appendTail(withoutAmounts);
+        String text = normalizeChineseAmounts(withoutAmounts.toString())
+            .replaceAll("[^\\p{IsHan}A-Za-z0-9_-]+", " ");
+        for (String stop : STOP_PHRASES) text = text.replace(stop, " ");
+        Matcher hanMatcher = HAN_RUN.matcher(text);
+        while (hanMatcher.find() && terms.size() < 24) {
+            String run = hanMatcher.group();
+            if (!STOP_TERMS.contains(run)) terms.add(run);
+            for (int i = 0; i + 2 <= run.length() && terms.size() < 24; i++) {
+                String term = run.substring(i, i + 2);
+                if (!STOP_TERMS.contains(term)) terms.add(term);
+            }
+        }
+        Matcher latinMatcher = LATIN_TERM.matcher(text);
+        while (latinMatcher.find() && terms.size() < 24) terms.add(latinMatcher.group().toLowerCase(Locale.ROOT));
+        return terms.stream()
+            .filter(term -> !term.matches("\\d+"))
+            .filter(term -> !STOP_TERMS.contains(term))
+            .limit(24)
+            .toList();
+    }
+
+    public static boolean isFormalFinanceDomain(String domain) {
+        return FINANCE_REIMBURSEMENT.equals(domain) || FINANCE_POLICY.equals(domain);
+    }
+
+    public static String sourceKind(String sourceAuthority, String sourceType, String sourceName) {
+        String authority = clean(sourceAuthority).toUpperCase(Locale.ROOT);
+        String type = clean(sourceType).toUpperCase(Locale.ROOT);
+        String name = clean(sourceName).toLowerCase(Locale.ROOT);
+        if (authority.contains("EXPERIENCE") || authority.contains("EMPLOYEE_SUBMISSION")
+            || type.contains("INTERVIEW")) return "EXPERIENCE";
+        if (authority.contains("CASE") || type.contains("CASE")) return "CASE";
+        if (name.contains("操作手册") || authority.contains("OPERATING_MANUAL")) return "OPERATING_MANUAL";
+        if (name.contains("sop") || authority.contains("APPROVED_SOP")) return "APPROVED_SOP";
+        if (authority.contains("POLICY") || authority.contains("COMPANY_POLICY")
+            || authority.contains("FORMAL")) return "FORMAL_POLICY";
+        return "REFERENCE_MATERIAL";
+    }
+
+    public static boolean authorityAllowed(QueryPlan plan, String sourceKind) {
+        return plan == null || plan.requiredAuthorities().isEmpty()
+            || plan.requiredAuthorities().contains(sourceKind);
+    }
+
+    public static boolean domainMatches(QueryPlan plan, String title, String content) {
+        if (plan == null || !isFormalFinanceDomain(plan.domain())) return true;
+        String source = clean(title) + " " + clean(content);
+        if (FINANCE_REIMBURSEMENT.equals(plan.domain())) {
+            return containsAny(source, "费用报销", "发票登记", "报销单", "财务复审", "个人借款")
+                || (containsAny(title, "报销", "发票", "财务", "借款")
+                    && containsAny(source, "流程", "提交", "登记", "审批", "复审"));
+        }
+        return containsAny(title, "财务", "报销", "借款", "发票", "采购")
+            || containsAny(source, "财务管理", "费用报销", "个人借款", "采购审批");
+    }
+
+    private static String domain(String query) {
+        if (containsAny(query, "报销", "走账") && containsAny(query, "采购", "自采", "发票", "费用", "垫付")) {
+            return FINANCE_REIMBURSEMENT;
+        }
+        if (containsAny(query, "报销", "借款", "发票", "财务", "采购", "自采", "报备", "审批")) {
+            return FINANCE_POLICY;
+        }
+        return GENERAL_KNOWLEDGE;
+    }
+
+    private static QueryEntity entities(String query) {
+        Double amount = null;
+        Matcher arabic = ARABIC_AMOUNT.matcher(query);
+        if (arabic.find()) amount = Double.valueOf(arabic.group(1));
+        if (amount == null) {
+            Matcher chinese = CHINESE_HUNDREDS.matcher(query);
+            if (chinese.find()) amount = chineseDigit(chinese.group(1)) * 100.0;
+        }
+        String comparator = containsAny(query, "以下", "以内", "不超过", "最多") ? "LTE"
+            : containsAny(query, "超过", "大于") ? "GT"
+            : containsAny(query, "以上", "不少于", "至少") ? "GTE" : "UNSPECIFIED";
+        String basis = containsAny(query, "单笔", "每笔", "一笔") ? "PER_TRANSACTION"
+            : containsAny(query, "月累计", "每月", "当月", "一个月", "月度") ? "MONTHLY_TOTAL" : "UNKNOWN";
+        return new QueryEntity(amount, amount == null ? null : "CNY", comparator, basis);
+    }
+
+    private static String normalize(String original, String domain, QueryEntity entities) {
+        if (!isFormalFinanceDomain(domain)) return original;
+        LinkedHashSet<String> parts = new LinkedHashSet<>();
+        if (containsAny(original, "采购", "自采", "零星", "小额")) parts.add("零星采购");
+        parts.add(FINANCE_REIMBURSEMENT.equals(domain) ? "费用报销流程" : "财务制度");
+        if (containsAny(original, "发票")) parts.add("发票登记");
+        if (entities.amount() != null) {
+            String amount = entities.amount() % 1 == 0
+                ? String.valueOf(entities.amount().longValue()) : String.valueOf(entities.amount());
+            parts.add(amount + "元" + comparatorText(entities.comparator()));
+        }
+        if (!"UNKNOWN".equals(entities.amountBasis())) parts.add(entities.amountBasis());
+        return String.join(" ", parts);
+    }
+
+    private static List<String> alternates(String original, String domain) {
+        if (!isFormalFinanceDomain(domain)) return List.of();
+        LinkedHashSet<String> values = new LinkedHashSet<>();
+        values.add("费用报销 发票登记 报销单 财务复审");
+        if (containsAny(original, "采购", "自采", "零星", "小额")) values.add("零星采购 自采 费用报销");
+        if (containsAny(original, "借款")) values.add("个人借款 费用报销冲抵");
+        return values.stream().limit(2).toList();
+    }
+
+    private static List<String> evidenceNeeds(String original, String domain, QueryEntity entities) {
+        if (!isFormalFinanceDomain(domain)) return List.of("DIRECT_ANSWER");
+        LinkedHashSet<String> needs = new LinkedHashSet<>();
+        if (entities.amount() != null) {
+            needs.add("AMOUNT_THRESHOLD");
+            needs.add("AMOUNT_BASIS");
+            needs.add("APPLICABILITY");
+        }
+        if (containsAny(original, "怎么", "如何", "流程", "走账", "报销")) needs.add("ENTRY_STEPS");
+        needs.add("REQUIRED_MATERIALS");
+        needs.add("APPROVAL_REVIEW");
+        needs.add("EXCEPTIONS");
+        return List.copyOf(needs);
+    }
+
+    private static String normalizeChineseAmounts(String value) {
+        Matcher matcher = CHINESE_HUNDREDS.matcher(value);
+        StringBuffer output = new StringBuffer();
+        while (matcher.find()) matcher.appendReplacement(output, chineseDigit(matcher.group(1)) + "00元");
+        matcher.appendTail(output);
+        return output.toString();
+    }
+
+    private static int chineseDigit(String value) {
+        return "一二三四五六七八九".indexOf(value) + 1;
+    }
+
+    private static String comparatorText(String comparator) {
+        return switch (Optional.ofNullable(comparator).orElse("")) {
+            case "LTE" -> "以下";
+            case "GT" -> "以上";
+            case "GTE" -> "及以上";
+            default -> "";
+        };
+    }
+
+    private static boolean containsAny(String value, String... terms) {
+        String source = clean(value).toLowerCase(Locale.ROOT);
+        for (String term : terms) if (source.contains(term.toLowerCase(Locale.ROOT))) return true;
+        return false;
+    }
+
+    private static String clean(String value) {
+        return Optional.ofNullable(value).orElse("").trim();
+    }
+}

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

@@ -46,6 +46,10 @@ import org.dromara.aihr.knowledge.parse.ParsedDocument.LocatedSegment;
 import org.dromara.aihr.knowledge.parse.AihrExtractionQuality;
 import org.dromara.aihr.knowledge.parse.TikaKnowledgeDocumentParser;
 import org.dromara.aihr.knowledge.service.AihrFormalPolicyClassifier;
+import org.dromara.aihr.knowledge.service.AihrRagQueryPlanner;
+import org.dromara.aihr.knowledge.domain.AihrKnowledgeQueryDto.ChannelContribution;
+import org.dromara.aihr.knowledge.domain.AihrKnowledgeQueryDto.QueryPlan;
+import org.dromara.aihr.knowledge.domain.AihrKnowledgeQueryDto.RetrievalCandidate;
 import org.dromara.aihr.knowledge.quality.AihrKnowledgeLifecycle.UsageType;
 import org.dromara.aihr.knowledge.quality.AihrKnowledgeLifecycle.ReasonCode;
 import org.dromara.aihr.knowledge.quality.AihrKnowledgeLifecycleService;
@@ -126,6 +130,10 @@ public class AihrSopSeedService {
     private static final String LOCAL_EMBEDDING_BASE_URL = "local";
     private static final int LOCAL_EMBEDDING_DIMENSION = 1536;
     private static final int MAX_EVIDENCE_SNIPPETS = 3;
+    private static final int RETRIEVAL_CANDIDATE_K = 30;
+    private static final int RETRIEVAL_RERANK_K = 50;
+    private static final int RETRIEVAL_EVIDENCE_K = 3;
+    private static final int RETRIEVAL_DISPLAY_K = 3;
     private static final KnowledgeDocumentParser KNOWLEDGE_DOCUMENT_PARSER = new TikaKnowledgeDocumentParser();
     private static final String NO_CLEAR_SOP_EVIDENCE = "未在已发布 SOP 中找到明确依据";
     private static final String LEGACY_NO_CLEAR_SOP_EVIDENCE = "未在已入库 SOP 中找到明确依据";
@@ -227,11 +235,16 @@ public class AihrSopSeedService {
 
     public SearchResponse searchAuthorized(SearchRequest request, String requesterExtPartyId, Set<Long> allowedKnowledgeIds) {
         String queryText = request == null ? "" : Optional.ofNullable(request.queryText()).orElse("").trim();
-        return searchAuthorized(request, requesterExtPartyId, allowedKnowledgeIds, queryText);
+        return searchAuthorizedDetailed(request, requesterExtPartyId, allowedKnowledgeIds, queryText).response();
     }
 
     public SearchResponse searchAuthorized(SearchRequest request, String requesterExtPartyId,
                                            Set<Long> allowedKnowledgeIds, String retrievalQueryText) {
+        return searchAuthorizedDetailed(request, requesterExtPartyId, allowedKnowledgeIds, retrievalQueryText).response();
+    }
+
+    public SearchExecution searchAuthorizedDetailed(SearchRequest request, String requesterExtPartyId,
+                                                     Set<Long> allowedKnowledgeIds, String retrievalQueryText) {
         if (allowedKnowledgeIds == null || allowedKnowledgeIds.isEmpty()) {
             throw new ServiceException("当前请求没有可访问的知识空间", 403);
         }
@@ -240,18 +253,20 @@ public class AihrSopSeedService {
         String position = firstNonBlank(request == null ? null : request.position(), "生活顾问");
         String source = normalizeSearchSource(request == null ? null : request.source());
         String requester = isBlank(requesterExtPartyId) ? "" : requesterExtPartyId.trim();
+        QueryPlan plan = AihrRagQueryPlanner.plan(queryText);
         if (isBlank(queryText)) {
-            return withReviewId(noEvidenceResponse("", "sop"), source, requester);
+            return new SearchExecution(withReviewId(noEvidenceResponse("", "sop"), source, requester), plan, List.of());
         }
         boolean formalPolicyOnly = AihrFormalPolicyClassifier.FORMAL_POLICY_SOURCE.equals(source);
-        SearchResponse response = dbSearch(
+        SearchExecution execution = dbSearchDetailed(
             "", queryText, retrievalQuery, request == null ? null : request.limit(), allowedKnowledgeIds,
-            formalPolicyOnly);
+            formalPolicyOnly, plan);
+        SearchResponse response = execution.response();
         if (response == null || isNoEvidenceAnswer(response.answer())) {
             recordKnowledgeGap(queryText, "sop", position, source);
-            response = noEvidenceResponse(queryText, "sop");
+            if (response == null) response = noEvidenceResponse(queryText, "sop");
         }
-        return withReviewId(response, source, requester);
+        return new SearchExecution(withReviewId(response, source, requester), plan, execution.candidates());
     }
 
