feat: add rag evidence postprocess blocks

openspec/changes/archive/2026-07-04-rag-evidence-postprocess/.openspec.yaml

@@ -0,0 +1,2 @@
+schema: spec-driven
+created: 2026-07-04

openspec/changes/archive/2026-07-04-rag-evidence-postprocess/design.md

@@ -0,0 +1,53 @@
+## Context
+
+The previous change converted L0 into a hint provider and made L0/L1 cooperate. The next step is to stop treating the tool output as an unstructured primary/supplement pair. The Agent can keep receiving compatible fields, but retrieval internals and traces should have structured evidence blocks.
+
+This change is a bridge toward later DocumentPostProcessor-style behavior. It should be small enough to archive independently and should not introduce Spring AI dependencies.
+
+## Goals / Non-Goals
+
+**Goals:**
+
+- Add `EvidenceBlock` DTOs to `LookupResult`.
+- Create evidence blocks from L0 matches and L1 results.
+- Deduplicate evidence by stable source key.
+- Capture source, title, breadcrumb, score, retrieval layer, hit reasons, and content preview.
+- Persist evidence blocks and postprocess counts in `tool_invocation.retrieval_details`.
+
+**Non-Goals:**
+
+- Do not implement neighbor chunk expansion yet.
+- Do not replace primary/supplement output.
+- Do not add cross-encoder or LLM rerank.
+- Do not migrate to Spring AI DocumentPostProcessor yet.
+
+## Decisions
+
+### Decision: Add evidence blocks while preserving existing result fields
+
+`LookupResult` will gain `List<EvidenceBlock> evidenceBlocks`. Existing `primary`, `supplement`, `found`, `relevanceLevel`, and `completenessHint` remain compatible.
+
+Rationale: this lets the Agent continue using the current shape while tests and traces begin validating the new evidence model.
+
+### Decision: Keep postprocess rule-based
+
+The evidence builder will use deterministic rules:
+
+- L0 entries become `L0` evidence.
+- L1 candidates become `L1` evidence.
+- Same source key is deduplicated.
+- Hit reasons are collected from L0 hints, L1 rank, category filters, and fallback state.
+
+Rationale: this is explainable, cheap, and suitable before introducing framework postprocessors.
+
+### Decision: Persist compact evidence summaries
+
+`ToolInvocationRecorder` will store compact evidence block metadata, not full content, inside `retrieval_details`.
+
+Rationale: `tool_invocation` should remain useful for trace review without duplicating large chunks.
+
+## Risks / Trade-offs
+
+- Evidence source keys may be imperfect before full metadata normalization -> fall back to file path, metadata, title, then rank.
+- Agent prompts may ignore `evidenceBlocks` initially -> keep primary/supplement compatibility.
+- Adding content previews increases tool output size -> cap evidence content length.

openspec/changes/archive/2026-07-04-rag-evidence-postprocess/proposal.md

@@ -0,0 +1,28 @@
+## Why
+
+The current `lookup_knowledge` result is still shaped as one L0 primary result plus one L1 supplement. That makes retrieval evidence hard to inspect, hard to deduplicate, and hard to reuse later by verifier/evaluator code. The RAG refactor needs a structured evidence layer before Spring AI retriever migration.
+
+## What Changes
+
+- Add structured evidence blocks to `LookupResult`.
+- Build evidence blocks from L0 hint matches and L1 candidates.
+- Deduplicate evidence by source identity where possible.
+- Add hit reasons such as L0 matched keywords, L1 semantic rank, category filter, and fallback.
+- Persist evidence block summaries and postprocess counts in `tool_invocation.retrieval_details`.
+- Keep existing `primary` and `supplement` fields for compatibility.
+
+## Capabilities
+
+### New Capabilities
+
+- None.
+
+### Modified Capabilities
+
+- `rag-knowledge-retrieval`: Add evidence block post-processing requirements for explicit RAG knowledge retrieval.
+
+## Impact
+
+- Affects `LookupResult`, `LookupKnowledgeTool`, and `ToolInvocationRecorder`.
+- Updates tool tests and trace recorder tests.
+- Does not change Milvus schema, document upload, chunking, or Spring AI integration.

openspec/changes/archive/2026-07-04-rag-evidence-postprocess/specs/rag-knowledge-retrieval/spec.md

@@ -0,0 +1,35 @@
+## ADDED Requirements
+
+### Requirement: Knowledge retrieval SHALL return structured evidence blocks
+The `lookup_knowledge` retrieval flow SHALL expose retrieved evidence as structured evidence blocks in addition to the existing compatibility fields.
+
+#### Scenario: Evidence block contains source metadata
+- **WHEN** a `lookup_knowledge` call returns evidence
+- **THEN** each evidence block SHALL include source, title when available, breadcrumb when available, retrieval layer, content, and hit reasons
+
+#### Scenario: Compatibility fields remain available
+- **WHEN** evidence blocks are returned
+- **THEN** the existing `primary` and `supplement` result fields SHALL remain available when their source evidence exists
+
+### Requirement: Knowledge retrieval SHALL deduplicate evidence blocks
+The retrieval flow SHALL remove duplicate evidence blocks before returning them to the Agent.
+
+#### Scenario: Duplicate source deduplication
+- **WHEN** L0 and L1 produce evidence with the same source identity
+- **THEN** the retrieval flow SHALL keep a single evidence block for that source
+- **AND** the evidence block SHALL preserve hit reasons from both retrieval paths when available
+
+#### Scenario: Postprocess count tracking
+- **WHEN** evidence post-processing completes
+- **THEN** the tool invocation details SHALL record candidate count and final evidence block count
+
+### Requirement: Knowledge retrieval SHALL persist evidence block summaries
+The system SHALL persist compact evidence block summaries in `tool_invocation.retrieval_details`.
+
+#### Scenario: Evidence summaries are persisted
+- **WHEN** a `lookup_knowledge` call records a tool invocation
+- **THEN** `retrieval_details` SHALL include evidence block summaries containing source, title, retrieval layer, score when available, and hit reasons
+
+#### Scenario: Full content is not duplicated into retrieval details
+- **WHEN** evidence block summaries are persisted
+- **THEN** full evidence content SHALL be omitted or truncated so the trace record remains compact