     /**
@@ -3418,6 +3433,10 @@ public class AihrSopSeedService {
             + String.join(",", java.util.Collections.nCopies(allowedKnowledgeIds.size(), "?")) + ")";
     }
 
+    private static String sqlPlaceholders(int count) {
+        return String.join(",", java.util.Collections.nCopies(count, "?"));
+    }
+
     private static void addKnowledgeScopeArgs(List<Object> args, Set<Long> allowedKnowledgeIds) {
         if (allowedKnowledgeIds != null) {
             args.addAll(allowedKnowledgeIds);
@@ -3431,21 +3450,24 @@ public class AihrSopSeedService {
         return """
             and exists (
                 select 1
-                  from aihr_knowledge_attach governed_attach
+                  from aihr_chunk_revision governed_chunk
-                  join aihr_knowledge_source_governance governance
+                  join aihr_data_asset governed_asset
-                    on governance.tenant_id = governed_attach.tenant_id
+                    on governed_asset.tenant_id = governed_chunk.tenant_id
-                   and governance.attachment_id = governed_attach.id
+                   and governed_asset.id = governed_chunk.asset_id
-                 where governed_attach.tenant_id = %1$s.tenant_id
+                   and coalesce(governed_asset.published_version_id, governed_asset.current_version_id) = governed_chunk.version_id
-                   and governed_attach.knowledge_id = %1$s.knowledge_id
+                  join aihr_dataset_membership governed_dataset
-                   and governed_attach.doc_id = %1$s.doc_id
+                    on governed_dataset.tenant_id = governed_chunk.tenant_id
-                   and governed_attach.status = 2
+                   and governed_dataset.version_id = governed_chunk.version_id
-                   and governance.source_doc_id = %1$s.doc_id
+                   and governed_dataset.dataset_code = 'production'
-                   and governance.authority_type = 'FORMAL_POLICY'
+                   and governed_dataset.status = 'ACTIVE'
-                   and governance.lifecycle_status = 'APPROVED'
+                 where governed_chunk.tenant_id = %1$s.tenant_id
-                   and governance.effective_date <= current_date()
+                   and governed_chunk.published_fragment_id = %1$s.id
-                   and (governance.expires_date is null or governance.expires_date >= current_date())
+                   and governed_asset.lifecycle_status = 'PUBLISHED'
-                   and char_length(trim(governance.source_version)) > 0
+                   and governed_asset.trust_level = 'HUMAN_VERIFIED'
-                   and governance.content_sha256 regexp '^[0-9a-f]{64}$'
+                   and upper(governed_asset.source_authority) in
+                       ('COMPANY_POLICY', 'FORMAL_POLICY', 'OPERATING_MANUAL', 'APPROVED_SOP')
+                   and (governed_asset.effective_from is null or governed_asset.effective_from <= current_date())
+                   and (governed_asset.effective_to is null or governed_asset.effective_to >= current_date())
             )
             """.formatted(fragmentAlias);
     }
@@ -3473,6 +3495,305 @@ public class AihrSopSeedService {
         }
     }
 
+    private SearchExecution dbSearchDetailed(String category, String queryText, String retrievalQueryText,
+                                             Integer displayLimit, Set<Long> allowedKnowledgeIds,
+                                             boolean formalPolicyOnly, QueryPlan plan) {
+        try {
+            boolean requiresFormal = formalPolicyOnly || !plan.requiredAuthorities().isEmpty();
+            List<CandidateAccumulator> candidates = retrieveCandidates(
+                category, queryText, retrievalQueryText, allowedKnowledgeIds, plan, requiresFormal);
+            if (candidates.isEmpty()) return new SearchExecution(null, plan, List.of());
+            enrichCandidateGovernance(candidates);
+            Set<Long> rejectedFeedback = downFeedbackFragmentIds(queryText);
+            for (CandidateAccumulator candidate : candidates) {
+                if (candidate.hit.fragmentId() != null && rejectedFeedback.contains(candidate.hit.fragmentId())) {
+                    candidate.reject("REJECTED_DOWNSTREAM_FEEDBACK");
+                } else if (requiresFormal && !formalSourceKind(candidate.sourceKind)) {
+                    candidate.reject("REJECTED_SOURCE_AUTHORITY");
+                } else if (!AihrRagQueryPlanner.domainMatches(plan, candidate.hit.title(), candidate.hit.content())) {
+                    candidate.reject("REJECTED_DOMAIN_MISMATCH");
+                }
+            }
+            List<CandidateAccumulator> eligible = candidates.stream()
+                .filter(candidate -> candidate.decisionReason == null)
+                .sorted(Comparator.comparingDouble(CandidateAccumulator::fusionScore).reversed())
+                .toList();
+            List<CandidateAccumulator> rerankPool = eligible.stream().limit(RETRIEVAL_RERANK_K).toList();
+            eligible.stream().skip(RETRIEVAL_RERANK_K)
+                .forEach(candidate -> candidate.reject("OUTSIDE_RERANK_K"));
+            List<CandidateAccumulator> ranked = rerankCandidates(queryText, rerankPool);
+            for (int index = 0; index < ranked.size(); index++) ranked.get(index).candidateRank = index + 1;
+
+            List<CandidateAccumulator> evidence = ranked.stream().limit(RETRIEVAL_EVIDENCE_K).toList();
+            if (plan.needsClarification()) {
+                evidence.forEach(candidate -> candidate.reject("NEEDS_CLARIFICATION_AMOUNT_BASIS"));
+                String message = "已找到正式的费用报销资料，但当前问题中的金额口径不明确：请确认“"
+                    + amountText(plan) + "”是单笔金额还是月累计额度。当前资料不能据此推导零星采购必然可以报销。";
+                return new SearchExecution(clarificationResponse(queryText, category, message), plan,
+                    candidateTraces(candidates));
+            }
+            if (requiresFormal && evidence.isEmpty()) {
+                return new SearchExecution(noEvidenceResponse(queryText, category,
+                    List.of("没有找到当前有效且经过人工审核的正式制度、操作手册或已批准 SOP。")),
+                    plan, candidateTraces(candidates));
+            }
+            if (requiresAmountEvidence(plan) && !supportsAmountRule(plan, evidence)) {
+                evidence.forEach(candidate -> candidate.reject("INSUFFICIENT_AMOUNT_EVIDENCE"));
+                String message = "已找到正式的费用报销操作资料，但当前正式资料没有证明“" + amountText(plan)
+                    + "”对应零星采购的适用额度和口径。请由财务负责人确认单笔/月累计规则后再办理。";
+                return new SearchExecution(clarificationResponse(queryText, category, message), plan,
+                    candidateTraces(candidates));
+            }
+            if (!hasRequiredFinanceEvidence(plan, evidence)) {
+                evidence.forEach(candidate -> candidate.reject("INSUFFICIENT_REQUIRED_EVIDENCE"));
+                return new SearchExecution(noEvidenceResponse(queryText, category,
+                    List.of("正式资料尚未同时覆盖报销入口、必要材料和审批/复审要求。")),
+                    plan, candidateTraces(candidates));
+            }
+            evidence.forEach(candidate -> {
+                candidate.selectedForEvidence = true;
+                candidate.decisionReason = "SELECTED_EVIDENCE";
+            });
+            ranked.stream().skip(evidence.size()).forEach(candidate -> {
+                if (candidate.decisionReason == null) candidate.decisionReason = "OUTSIDE_EVIDENCE_K";
+            });
+            SearchResponse response = responseFromHits(category, queryText,
+                evidence.stream().map(candidate -> candidate.hit).toList(), List.of(), displayLimitValue(displayLimit));
+            return new SearchExecution(response, plan, candidateTraces(candidates));
+        } catch (DataAccessException ex) {
+            return new SearchExecution(null, plan, List.of());
+        }
+    }
+
+    private List<CandidateAccumulator> retrieveCandidates(String category, String queryText,
+                                                          String retrievalQueryText,
+                                                          Set<Long> allowedKnowledgeIds, QueryPlan plan,
+                                                          boolean includePublishedCandidates) {
+        List<QueryVariant> variants = queryVariants(queryText, retrievalQueryText, plan);
+        Map<String, CandidateAccumulator> candidates = new LinkedHashMap<>();
+        for (QueryVariant variant : variants) {
+            addCandidateHits(candidates, likeHits(category, variant.text(), RETRIEVAL_CANDIDATE_K,
+                allowedKnowledgeIds, false, includePublishedCandidates), "KEYWORD", variant.code(), true);
+            addCandidateHits(candidates, fulltextHits(category, variant.text(), RETRIEVAL_CANDIDATE_K,
+                allowedKnowledgeIds, false, includePublishedCandidates), "FULLTEXT", variant.code(), true);
+        }
+        Map<String, List<KnowledgeHit>> vectorHits = vectorHits(
+            category, variants, RETRIEVAL_CANDIDATE_K, allowedKnowledgeIds, false, includePublishedCandidates);
+        for (QueryVariant variant : variants) {
+            addCandidateHits(candidates, vectorHits.getOrDefault(variant.code(), List.of()),
+                "VECTOR", variant.code(), false);
+        }
+        applyFamilyFusion(candidates.values(), true);
+        applyFamilyFusion(candidates.values(), false);
+        return candidates.values().stream()
+            .peek(CandidateAccumulator::finishFusion)
+            .sorted(Comparator.comparingDouble(CandidateAccumulator::fusionScore).reversed())
+            .toList();
+    }
+
+    private static List<QueryVariant> queryVariants(String queryText, String retrievalQueryText, QueryPlan plan) {
+        LinkedHashMap<String, String> variants = new LinkedHashMap<>();
+        addQueryVariant(variants, "ORIGINAL", queryText);
+        if (!Objects.equals(cleanQuery(queryText), cleanQuery(retrievalQueryText))) {
+            addQueryVariant(variants, "GLOSSARY", retrievalQueryText);
+        }
+        addQueryVariant(variants, "NORMALIZED", plan.normalizedQuery());
+        for (int index = 0; index < plan.alternateQueries().size(); index++) {
+            addQueryVariant(variants, "ALTERNATE_" + (index + 1), plan.alternateQueries().get(index));
+        }
+        return variants.entrySet().stream().map(entry -> new QueryVariant(entry.getKey(), entry.getValue())).toList();
+    }
+
+    private static void addQueryVariant(Map<String, String> variants, String code, String value) {
+        String query = cleanQuery(value);
+        if (query.isBlank() || variants.values().stream().anyMatch(existing -> cleanQuery(existing).equals(query))) return;
+        variants.put(code, value.trim());
+    }
+
+    private static String cleanQuery(String value) {
+        return Optional.ofNullable(value).orElse("").replaceAll("\\s+", " ").trim().toLowerCase(Locale.ROOT);
+    }
+
+    private static void addCandidateHits(Map<String, CandidateAccumulator> candidates, List<KnowledgeHit> hits,
+                                         String channel, String queryVariant, boolean sparse) {
+        for (int index = 0; index < hits.size(); index++) {
+            KnowledgeHit hit = hits.get(index);
+            String key = hit.fragmentId() == null ? hit.docId() + "#" + hit.idx() : String.valueOf(hit.fragmentId());
+            CandidateAccumulator candidate = candidates.computeIfAbsent(key, ignored -> new CandidateAccumulator(hit));
+            candidate.add(channel, index + 1, hit.score(), queryVariant, sparse);
+        }
+    }
+
+    private static void applyFamilyFusion(java.util.Collection<CandidateAccumulator> candidates, boolean sparse) {
+        List<CandidateAccumulator> ranked = candidates.stream()
+            .filter(candidate -> sparse ? candidate.sparseSignal > 0 : candidate.vectorSignal > 0)
+            .sorted(Comparator.comparingDouble((CandidateAccumulator candidate) ->
+                    sparse ? candidate.sparseSignal : candidate.vectorSignal)
+                .reversed())
+            .toList();
+        for (int index = 0; index < ranked.size(); index++) {
+            double contribution = reciprocalRank(index + 1);
+            if (sparse) ranked.get(index).sparseFusion = contribution;
+            else ranked.get(index).vectorFusion = contribution;
+        }
+    }
+
+    private static double reciprocalRank(int rank) {
+        return 1.0 / (60 + Math.max(1, rank));
+    }
+
+    private void enrichCandidateGovernance(List<CandidateAccumulator> candidates) {
+        List<Long> fragmentIds = candidates.stream().map(candidate -> candidate.hit.fragmentId())
+            .filter(Objects::nonNull).distinct().toList();
+        if (fragmentIds.isEmpty()) return;
+        List<Object> args = new ArrayList<>();
+        args.add(tenantId());
+        args.addAll(fragmentIds);
+        jdbcTemplate.query("""
+            select f.id fragment_id, asset.source_authority, asset.source_type, asset.source_name
+              from aihr_knowledge_fragment f
+              join aihr_chunk_revision chunk
+                on chunk.tenant_id = f.tenant_id and chunk.published_fragment_id = f.id
+              join aihr_data_asset asset
+                on asset.tenant_id = chunk.tenant_id and asset.id = chunk.asset_id
+               and coalesce(asset.published_version_id, asset.current_version_id) = chunk.version_id
+             where f.tenant_id = ? and f.id in (%s)
+            """.formatted(sqlPlaceholders(fragmentIds.size())), rs -> {
+                Long fragmentId = rs.getLong("fragment_id");
+                candidates.stream().filter(candidate -> Objects.equals(candidate.hit.fragmentId(), fragmentId))
+                    .findFirst().ifPresent(candidate -> {
+                        candidate.sourceAuthority = rsString(rs, "source_authority");
+                        candidate.sourceKind = AihrRagQueryPlanner.sourceKind(candidate.sourceAuthority,
+                            rsString(rs, "source_type"), rsString(rs, "source_name"));
+                    });
+            }, args.toArray());
+    }
+
+    private static String rsString(java.sql.ResultSet rs, String column) {
+        try {
+            return rs.getString(column);
+        } catch (java.sql.SQLException ex) {
+            throw new IllegalStateException(ex);
+        }
+    }
+
+    private List<CandidateAccumulator> rerankCandidates(String queryText, List<CandidateAccumulator> candidates) {
+        if (candidates.size() < 2) return candidates;
+        Optional<ChatRuntime> runtime = rerankRuntime();
+        if (runtime.isEmpty()) return candidates;
+        try {
+            ObjectNode body = objectMapper.createObjectNode();
+            body.put("model", runtime.get().modelName());
+            body.put("query", AihrSensitiveText.forModel(queryText));
+            body.put("top_n", candidates.size());
+            body.put("return_documents", false);
+            ArrayNode documents = body.putArray("documents");
+            candidates.forEach(candidate -> documents.add(
+                AihrSensitiveText.forModel(truncate(candidate.hit.content(), 2000))));
+            HttpRequest request = 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(request, 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 candidates;
+            List<CandidateAccumulator> reranked = new ArrayList<>();
+            for (JsonNode item : results) {
+                int index = item.path("index").asInt(-1);
+                if (index < 0 || index >= candidates.size()) continue;
+                CandidateAccumulator candidate = candidates.get(index);
+                candidate.rerankApplied = true;
+                candidate.rerankModel = runtime.get().modelName();
+                if (item.path("relevance_score").isNumber()) {
+                    candidate.rerankScore = item.path("relevance_score").asDouble();
+                }
+                if (!reranked.contains(candidate)) reranked.add(candidate);
+            }
+            for (CandidateAccumulator candidate : candidates) if (!reranked.contains(candidate)) reranked.add(candidate);
+            return reranked;
+        } catch (Exception ex) {
+            log.warn("candidate rerank failed, keep family RRF order（处理错误已隐藏）");
+            return candidates;
+        }
+    }
+
+    private static boolean formalSourceKind(String sourceKind) {
+        return Set.of("FORMAL_POLICY", "OPERATING_MANUAL", "APPROVED_SOP").contains(sourceKind);
+    }
+
+    private static boolean requiresAmountEvidence(QueryPlan plan) {
+        return plan != null && AihrRagQueryPlanner.isFormalFinanceDomain(plan.domain())
+            && plan.entities() != null && plan.entities().amount() != null;
+    }
+
+    private static boolean supportsAmountRule(QueryPlan plan, List<CandidateAccumulator> evidence) {
+        if (!requiresAmountEvidence(plan)) return true;
+        String amount = amountValue(plan);
+        String content = evidence.stream().map(candidate -> candidate.hit.content()).filter(Objects::nonNull)
+            .reduce("", (left, right) -> left + " " + right);
+        boolean amountFound = content.contains(amount + "元") || content.contains(amount + " 元")
+            || content.contains(chineseAmount(plan.entities().amount()));
+        boolean basisFound = "PER_TRANSACTION".equals(plan.entities().amountBasis())
+            ? containsAny(content, "单笔", "每笔", "一笔")
+            : "MONTHLY_TOTAL".equals(plan.entities().amountBasis())
+                ? containsAny(content, "月累计", "每月", "当月", "月度额度") : false;
+        return amountFound && basisFound;
+    }
+
+    private static boolean hasRequiredFinanceEvidence(QueryPlan plan, List<CandidateAccumulator> evidence) {
+        if (plan == null || !AihrRagQueryPlanner.isFormalFinanceDomain(plan.domain())) return !evidence.isEmpty();
+        String content = evidence.stream().map(candidate -> candidate.hit.content()).filter(Objects::nonNull)
+            .reduce("", (left, right) -> left + " " + right);
+        return containsAny(content, "费用报销", "报销单", "发票登记")
+            && containsAny(content, "发票", "票据", "附件")
+            && containsAny(content, "提交", "审批", "复审", "流程结束");
+    }
+
+    private static String amountValue(QueryPlan plan) {
+        double value = plan.entities().amount();
+        return value % 1 == 0 ? String.valueOf((long) value) : String.valueOf(value);
+    }
+
+    private static String amountText(QueryPlan plan) {
+        if (plan == null || plan.entities() == null || plan.entities().amount() == null) return "该金额";
+        String comparator = switch (Optional.ofNullable(plan.entities().comparator()).orElse("")) {
+            case "LTE" -> "以下";
+            case "GT" -> "以上";
+            case "GTE" -> "及以上";
+            default -> "";
+        };
+        return amountValue(plan) + "元" + comparator;
+    }
+
+    private static String chineseAmount(Double amount) {
+        if (amount == null || amount % 100 != 0 || amount < 100 || amount > 900) return "";
+        return "一二三四五六七八九".substring((int) (amount / 100) - 1, (int) (amount / 100)) + "百";
+    }
+
+    private static int displayLimitValue(Integer limit) {
+        if (limit == null) return RETRIEVAL_DISPLAY_K;
+        return Math.max(1, Math.min(limit, RETRIEVAL_DISPLAY_K));
+    }
+
+    private static SearchResponse clarificationResponse(String queryText, String category, String message) {
+        return new SearchResponse(queryText, firstNonBlank(category, "sop"), message, "正式证据不足，未生成流程",
+            List.of(), List.of(), List.of(), records(), List.of(), List.of(message), null, null);
+    }
+
+    private static List<RetrievalCandidate> candidateTraces(List<CandidateAccumulator> candidates) {
+        return candidates.stream()
+            .sorted(Comparator.comparing((CandidateAccumulator candidate) -> candidate.candidateRank,
+                    Comparator.nullsLast(Integer::compareTo))
+                .thenComparing(Comparator.comparingDouble(CandidateAccumulator::fusionScore).reversed()))
+            .map(CandidateAccumulator::toTrace)
+            .toList();
+    }
+
     private static String normalizeFeedbackRequestId(String value) {
         if (isBlank(value)) {
             return null;
@@ -3571,6 +3892,14 @@ public class AihrSopSeedService {
         return "and " + AihrKnowledgeRolloutService.servingFragmentExistsSql(fragmentAlias);
     }
 
+    private static String candidateLifecycleScopeSql(String fragmentAlias, boolean includePublishedCandidates) {
+        if (!includePublishedCandidates) {
+            return servingLifecycleScopeSql(fragmentAlias);
+        }
+        return "and (" + AihrKnowledgeRolloutService.servingFragmentExistsSql(fragmentAlias)
+            + " or " + publishedLifecycleExistsSql(fragmentAlias) + ")";
+    }
+
     private static String publishedLifecycleExistsSql(String fragmentAlias) {
         return """
             exists (
@@ -3664,6 +3993,40 @@ public class AihrSopSeedService {
         return List.of();
     }
 
+    private Map<String, List<KnowledgeHit>> vectorHits(String category, List<QueryVariant> variants, int limit,
+                                                       Set<Long> allowedKnowledgeIds,
+                                                       boolean formalPolicyOnly) {
+        return vectorHits(category, variants, limit, allowedKnowledgeIds, formalPolicyOnly, false);
+    }
+
+    private Map<String, List<KnowledgeHit>> vectorHits(String category, List<QueryVariant> variants, int limit,
+                                                       Set<Long> allowedKnowledgeIds,
+                                                       boolean formalPolicyOnly,
+                                                       boolean includePublishedCandidates) {
+        if (variants == null || variants.isEmpty()) return Map.of();
+        List<String> queries = variants.stream().map(QueryVariant::text).toList();
+        for (EmbeddingRuntime runtime : embeddingRuntimes()) {
+            try {
+                List<String> embeddings = callEmbeddings(runtime, queries);
+                if (embeddings.size() != variants.size()) continue;
+                Map<String, List<KnowledgeHit>> results = new LinkedHashMap<>();
+                for (int index = 0; index < variants.size(); index++) {
+                    List<KnowledgeHit> hits = new ArrayList<>();
+                    for (VectorMatch match : queryQdrant(
+                        category, embeddings.get(index), limit, allowedKnowledgeIds)) {
+                        hydrateQdrantHit(match, allowedKnowledgeIds, formalPolicyOnly, includePublishedCandidates)
+                            .ifPresent(hits::add);
+                    }
+                    results.put(variants.get(index).code(), List.copyOf(hits));
+                }
+                return results;
+            } catch (Exception ignored) {
+                // Vector recall is additive; try the next configured runtime before sparse-only fallback.
+            }
+        }
+        return Map.of();
+    }
+
     private List<KnowledgeHit> fulltextHits(String category, String queryText, int limit) {
         return fulltextHits(category, queryText, limit, null);
     }
@@ -3674,6 +4037,12 @@ public class AihrSopSeedService {
 
     private List<KnowledgeHit> fulltextHits(String category, String queryText, int limit,
                                             Set<Long> allowedKnowledgeIds, boolean formalPolicyOnly) {
+        return fulltextHits(category, queryText, limit, allowedKnowledgeIds, formalPolicyOnly, false);
+    }
+
+    private List<KnowledgeHit> fulltextHits(String category, String queryText, int limit,
+                                            Set<Long> allowedKnowledgeIds, boolean formalPolicyOnly,
+                                            boolean includePublishedCandidates) {
         String sql = """
             select f.id as fragment_id,
                    coalesce(a.name, i.name) as title,
@@ -3695,7 +4064,8 @@ public class AihrSopSeedService {
              order by score desc, f.idx asc
              limit ?
             """.formatted(knowledgeScopeSql("f", allowedKnowledgeIds),
-                formalSourceScopeSql("f", formalPolicyOnly), servingLifecycleScopeSql("f"));
+                formalSourceScopeSql("f", formalPolicyOnly),
+                candidateLifecycleScopeSql("f", includePublishedCandidates));
         List<Object> args = new ArrayList<>();
         args.add(queryText);
         args.add(tenantId());
@@ -3727,6 +4097,12 @@ public class AihrSopSeedService {
 
     private List<KnowledgeHit> likeHits(String category, String queryText, int limit,
                                         Set<Long> allowedKnowledgeIds, boolean formalPolicyOnly) {
+        return likeHits(category, queryText, limit, allowedKnowledgeIds, formalPolicyOnly, false);
+    }
+
+    private List<KnowledgeHit> likeHits(String category, String queryText, int limit,
+                                        Set<Long> allowedKnowledgeIds, boolean formalPolicyOnly,
+                                        boolean includePublishedCandidates) {
         List<String> terms = searchTerms(queryText);
         if (terms.isEmpty()) {
             return List.of();
@@ -3758,7 +4134,8 @@ public class AihrSopSeedService {
              order by score desc, f.idx asc
              limit ?
             """.formatted(String.join(" + ", scoreParts), knowledgeScopeSql("f", allowedKnowledgeIds),
-                formalSourceScopeSql("f", formalPolicyOnly), servingLifecycleScopeSql("f"),
+                formalSourceScopeSql("f", formalPolicyOnly),
+                candidateLifecycleScopeSql("f", includePublishedCandidates),
                 String.join(" or ", whereParts));
         List<Object> args = new ArrayList<>();
         terms.stream().map(AihrSopSeedService::likePattern).forEach(args::add);
@@ -3791,6 +4168,12 @@ public class AihrSopSeedService {
 
     private Optional<KnowledgeHit> hydrateQdrantHit(VectorMatch match, Set<Long> allowedKnowledgeIds,
                                                     boolean formalPolicyOnly) {
+        return hydrateQdrantHit(match, allowedKnowledgeIds, formalPolicyOnly, false);
+    }
+
+    private Optional<KnowledgeHit> hydrateQdrantHit(VectorMatch match, Set<Long> allowedKnowledgeIds,
+                                                    boolean formalPolicyOnly,
+                                                    boolean includePublishedCandidates) {
         String sql = """
             select f.id as fragment_id,
                    coalesce(a.name, i.name) as title,
@@ -3811,7 +4194,8 @@ public class AihrSopSeedService {
                and f.idx = ?
              limit 1
             """.formatted(knowledgeScopeSql("f", allowedKnowledgeIds),
-                formalSourceScopeSql("f", formalPolicyOnly), servingLifecycleScopeSql("f"));
+                formalSourceScopeSql("f", formalPolicyOnly),
+                candidateLifecycleScopeSql("f", includePublishedCandidates));
         List<Object> args = new ArrayList<>();
         args.add(tenantId());
         addKnowledgeScopeArgs(args, allowedKnowledgeIds);
@@ -3862,15 +4246,21 @@ public class AihrSopSeedService {
     }
 
     private SearchResponse responseFromHits(String category, String queryText, List<KnowledgeHit> hits, List<String> extraCautions) {
+        return responseFromHits(category, queryText, hits, extraCautions, RETRIEVAL_DISPLAY_K);
+    }
+
+    private SearchResponse responseFromHits(String category, String queryText, List<KnowledgeHit> hits,
+                                            List<String> extraCautions, int displayLimit) {
         List<KnowledgeHit> evidenceHits = evidenceHits(queryText, hits);
         KnowledgeHit first = evidenceHits.get(0);
         String responseCategory = isBlank(category) ? first.category() : category;
-        List<SnippetResponse> snippets = evidenceHits.stream()
+        List<KnowledgeHit> displayHits = displayHits(evidenceHits, displayLimit);
+        List<SnippetResponse> snippets = displayHits.stream()
             .map(hit -> new SnippetResponse("第 " + hit.idx() + " 段：" + hit.title(),
                 displaySnippetText(hit.content()), hit.fragmentId(), hit.score(), hit.retrievalChannel()))
             .toList();
         Map<String, DocResponse> docs = new LinkedHashMap<>();
-        for (KnowledgeHit hit : evidenceHits) {
+        for (KnowledgeHit hit : displayHits) {
             docs.putIfAbsent(hit.title(), new DocResponse(hit.title(), hit.description(), "命中 " + scoreText(hit.score()), "pill-success"));
         }
         SearchDigest digest = searchDigest(queryText, evidenceHits);
@@ -3914,6 +4304,11 @@ public class AihrSopSeedService {
         return hits.stream().limit(MAX_EVIDENCE_SNIPPETS).toList();
     }
 
+    static List<KnowledgeHit> displayHits(List<KnowledgeHit> evidenceHits, Integer limit) {
+        if (evidenceHits == null || evidenceHits.isEmpty()) return List.of();
+        return evidenceHits.stream().limit(displayLimitValue(limit)).toList();
+    }
+
     private static boolean queryMentionsTitle(String queryText, String title) {
         String query = evidenceText(queryText);
         String docTitle = evidenceText(title);
@@ -4201,26 +4596,7 @@ public class AihrSopSeedService {
     }
 
     private static List<String> searchTerms(String queryText) {
-        String cleaned = Optional.ofNullable(queryText).orElse("")
+        return AihrRagQueryPlanner.sparseTerms(queryText);
-            .toLowerCase()
-            .replaceAll("[^\\p{IsHan}\\p{Alnum}]+", " ")
-            .trim();
-        if (cleaned.isEmpty()) {
-            return List.of();
-        }
-        LinkedHashSet<String> terms = new LinkedHashSet<>();
-        for (String token : cleaned.split("\\s+")) {
-            if (token.length() >= 2 && token.length() <= 16) {
-                terms.add(token);
-            }
-            for (int i = 0; i + 2 <= token.length() && terms.size() < 12; i++) {
-                terms.add(token.substring(i, i + 2));
-            }
-            if (terms.size() >= 12) {
-                break;
-            }
-        }
-        return terms.stream().limit(12).toList();
     }
 
     private static String likePattern(String term) {
@@ -5474,6 +5850,63 @@ public class AihrSopSeedService {
     static record FeedbackFilterResult(List<KnowledgeHit> hits, List<String> cautions) {
     }
 
+    public record SearchExecution(SearchResponse response, QueryPlan queryPlan,
+                                  List<RetrievalCandidate> candidates) {
+        public SearchExecution {
+            candidates = candidates == null ? List.of() : List.copyOf(candidates);
+        }
+    }
+
+    private record QueryVariant(String code, String text) {}
+
+    private static final class CandidateAccumulator {
+        private final KnowledgeHit hit;
+        private final List<ChannelContribution> contributions = new ArrayList<>();
+        private double sparseSignal;
+        private double vectorSignal;
+        private double sparseFusion;
+        private double vectorFusion;
+        private double fusionScore;
+        private String sourceAuthority = "UNKNOWN";
+        private String sourceKind = "REFERENCE_MATERIAL";
+        private Integer candidateRank;
+        private Double rerankScore;
+        private String rerankModel;
+        private boolean rerankApplied;
+        private boolean selectedForEvidence;
+        private String decisionReason;
+
+        private CandidateAccumulator(KnowledgeHit hit) {
+            this.hit = hit;
+        }
+
+        private void add(String channel, int channelRank, double rawScore, String queryVariant, boolean sparse) {
+            double signal = reciprocalRank(channelRank);
+            contributions.add(new ChannelContribution(channel, channelRank, rawScore, signal, queryVariant));
+            if (sparse) sparseSignal = Math.max(sparseSignal, signal);
+            else vectorSignal = Math.max(vectorSignal, signal);
+        }
+
+        private void finishFusion() {
+            fusionScore = sparseFusion + vectorFusion;
+        }
+
+        private double fusionScore() {
+            return fusionScore;
+        }
+
+        private void reject(String reason) {
+            selectedForEvidence = false;
+            if (decisionReason == null || "OUTSIDE_EVIDENCE_K".equals(decisionReason)) decisionReason = reason;
+        }
+
+        private RetrievalCandidate toTrace() {
+            return new RetrievalCandidate(hit.fragmentId(), hit.title(), sourceAuthority, sourceKind,
+                candidateRank, fusionScore, rerankScore, rerankModel, rerankApplied, selectedForEvidence,
+                firstNonBlank(decisionReason, "CANDIDATE"), List.copyOf(contributions));
+        }
+    }
+
     private record KnowledgeConfig(long knowledgeId, int blockSize, int overlap) {
     }
 

backend/ruoyi-modules/ruoyi-aihr/src/test/java/org/dromara/aihr/knowledge/service/AihrRagQueryPlannerTest.java

@@ -0,0 +1,61 @@
+package org.dromara.aihr.knowledge.service;
+
+import org.junit.jupiter.api.Tag;
+import org.junit.jupiter.api.Test;
+
+import java.util.List;
+
+import static org.assertj.core.api.Assertions.assertThat;
+
+@Tag("dev")
+class AihrRagQueryPlannerTest {
+
+    @Test
+    void removesNumericNgramNoiseButPreservesTypedAmountAndBusinessIntent() {
+        var query = "我有个报销500块钱，告诉我怎么走流程。";
+
+        var plan = AihrRagQueryPlanner.plan(query);
+        var terms = AihrRagQueryPlanner.sparseTerms(query);
+
+        assertThat(plan.originalQuery()).isEqualTo(query);
+        assertThat(plan.domain()).isEqualTo(AihrRagQueryPlanner.FINANCE_POLICY);
+        assertThat(plan.entities().amount()).isEqualTo(500D);
+        assertThat(plan.entities().currency()).isEqualTo("CNY");
+        assertThat(terms).contains("500元", "报销", "流程");
+        assertThat(terms).doesNotContain("50", "00", "销5", "0块", "我有", "告诉");
+    }
+
+    @Test
+    void plansAmbiguousPurchaseThresholdAsFormalFinanceClarification() {
+        List<String> queries = List.of(
+            "500元以下的零星采购怎么报销？",
+            "五百块的小额自采怎么走账？"
+        );
+
+        for (String query : queries) {
+            var plan = AihrRagQueryPlanner.plan(query);
+            assertThat(plan.domain()).isEqualTo(AihrRagQueryPlanner.FINANCE_REIMBURSEMENT);
+            assertThat(plan.entities().amount()).isEqualTo(500D);
+            assertThat(plan.entities().amountBasis()).isEqualTo("UNKNOWN");
+            assertThat(plan.needsClarification()).isTrue();
+            assertThat(plan.requiredAuthorities())
+                .containsExactly("FORMAL_POLICY", "OPERATING_MANUAL", "APPROVED_SOP");
+            assertThat(plan.evidenceNeeds())
+                .contains("AMOUNT_THRESHOLD", "AMOUNT_BASIS", "ENTRY_STEPS", "REQUIRED_MATERIALS",
+                    "APPROVAL_REVIEW", "EXCEPTIONS");
+        }
+    }
+
+    @Test
+    void preservesExplicitAmountBasisAndRejectsExperienceAsFormalAuthority() {
+        var plan = AihrRagQueryPlanner.plan("月累计500元以内的零星采购怎么报销？");
+
+        assertThat(plan.entities().amountBasis()).isEqualTo("MONTHLY_TOTAL");
+        assertThat(plan.needsClarification()).isFalse();
+        assertThat(AihrRagQueryPlanner.sourceKind("EMPLOYEE_SUBMISSION", "INTERVIEW", "员工访谈"))
+            .isEqualTo("EXPERIENCE");
+        assertThat(AihrRagQueryPlanner.authorityAllowed(plan, "EXPERIENCE")).isFalse();
+        assertThat(AihrRagQueryPlanner.sourceKind("COMPANY_POLICY", "PDF", "费用报销操作手册.pdf"))
+            .isEqualTo("OPERATING_MANUAL");
+    }
+}

backend/ruoyi-modules/ruoyi-aihr/src/test/java/org/dromara/aihr/service/AihrSopSeedServiceTest.java

@@ -152,6 +152,20 @@ public class AihrSopSeedServiceTest {
         assertEquals(3L, evidence.get(2).fragmentId());
     }
 
+    @Test
+    @Tag("dev")
+    public void displayLimitDoesNotReduceEvidencePool() {
+        List<AihrSopSeedService.KnowledgeHit> evidence = AihrSopSeedService.evidenceHits(
+            "费用报销的发票怎么登记？",
+            List.of(hit(1L, "费用报销操作手册"), hit(2L, "费用报销操作手册"), hit(3L, "费用报销操作手册")));
+
+        List<AihrSopSeedService.KnowledgeHit> displayed = AihrSopSeedService.displayHits(evidence, 1);
+
+        assertEquals(3, evidence.size());
+        assertEquals(1, displayed.size());
+        assertEquals(1L, displayed.get(0).fragmentId());
+    }
+
     @Test
     @Tag("dev")
     public void allDownVotedHitsAreSuppressed() {
@@ -667,12 +681,14 @@ public class AihrSopSeedServiceTest {
         assertTrue(code.contains("searchAuthorized(SearchRequest request, String requesterExtPartyId, Set<Long> allowedKnowledgeIds)"));
         assertTrue(code.contains(".knowledge_id in ("));
         assertTrue(code.contains("qdrantMatchAny(\"knowledge_id\", knowledgeIds)"));
-        assertTrue(code.contains("aihr_knowledge_source_governance"));
+        assertTrue(code.contains("governed_dataset.dataset_code = 'production'"));
-        assertTrue(code.contains("governance.source_doc_id"));
+        assertTrue(code.contains("governed_dataset.status = 'ACTIVE'"));
-        assertTrue(code.contains("governance.authority_type = 'FORMAL_POLICY'"));
+        assertTrue(code.contains("governed_asset.lifecycle_status = 'PUBLISHED'"));
-        assertTrue(code.contains("governance.lifecycle_status = 'APPROVED'"));
+        assertTrue(code.contains("governed_asset.trust_level = 'HUMAN_VERIFIED'"));
-        assertTrue(code.contains("governance.effective_date <= current_date()"));
+        assertTrue(code.contains("upper(governed_asset.source_authority) in"));
-        assertTrue(code.contains("governance.content_sha256 regexp '^[0-9a-f]{64}$'"));
+        assertTrue(code.contains("'COMPANY_POLICY', 'FORMAL_POLICY', 'OPERATING_MANUAL', 'APPROVED_SOP'"));
+        assertTrue(code.contains("governed_asset.effective_from is null"));
+        assertTrue(code.contains("governed_asset.effective_to is null"));
 
         AihrSopSeedService service = new AihrSopSeedService(
             new ObjectMapper(), mock(JdbcTemplate.class), null, "", null, null, mock(AihrModelSeedService.class));
@@ -680,6 +696,27 @@ public class AihrSopSeedServiceTest {
             () -> service.searchAuthorized(new AihrSopDto.SearchRequest("问题", "sop", null, null, 5), "", Set.of()));
     }
 
+    @Test
+    @Tag("dev")
+    public void formalCandidateRecallIncludesPublishedGovernedFragmentsWithoutChangingDefaultRollout() throws Exception {
+        Method candidateScope = AihrSopSeedService.class.getDeclaredMethod(
+            "candidateLifecycleScopeSql", String.class, boolean.class);
+        Method servingScope = AihrSopSeedService.class.getDeclaredMethod("servingLifecycleScopeSql", String.class);
+        Method publishedScope = AihrSopSeedService.class.getDeclaredMethod("publishedLifecycleExistsSql", String.class);
+        candidateScope.setAccessible(true);
+        servingScope.setAccessible(true);
+        publishedScope.setAccessible(true);
+
+        String serving = (String) servingScope.invoke(null, "f");
+        String published = (String) publishedScope.invoke(null, "f");
+        String defaultCandidates = (String) candidateScope.invoke(null, "f", false);
+        String formalCandidates = (String) candidateScope.invoke(null, "f", true);
+
+        assertEquals(serving, defaultCandidates);
+        assertTrue(formalCandidates.contains(serving.substring("and ".length())));
+        assertTrue(formalCandidates.contains(published));
+    }
+
     @Test
     @Tag("dev")
     public void legacySearchControllerDelegatesToAuthorizedQueryService() throws Exception {

backend/script/sql/update/aihr_20260815_rag_candidate_trace_mysql8.sql

@@ -0,0 +1,29 @@
+-- Generation-safe RAG candidate observability. Additive and idempotent on MySQL 8.
+
+CREATE TABLE IF NOT EXISTS `aihr_query_candidate_trace` (
+  `id` bigint NOT NULL AUTO_INCREMENT,
+  `tenant_id` varchar(20) NOT NULL,
+  `request_id` varchar(64) NOT NULL,
+  `fragment_id` bigint NOT NULL,
+  `candidate_rank` int DEFAULT NULL COMMENT 'Rank after rerank, null when rejected before rerank',
+  `channel` varchar(20) NOT NULL COMMENT 'KEYWORD/FULLTEXT/VECTOR',
+  `channel_rank` int NOT NULL,
+  `raw_score` decimal(18,10) DEFAULT NULL,
+  `fusion_score` decimal(18,10) DEFAULT NULL,
+  `rerank_score` decimal(18,10) DEFAULT NULL,
+  `rerank_model` varchar(100) DEFAULT NULL,
+  `rerank_applied` tinyint(1) NOT NULL DEFAULT 0,
+  `source_authority` varchar(30) DEFAULT NULL,
+  `source_kind` varchar(30) DEFAULT NULL,
+  `query_variant` varchar(30) NOT NULL COMMENT 'ORIGINAL/NORMALIZED/GLOSSARY/ALTERNATE_n',
+  `selected_for_evidence` tinyint(1) NOT NULL DEFAULT 0,
+  `decision_reason` varchar(64) NOT NULL,
+  `create_time` datetime NOT NULL DEFAULT CURRENT_TIMESTAMP,
+  PRIMARY KEY (`id`),
+  UNIQUE KEY `uk_aihr_query_candidate_channel`
+    (`tenant_id`, `request_id`, `fragment_id`, `channel`, `query_variant`),
+  KEY `idx_aihr_query_candidate_request` (`tenant_id`, `request_id`, `candidate_rank`),
+  KEY `idx_aihr_query_candidate_reason` (`tenant_id`, `decision_reason`, `create_time`),
+  KEY `idx_aihr_query_candidate_fragment` (`tenant_id`, `fragment_id`, `create_time`)
+) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_0900_ai_ci
+  COMMENT='Pre-generation retrieval candidates, channel contributions, rerank and gate decisions';

docs/BRD_PRODUCTION_MIGRATION_RUNBOOK.md

@@ -166,6 +166,7 @@ mysql --default-character-set=utf8mb4 "$DB_NAME" < backend/script/sql/update/aih
 mysql --default-character-set=utf8mb4 "$DB_NAME" < backend/script/sql/update/aihr_20260812_pipeline_run_sampling_mysql8.sql
 mysql --default-character-set=utf8mb4 "$DB_NAME" < backend/script/sql/update/aihr_20260813_golden_calibration_mysql8.sql
 mysql --default-character-set=utf8mb4 "$DB_NAME" < backend/script/sql/update/aihr_20260814_knowledge_privacy_derivative_mysql8.sql
+mysql --default-character-set=utf8mb4 "$DB_NAME" < backend/script/sql/update/aihr_20260815_rag_candidate_trace_mysql8.sql
 mysql --default-character-set=utf8mb4 "$DB_NAME" < backend/script/sql/update/aihr_20260815_knowledge_rollout_compat_mysql8.sql
 mysql --default-character-set=utf8mb4 "$DB_NAME" < backend/script/sql/update/aihr_20260816_knowledge_version_pointers_mysql8.sql
 mysql --default-character-set=utf8mb4 "$DB_NAME" < backend/script/sql/update/aihr_20260817_knowledge_review_assistance_mysql8.sql

mobile-uni/src/pages/user/assistant/index.vue

@@ -288,7 +288,7 @@ const addOutlineSection = () => outline.value?.sections.push({ title: '', bullet
 const addOutlineBullet = (sectionIndex: number) => outline.value?.sections[sectionIndex]?.bullets.push('');
 
 const citationLabel = (citation: PersonalCitation) => {
-  if (citation.domain === 'ENTERPRISE') return '企业 SOP';
+  if (citation.domain === 'ENTERPRISE') return '企业知识';
   return citation.sourceType === 'URL' ? '外部网页' : '我的资料';
 };
 

mobile-uni/src/pages/user/assistant/sessions.vue

@@ -39,7 +39,7 @@ const confirmDelete = (id: string) => uni.showModal({ title: '删除会话', con
 const remove = async (id: string) => { loading.value = true; try { await deletePersonalSession(id); detail.value = undefined; await loadList(); } catch (error) { message.value = error instanceof Error ? error.message : '删除会话失败'; } finally { loading.value = false; } };
 const scopeText = (scope: string) => scope.includes('ENTERPRISE') ? '我的资料 + 企业知识' : '我的资料';
 const citationLabel = (citation: PersonalCitation) => citation.domain === 'ENTERPRISE'
-  ? '企业 SOP'
+  ? '企业知识'
   : citation.sourceType === 'URL' ? '外部网页' : '我的资料';
 const openCitation = (citation: PersonalCitation) => {
   if (citation.itemId) uni.navigateTo({ url: personalItemRoute(citation.itemId) });

mobile-uni/src/pages/user/sop/index.vue

@@ -1326,7 +1326,13 @@ const citationDomainLabel = (snippet: KnowledgeSearchResponse['snippets'][number
   if (snippet.domain === 'PROJECT_SERVICE') return '项目记录';
   if (snippet.domain === 'PERSONAL') return '我的资料';
   if (snippet.sourceType === 'DATA_TOOL') return '数据工具';
-  return '企业 SOP';
+  return ({
+    FORMAL_POLICY: '正式制度',
+    OPERATING_MANUAL: '操作手册',
+    APPROVED_SOP: '已批准 SOP',
+    EXPERIENCE: '经验材料',
+    CASE: '案例'
+  }[snippet.sourceKind || ''] || '企业知识');
 };
 
 const serviceMemoryStatusLabel = (status?: string | null) => ({

mobile-uni/src/services/agent.ts

@@ -86,6 +86,8 @@ export const normalizeAgentResponse = (
     fragmentId: citation.fragmentId,
     spaceCode: citation.spaceCode,
     sourceType: citation.sourceType,
+    sourceAuthority: citation.sourceAuthority,
+    sourceKind: citation.sourceKind,
     domain: citation.domain,
     status: citation.status,
     occurredAt: citation.occurredAt,

mobile-uni/src/services/knowledge.ts

@@ -276,6 +276,8 @@ const normalizeUnifiedResponse = (response: UnifiedKnowledgeResponse): Knowledge
     fragmentId: item.fragmentId,
     spaceCode: item.spaceCode,
     sourceType: item.sourceType,
+    sourceAuthority: item.sourceAuthority,
+    sourceKind: item.sourceKind,
     domain: item.domain,
     status: item.status,
     occurredAt: item.occurredAt,

mobile-uni/src/types/api.ts

@@ -115,6 +115,8 @@ export interface KnowledgeSnippet {
   fragmentId?: number | null;
   spaceCode?: string | null;
   sourceType?: 'DOCUMENT' | 'DATA_TOOL' | string;
+  sourceAuthority?: string | null;
+  sourceKind?: 'FORMAL_POLICY' | 'OPERATING_MANUAL' | 'APPROVED_SOP' | 'EXPERIENCE' | 'CASE' | string | null;
   domain?: 'PERSONAL' | 'PROJECT_SERVICE' | 'ENTERPRISE' | string | null;
   status?: string | null;
   occurredAt?: string | null;
@@ -171,6 +173,8 @@ export interface KnowledgeMediaSelectedFile {
 export interface KnowledgeCitation {
   spaceCode?: string | null;
   sourceType: 'DOCUMENT' | 'DATA_TOOL' | string;
+  sourceAuthority?: string | null;
+  sourceKind?: 'FORMAL_POLICY' | 'OPERATING_MANUAL' | 'APPROVED_SOP' | 'EXPERIENCE' | 'CASE' | string | null;
   docId?: string | null;
   title: string;
   snippet: string;

scripts/evaluate-knowledge-quality.mjs

@@ -11,6 +11,29 @@ function hitSource(hit) {
   return typeof hit === 'object' && hit !== null ? String(hit.sourceName || hit.title || '') : '';
 }
 
+export function normalizeSourceName(value) {
+  return String(value || '')
+    .replace(/^第\s*\d+\s*段[：:]\s*/u, '')
+    .replace(/\.(?:pdf|docx?|pptx?|xlsx?|txt|md)$/iu, '')
+    .replace(/\s+/gu, '')
+    .toLowerCase();
+}
+
+function sourceMatches(actual, expected) {
+  const left = normalizeSourceName(actual);
+  const right = normalizeSourceName(expected);
+  return Boolean(left && right && (left.includes(right) || right.includes(left)));
+}
+
+function requiredSourceMatches(hit, source, authorities, kinds) {
+  if (!sourceMatches(hitSource(hit), source)) return false;
+  if (typeof hit !== 'object' || hit === null) return authorities.length === 0 && kinds.length === 0;
+  const authority = String(hit.sourceAuthority || '').toUpperCase();
+  const kind = String(hit.sourceKind || '').toUpperCase();
+  return (authorities.length === 0 || authorities.includes(authority))
+    && (kinds.length === 0 || kinds.includes(kind));
+}
+
 export function reciprocalRank(retrieved, relevant) {
   const relevantSet = new Set(relevant.map(String));
   const rank = retrieved.findIndex((hit) => relevantSet.has(hitId(hit)));
@@ -27,38 +50,68 @@ export function recallAtK(retrieved, relevant) {
 export function ndcgAtK(retrieved, relevant) {
   const relevantSet = new Set(relevant.map(String));
   if (relevantSet.size === 0) return 0;
-  const dcg = retrieved.reduce((sum, hit, index) =>
+  const credited = new Set();
-    sum + (relevantSet.has(hitId(hit)) ? 1 / Math.log2(index + 2) : 0), 0);
+  const dcg = retrieved.reduce((sum, hit, index) => {
+    const id = hitId(hit);
+    if (!relevantSet.has(id) || credited.has(id)) return sum;
+    credited.add(id);
+    return sum + 1 / Math.log2(index + 2);
+  }, 0);
   const idealLength = Math.min(relevantSet.size, retrieved.length);
   const ideal = Array.from({ length: idealLength }, (_, index) => 1 / Math.log2(index + 2))
     .reduce((sum, value) => sum + value, 0);
   return ideal === 0 ? 0 : dcg / ideal;
 }
 
-export function evaluateCase(retrieved, testCase, k) {
+export function evaluateCase(retrieved, testCase, k, actualNoEvidence = retrieved.length === 0) {
   const ranked = retrieved.slice(0, k);
   const relevant = testCase.relevantFragmentIds || [];
   const forbiddenIds = new Set((testCase.forbiddenFragmentIds || []).map(String));
-  const forbiddenSources = new Set((testCase.forbiddenSourceNames || []).map((value) => value.toLowerCase()));
+  const forbiddenSources = testCase.forbiddenSourceNames || [];
-  const requiredSources = new Set((testCase.requiredSourceNames || []).map((value) => value.toLowerCase()));
+  const requiredSources = testCase.requiredSourceNames || [];
-  const returnedSources = new Set(ranked.map(hitSource).filter(Boolean).map((value) => value.toLowerCase()));
+  const requiredAuthorities = (testCase.requiredSourceAuthorities || []).map((value) => String(value).toUpperCase());
-  const forbiddenLeak = ranked.some((hit) => forbiddenIds.has(hitId(hit))
+  const requiredKinds = (testCase.requiredSourceKinds || []).map((value) => String(value).toUpperCase());
-    || forbiddenSources.has(hitSource(hit).toLowerCase()));
+  const visibleEvidence = ranked.filter((hit) => typeof hit !== 'object' || hit === null
-  const missingRequiredSource = requiredSources.size > 0
+    || hit.selectedForEvidence === undefined || hit.selectedForEvidence === true);
-    && ![...requiredSources].some((source) => returnedSources.has(source));
+  const forbiddenLeak = visibleEvidence.some((hit) => forbiddenIds.has(hitId(hit))
+    || forbiddenSources.some((source) => sourceMatches(hitSource(hit), source)));
+  const missingRequiredSource = requiredSources.length > 0
+    && !requiredSources.some((source) => ranked.some((hit) =>
+      requiredSourceMatches(hit, source, requiredAuthorities, requiredKinds)));
+  const sourceRanking = requiredSources.length > 0 && relevant.length === 0
+    ? ranked.map((hit, index) => requiredSources.some((source) =>
+      requiredSourceMatches(hit, source, requiredAuthorities, requiredKinds))
+        ? normalizeSourceName(hitSource(hit)) : `__nonmatching_source_${index}`) : null;
+  const normalizedRelevantSources = requiredSources.map(normalizeSourceName);
+  const metricHits = sourceRanking || ranked;
+  const metricRelevant = sourceRanking ? normalizedRelevantSources : relevant;
   const expectedNoEvidence = Boolean(testCase.expectedNoEvidence);
+  const retrievalEvaluated = metricRelevant.length > 0;
   return {
     id: testCase.id,
     caseType: testCase.caseType || (expectedNoEvidence ? 'no-answer' : 'answerable'),
-    answerable: !expectedNoEvidence,
+    answerable: retrievalEvaluated,
-    recallAtK: expectedNoEvidence ? null : recallAtK(ranked, relevant),
+    retrievalEvaluated,
-    reciprocalRank: expectedNoEvidence ? null : reciprocalRank(ranked, relevant),
+    expectedNoEvidence,
-    ndcgAtK: expectedNoEvidence ? null : ndcgAtK(ranked, relevant),
+    recallAtK: retrievalEvaluated ? recallAtK(metricHits, metricRelevant) : null,
+    recallAt5: retrievalEvaluated ? recallAtK(metricHits.slice(0, 5), metricRelevant) : null,
+    recallAt10: retrievalEvaluated ? recallAtK(metricHits.slice(0, 10), metricRelevant) : null,
+    recallAt20: retrievalEvaluated ? recallAtK(metricHits.slice(0, 20), metricRelevant) : null,
+    reciprocalRank: retrievalEvaluated ? reciprocalRank(metricHits, metricRelevant) : null,
+    ndcgAtK: retrievalEvaluated ? ndcgAtK(metricHits, metricRelevant) : null,
     forbiddenLeak,
     wrongSource: forbiddenLeak || missingRequiredSource,
     returned: ranked.length,
-    noEvidence: ranked.length === 0,
+    noEvidence: actualNoEvidence,
-    noAnswerFalsePositive: expectedNoEvidence && ranked.length > 0
+    noAnswerFalsePositive: expectedNoEvidence && !actualNoEvidence,
+    topCandidates: ranked.slice(0, 5).map((hit, index) => ({
+      rank: index + 1,
+      fragmentId: typeof hit === 'object' && hit !== null ? hit.fragmentId : hit,
+      sourceName: hitSource(hit),
+      sourceAuthority: typeof hit === 'object' && hit !== null ? hit.sourceAuthority : null,
+      sourceKind: typeof hit === 'object' && hit !== null ? hit.sourceKind : null,
+      selectedForEvidence: typeof hit === 'object' && hit !== null ? hit.selectedForEvidence : undefined
+    }))
   };
 }
 
@@ -71,8 +124,8 @@ function average(rows, key) {
 }
 
 export function summarize(results) {
-  const answerable = results.filter((row) => row.answerable);
+  const answerable = results.filter((row) => row.retrievalEvaluated ?? row.answerable);
-  const noAnswer = results.filter((row) => !row.answerable);
+  const noAnswer = results.filter((row) => row.expectedNoEvidence ?? !row.answerable);
   const byCaseType = Object.fromEntries([...new Set(results.map((row) => row.caseType))].sort().map((caseType) => {
     const rows = results.filter((row) => row.caseType === caseType);
     return [caseType, { cases: rows.length, forbiddenLeakageRate: rate(rows, (row) => row.forbiddenLeak) }];
@@ -82,6 +135,9 @@ export function summarize(results) {
     answerableCases: answerable.length,
     noAnswerCases: noAnswer.length,
     recallAtK: average(answerable, 'recallAtK'),
+    recallAt5: average(answerable, 'recallAt5'),
+    recallAt10: average(answerable, 'recallAt10'),
+    recallAt20: average(answerable, 'recallAt20'),
     mrr: average(answerable, 'reciprocalRank'),
     ndcgAtK: average(answerable, 'ndcgAtK'),
     forbiddenLeakageRate: rate(results, (row) => row.forbiddenLeak),
@@ -155,8 +211,10 @@ export function validateDataset(dataset) {
   for (const item of dataset.cases) {
     if (!item.id || ids.has(item.id) || !item.query) throw new Error(`invalid or duplicate case: ${item.id || '<missing>'}`);
     ids.add(item.id);
-    if (!item.expectedNoEvidence && (!Array.isArray(item.relevantFragmentIds) || item.relevantFragmentIds.length === 0)) {
+    if (!item.expectedNoEvidence
-      throw new Error(`answerable case must define relevantFragmentIds: ${item.id}`);
+        && (!Array.isArray(item.relevantFragmentIds) || item.relevantFragmentIds.length === 0)
+        && (!Array.isArray(item.requiredSourceNames) || item.requiredSourceNames.length === 0)) {
+      throw new Error(`answerable case must define relevantFragmentIds or requiredSourceNames: ${item.id}`);
     }
     if (item.expectedNoEvidence && (item.relevantFragmentIds || []).length > 0) {
       throw new Error(`no-answer case cannot define relevantFragmentIds: ${item.id}`);
@@ -168,23 +226,42 @@ async function requestSearch(options, testCase, fetchImpl) {
   const headers = { 'content-type': 'application/json' };
   if (options.token) headers.Authorization = options.token.startsWith('Bearer ') ? options.token : `Bearer ${options.token}`;
   if (options.clientid) headers.clientid = options.clientid;
-  const response = await fetchImpl(`${options.baseUrl.replace(/\/$/, '')}/api/knowledge/search`, {
+  const response = await fetchImpl(`${options.baseUrl.replace(/\/$/, '')}/api/knowledge/query`, {
     method: 'POST', headers,
     body: JSON.stringify({
       queryText: testCase.query,
       category: testCase.category || undefined,
       position: testCase.position || undefined,
-      source: 'knowledge_search',
+      source: testCase.source || 'mobile_uni_agent',
-      limit: options.k
+      // Response display remains capped at 10; retrievalCandidates is the decoupled pre-generation pool.
+      limit: Math.min(options.k, 10)
     })
   });
   const body = await response.json();
   if (!response.ok || body.code !== 200) {
-    throw new Error(`case ${testCase.id} search failed with HTTP ${response.status} code ${body.code}`);
+    throw new Error(`case ${testCase.id} search failed with HTTP ${response.status} code ${body.code}: ${body.msg || ''}`);
   }
-  const snippets = body.data?.snippets || [];
+  const candidates = body.data?.retrievalCandidates;
-  return snippets.filter((snippet) => snippet.fragmentId !== null && snippet.fragmentId !== undefined)
+  const fallback = body.data?.citations || body.data?.legacy?.snippets || [];
-    .map((snippet) => ({ fragmentId: snippet.fragmentId, sourceName: snippet.title || '' }));
+  const rows = Array.isArray(candidates) && candidates.length > 0
+    ? [...candidates].sort((left, right) => {
+        const leftRank = Number.isInteger(left.candidateRank) ? left.candidateRank : Number.MAX_SAFE_INTEGER;
+        const rightRank = Number.isInteger(right.candidateRank) ? right.candidateRank : Number.MAX_SAFE_INTEGER;
+        return leftRank - rightRank;
+      })
+    : fallback;
+  return {
+    noEvidence: Boolean(body.data?.noEvidence),
+    retrieved: rows.filter((item) => item.fragmentId !== null && item.fragmentId !== undefined)
+      .map((item) => ({
+        fragmentId: item.fragmentId,
+        sourceName: item.title || '',
+        sourceAuthority: item.sourceAuthority || null,
+        sourceKind: item.sourceKind || null,
+        candidateRank: item.candidateRank,
+        selectedForEvidence: item.selectedForEvidence
+      }))
+  };
 }
 
 export async function run(options, dependencies = {}) {
@@ -197,8 +274,8 @@ export async function run(options, dependencies = {}) {
   const fetchImpl = dependencies.fetchImpl || globalThis.fetch;
   const results = [];
   for (const testCase of dataset.cases) {
-    const retrieved = await requestSearch(options, testCase, fetchImpl);
+    const result = await requestSearch(options, testCase, fetchImpl);
-    results.push(evaluateCase(retrieved, testCase, options.k));
+    results.push(evaluateCase(result.retrieved, testCase, options.k, result.noEvidence));
   }
   const summary = summarize(results);
   const failures = thresholdFailures(summary, options);

scripts/evaluate-knowledge-quality.test.mjs

@@ -1,13 +1,15 @@
 import test from 'node:test';
 import assert from 'node:assert/strict';
 import {
-  evaluateCase, ndcgAtK, parseArgs, recallAtK, reciprocalRank, summarize, thresholdFailures, validateDataset
+  evaluateCase, ndcgAtK, normalizeSourceName, parseArgs, recallAtK, reciprocalRank, summarize,
+  thresholdFailures, validateDataset
 } from './evaluate-knowledge-quality.mjs';
 
 test('ranking metrics use fragment ids and preserve top-k order', () => {
   assert.equal(recallAtK(['a', 'b', 'c'], ['b', 'd']), 0.5);
   assert.equal(reciprocalRank(['a', 'b'], ['b']), 0.5);
   assert.ok(ndcgAtK(['b', 'a'], ['b', 'c']) > 0.61);
+  assert.equal(ndcgAtK(['b', 'b', 'b'], ['b']), 1);
 });
 
 test('forbidden fragments and sources are surfaced as leakage', () => {
@@ -25,7 +27,7 @@ test('forbidden fragments and sources are surfaced as leakage', () => {
 test('no-answer cases measure false positives separately from answerable recall', () => {
   const answerable = evaluateCase([{ fragmentId: 1, sourceName: 'SOP' }], {
     id: 'answerable', relevantFragmentIds: [1], requiredSourceNames: ['SOP']
-  }, 5);
+  }, 5, false);
   const noAnswer = evaluateCase([{ fragmentId: 9, sourceName: 'Unrelated' }], {
     id: 'no-answer', expectedNoEvidence: true, relevantFragmentIds: [], caseType: 'no-answer'
   }, 5);
@@ -35,6 +37,45 @@ test('no-answer cases measure false positives separately from answerable recall'
   assert.equal(summary.noAnswerPrecision, 0);
 });
 
+test('candidate recall is independent from the final no-evidence gate and display prefixes', () => {
+  const result = evaluateCase([
+    {
+      fragmentId: 122126,
+      sourceName: '第 3 段：银城物业费用报销发票管理操作手册.pdf',
+      sourceAuthority: 'COMPANY_POLICY',
+      sourceKind: 'OPERATING_MANUAL',
+      selectedForEvidence: false
+    }
+  ], {
+    id: 'finance',
+    requiredSourceNames: ['银城物业费用报销发票管理操作手册'],
+    requiredSourceAuthorities: ['COMPANY_POLICY'],
+    requiredSourceKinds: ['OPERATING_MANUAL'],
+    expectedNoEvidence: false
+  }, 20, true);
+  assert.equal(result.recallAt5, 1);
+  assert.equal(result.noEvidence, true);
+  assert.equal(normalizeSourceName('第 12 段：制度.PDF'), '制度');
+});
+
+test('same-name ungoverned copies do not satisfy formal source recall', () => {
+  const result = evaluateCase([
+    {
+      fragmentId: 102169,
+      sourceName: '银城物业费用报销发票管理操作手册.pdf',
+      sourceAuthority: 'UNKNOWN',
+      sourceKind: 'REFERENCE_MATERIAL'
+    }
+  ], {
+    id: 'finance-authority',
+    requiredSourceNames: ['银城物业费用报销发票管理操作手册'],
+    requiredSourceAuthorities: ['COMPANY_POLICY'],
+    requiredSourceKinds: ['OPERATING_MANUAL']
+  }, 20, true);
+  assert.equal(result.recallAt20, 0);
+  assert.equal(result.wrongSource, true);
+});
+
 test('threshold failures can fail a release gate', () => {
   const failures = thresholdFailures({
     recallAtK: 0.7, mrr: 0.5, ndcgAtK: 0.6, forbiddenLeakageRate: 0.1,
@@ -49,6 +90,7 @@ test('threshold failures can fail a release gate', () => {
 test('dataset contract distinguishes answerable and no-answer cases', () => {
   assert.doesNotThrow(() => validateDataset({ schemaVersion: 1, cases: [
     { id: 'a', query: 'q', relevantFragmentIds: [1] },
+    { id: 'source', query: 'q-source', requiredSourceNames: ['正式手册'] },
     { id: 'b', query: 'q2', relevantFragmentIds: [], expectedNoEvidence: true }
   ] }));
   assert.throws(() => validateDataset({ schemaVersion: 1, cases: [

scripts/release-preflight.sh

@@ -275,6 +275,7 @@ FROM (
   UNION ALL SELECT 'aihr_review_batch'
   UNION ALL SELECT 'aihr_review_batch_item'
   UNION ALL SELECT 'aihr_knowledge_generation'
+  UNION ALL SELECT 'aihr_query_candidate_trace'
   UNION ALL SELECT 'aihr_generation_version'
   UNION ALL SELECT 'aihr_version_diff'
   UNION ALL SELECT 'aihr_version_rollback'
@@ -948,7 +949,8 @@ REMOTE
 
   [[ "$media_reprocess_contract" = "source_attach_id|tenant_id,source_attach_id,id" ]] \
     || fail "remote media-reprocess schema is invalid: got ${media_reprocess_contract:-missing}"
-  echo "remote_schema=82/82 $remote_ssh:$remote_db"
+  echo "remote_schema=83/83 $remote_ssh:$remote_db"
+  echo "remote_rag_candidate_trace=1/1 $remote_ssh:$remote_db"
   echo "remote_media_reprocess_contract=1/1 $remote_ssh:$remote_db"
   echo "remote_work_report_idempotency=3/3 $remote_ssh:$remote_db"
   echo "remote_learning_task_question_bank=6/6 $remote_ssh:$remote_db"

scripts/reset-dev-db.sh

@@ -36,6 +36,7 @@ docker exec -i wygj-mysql mysql -uroot -proot --default-character-set=utf8mb4 ry
 docker exec -i wygj-mysql mysql -uroot -proot --default-character-set=utf8mb4 ry-vue < "$ROOT_DIR/backend/script/sql/update/aihr_20260812_pipeline_run_sampling_mysql8.sql"
 docker exec -i wygj-mysql mysql -uroot -proot --default-character-set=utf8mb4 ry-vue < "$ROOT_DIR/backend/script/sql/update/aihr_20260813_golden_calibration_mysql8.sql"
 docker exec -i wygj-mysql mysql -uroot -proot --default-character-set=utf8mb4 ry-vue < "$ROOT_DIR/backend/script/sql/update/aihr_20260814_knowledge_privacy_derivative_mysql8.sql"
+docker exec -i wygj-mysql mysql -uroot -proot --default-character-set=utf8mb4 ry-vue < "$ROOT_DIR/backend/script/sql/update/aihr_20260815_rag_candidate_trace_mysql8.sql"
 docker exec -i wygj-mysql mysql -uroot -proot --default-character-set=utf8mb4 ry-vue < "$ROOT_DIR/backend/script/sql/update/aihr_20260815_knowledge_rollout_compat_mysql8.sql"
 docker exec -i wygj-mysql mysql -uroot -proot --default-character-set=utf8mb4 ry-vue < "$ROOT_DIR/backend/script/sql/update/aihr_20260816_knowledge_version_pointers_mysql8.sql"
 docker exec -i wygj-mysql mysql -uroot -proot --default-character-set=utf8mb4 ry-vue < "$ROOT_DIR/backend/script/sql/update/aihr_20260817_knowledge_review_assistance_mysql8.sql"

scripts/tests/aihr-schema-migrations.test.sh

@@ -655,4 +655,13 @@ test "$agent_run_unique" = '0|run_id'
 agent_user_index="$(mysql "$DB" -N -B -e "SELECT GROUP_CONCAT(column_name ORDER BY seq_in_index) FROM information_schema.statistics WHERE table_schema='$DB' AND table_name='aihr_agent_run' AND index_name='idx_aihr_agent_run_user';")"
 test "$agent_user_index" = 'tenant_id,user_id,create_time'
 
+rag_candidate_trace_migration="$ROOT_DIR/backend/script/sql/update/aihr_20260815_rag_candidate_trace_mysql8.sql"
+test -f "$rag_candidate_trace_migration"
+mysql "$DB" < "$rag_candidate_trace_migration"
+mysql "$DB" < "$rag_candidate_trace_migration"
+rag_candidate_trace_columns="$(mysql "$DB" -N -B -e "SELECT GROUP_CONCAT(column_name ORDER BY ordinal_position) FROM information_schema.columns WHERE table_schema='$DB' AND table_name='aihr_query_candidate_trace';")"
+test "$rag_candidate_trace_columns" = 'id,tenant_id,request_id,fragment_id,candidate_rank,channel,channel_rank,raw_score,fusion_score,rerank_score,rerank_model,rerank_applied,source_authority,source_kind,query_variant,selected_for_evidence,decision_reason,create_time'
+rag_candidate_trace_unique="$(mysql "$DB" -N -B -e "SELECT CONCAT(MIN(non_unique),'|',GROUP_CONCAT(column_name ORDER BY seq_in_index)) FROM information_schema.statistics WHERE table_schema='$DB' AND table_name='aihr_query_candidate_trace' AND index_name='uk_aihr_query_candidate_channel';")"
+test "$rag_candidate_trace_unique" = '0|tenant_id,request_id,fragment_id,channel,query_variant'
+
 echo "PASS: AIHR legacy schema migrations are idempotent, including Agent run audit, practice growth curriculum snapshots, knowledge-space categories, practice evidence, knowledge feedback, candidate interview review, position/SOP qualification contracts, assistant capture, work-report idempotency, direct feedback, broadcast publish/withdraw audit with targeting, delivery controls, topic evidence and file access audit, v1 snapshots, and trusted broadcast question context"

scripts/verify-finance-rag-local.mjs

@@ -0,0 +1,77 @@
+#!/usr/bin/env node
+
+import { run } from './evaluate-knowledge-quality.mjs';
+import { execFileSync } from 'node:child_process';
+import { fileURLToPath } from 'node:url';
+
+const baseUrl = process.env.AIHR_BASE_URL || 'https://wygj-api.localhost';
+const clientid = '428a8310cd442757ae699df5d894f051';
+const dataset = new URL('../tests/fixtures/data_quality/retrieval-finance-regression.json', import.meta.url);
+
+function localUniqueActivePhone() {
+  const sql = `
+    select min(s.person_phone)
+      from aihr_org_snapshot s
+     where s.tenant_id = '000000'
+       and lower(coalesce(s.employment_status, 'active')) in ('active', '在职')
+       and s.person_phone regexp '^[0-9]{11}$'
+       and (select count(distinct p.ext_party_id) from aihr_org_snapshot p
+             where p.tenant_id = s.tenant_id and p.person_phone = s.person_phone
+               and lower(coalesce(p.employment_status, 'active')) in ('active', '在职')) = 1
+       and (select count(distinct e.person_phone) from aihr_org_snapshot e
+             where e.tenant_id = s.tenant_id and e.ext_party_id = s.ext_party_id
+               and lower(coalesce(e.employment_status, 'active')) in ('active', '在职')) = 1`;
+  return execFileSync('docker', [
+    'exec', 'wygj-mysql', 'mysql', '-uroot', '-proot', '--default-character-set=utf8mb4',
+    'ry-vue', '-Nse', sql
+  ], { encoding: 'utf8' }).trim();
+}
+
+const phone = process.env.AIHR_VERIFY_PHONE || localUniqueActivePhone();
+if (!phone) throw new Error('no bidirectionally unique active local employee is available');
+
+execFileSync('docker', [
+  'exec', 'wygj-redis', 'redis-cli', '-a', 'ruoyi123', 'del',
+  `{global:rate_limit:/resource/sms/code:${phone}}:value`,
+  `global:rate_limit:/resource/sms/code:${phone}`,
+  `{global:rate_limit:/resource/sms/code:${phone}}:permits`
+], { stdio: 'ignore' });
+
+if (new URL(baseUrl).hostname.endsWith('.localhost')) {
+  process.env.NODE_TLS_REJECT_UNAUTHORIZED = '0';
+}
+
+async function api(endpoint, init = {}) {
+  const response = await fetch(`${baseUrl.replace(/\/$/, '')}${endpoint}`, {
+    ...init,
+    headers: { clientid, ...(init.headers || {}) },
+    signal: AbortSignal.timeout(30_000)
+  });
+  const payload = await response.json();
+  if (!response.ok || Number(payload.code) !== 200) {
+    throw new Error(`${endpoint} failed: HTTP ${response.status}, code ${payload.code}, ${payload.msg || ''}`);
+  }
+  return payload.data;
+}
+
+await api(`/resource/sms/code?phonenumber=${encodeURIComponent(phone)}`);
+const login = await api('/auth/mobile/sms-login', {
+  method: 'POST',
+  headers: { 'content-type': 'application/json' },
+  body: JSON.stringify({
+    phonenumber: phone,
+    smsCode: process.env.AIHR_VERIFY_SMS_CODE || '123456',
+    tenantId: '000000'
+  })
+});
+if (!login?.access_token) throw new Error('mobile login returned no access token');
+
+const result = await run({
+  dataset: fileURLToPath(dataset),
+  baseUrl,
+  token: login.access_token,
+  clientid,
+  k: 20
+});
+console.log(JSON.stringify(result, null, 2));
+if (result.thresholdPassed === false) process.exitCode = 2;

tests/fixtures/data_quality/retrieval-finance-regression.json

@@ -0,0 +1,64 @@
+{
+  "schemaVersion": 1,
+  "datasetCode": "finance-reimbursement-rag-regression-v1",
+  "cases": [
+    {
+      "id": "finance-500-colloquial",
+      "query": "我有个报销500块钱，告诉我怎么走流程。",
+      "caseType": "formal-finance-ambiguous-basis",
+      "requiredSourceNames": ["银城物业费用报销发票管理操作手册"],
+      "requiredSourceAuthorities": ["COMPANY_POLICY"],
+      "requiredSourceKinds": ["OPERATING_MANUAL"],
+      "forbiddenSourceNames": ["200块私了", "500块维修"],
+      "expectedNoEvidence": true
+    },
+    {
+      "id": "finance-500-small-purchase",
+      "query": "500元以下的零星采购怎么报销？",
+      "caseType": "formal-finance-ambiguous-basis",
+      "requiredSourceNames": ["银城物业费用报销发票管理操作手册"],
+      "requiredSourceAuthorities": ["COMPANY_POLICY"],
+      "requiredSourceKinds": ["OPERATING_MANUAL"],
+      "forbiddenSourceNames": ["200块私了", "500块维修"],
+      "expectedNoEvidence": true
+    },
+    {
+      "id": "finance-500-chinese-amount",
+      "query": "五百块的小额自采怎么走账？",
+      "caseType": "formal-finance-ambiguous-basis",
+      "requiredSourceNames": ["银城物业费用报销发票管理操作手册"],
+      "requiredSourceAuthorities": ["COMPANY_POLICY"],
+      "requiredSourceKinds": ["OPERATING_MANUAL"],
+      "forbiddenSourceNames": ["200块私了", "500块维修"],
+      "expectedNoEvidence": true
+    },
+    {
+      "id": "finance-private-settlement-no-answer",
+      "query": "200块维修费用能不能直接私了后报销？",
+      "caseType": "no-answer",
+      "expectedNoEvidence": true,
+      "forbiddenSourceNames": ["200块私了", "500块维修"]
+    },
+    {
+      "id": "finance-over-limit-unsupported",
+      "query": "零星采购超过500元怎么报销？",
+      "caseType": "unsupported-threshold",
+      "expectedNoEvidence": true,
+      "forbiddenSourceNames": ["200块私了", "500块维修"]
+    },
+    {
+      "id": "finance-monthly-basis-conflict",
+      "query": "月累计500元以内的自采都能报销吗？",
+      "caseType": "amount-basis-conflict",
+      "expectedNoEvidence": true,
+      "forbiddenSourceNames": ["员工访谈", "200块私了", "500块维修"]
+    },
+    {
+      "id": "finance-per-transaction-basis-conflict",
+      "query": "单笔500元以内的自采都能报销吗？",
+      "caseType": "amount-basis-conflict",
+      "expectedNoEvidence": true,
+      "forbiddenSourceNames": ["员工访谈", "200块私了", "500块维修"]
+    }
+  ]
+}