openspec/changes/archive/2026-07-04-rag-evidence-postprocess/tasks.md

@@ -0,0 +1,26 @@
+## 1. Evidence Model
+
+- [x] 1.1 Add an `EvidenceBlock` DTO.
+- [x] 1.2 Add evidence block list and postprocess count fields to `LookupResult`.
+
+## 2. Evidence Postprocess
+
+- [x] 2.1 Build evidence blocks from L0 matches and L1 candidates in `LookupKnowledgeTool`.
+- [x] 2.2 Deduplicate evidence by stable source key.
+- [x] 2.3 Preserve existing primary/supplement compatibility behavior.
+
+## 3. Trace Recording
+
+- [x] 3.1 Extend `ToolInvocationRecorder.LookupKnowledgeRecord` with evidence block summaries and postprocess counts.
+- [x] 3.2 Persist evidence block summaries in `retrieval_details`.
+
+## 4. Tests
+
+- [x] 4.1 Add or update tests for evidence block creation and deduplication.
+- [x] 4.2 Add or update tests for persisted evidence block summaries.
+- [x] 4.3 Run targeted tests and the RAG retrieval baseline evaluator.
+
+## 5. Validation
+
+- [x] 5.1 Run OpenSpec validation for the change.
+- [x] 5.2 Review git diff to confirm only expected code/spec/test files changed.

openspec/specs/rag-knowledge-retrieval/spec.md

@@ -48,3 +48,37 @@ The system SHALL persist L0 hint details in `tool_invocation.retrieval_details`
 #### Scenario: Retrieval layer reflects cooperating retrieval
 - **WHEN** both L0 hint data and L1 candidates participate in a lookup
 - **THEN** the recorded retrieval layer SHALL be `L0+L1`
+
+### Requirement: Knowledge retrieval SHALL return structured evidence blocks
+The `lookup_knowledge` retrieval flow SHALL expose retrieved evidence as structured evidence blocks in addition to the existing compatibility fields.
+
+#### Scenario: Evidence block contains source metadata
+- **WHEN** a `lookup_knowledge` call returns evidence
+- **THEN** each evidence block SHALL include source, title when available, breadcrumb when available, retrieval layer, content, and hit reasons
+
+#### Scenario: Compatibility fields remain available
+- **WHEN** evidence blocks are returned
+- **THEN** the existing `primary` and `supplement` result fields SHALL remain available when their source evidence exists
+
+### Requirement: Knowledge retrieval SHALL deduplicate evidence blocks
+The retrieval flow SHALL remove duplicate evidence blocks before returning them to the Agent.
+
+#### Scenario: Duplicate source deduplication
+- **WHEN** L0 and L1 produce evidence with the same source identity
+- **THEN** the retrieval flow SHALL keep a single evidence block for that source
+- **AND** the evidence block SHALL preserve hit reasons from both retrieval paths when available
+
+#### Scenario: Postprocess count tracking
+- **WHEN** evidence post-processing completes
+- **THEN** the tool invocation details SHALL record candidate count and final evidence block count
+
+### Requirement: Knowledge retrieval SHALL persist evidence block summaries
+The system SHALL persist compact evidence block summaries in `tool_invocation.retrieval_details`.
+
+#### Scenario: Evidence summaries are persisted
+- **WHEN** a `lookup_knowledge` call records a tool invocation
+- **THEN** `retrieval_details` SHALL include evidence block summaries containing source, title, retrieval layer, score when available, and hit reasons
+
+#### Scenario: Full content is not duplicated into retrieval details
+- **WHEN** evidence block summaries are persisted
+- **THEN** full evidence content SHALL be omitted or truncated so the trace record remains compact

src/main/java/com/superbiz/agent/dto/EvidenceBlock.java

@@ -0,0 +1,28 @@
+package com.superbiz.agent.dto;
+
+import lombok.Builder;
+import lombok.Data;
+
+import java.util.List;
+
+/**
+ * Structured evidence returned by knowledge retrieval.
+ */
+@Data
+@Builder
+public class EvidenceBlock {
+
+    private String source;
+
+    private String title;
+
+    private String breadcrumb;
+
+    private String retrievalLayer;
+
+    private String content;
+
+    private Double score;
+
+    private List<String> hitReasons;
+}

src/main/java/com/superbiz/agent/dto/LookupResult.java

@@ -27,6 +27,21 @@ public class LookupResult {
      */
     private SupplementResult supplement;
 
+    /**
+     * Structured evidence blocks after retrieval post-processing.
+     */
+    private List<EvidenceBlock> evidenceBlocks;
+
+    /**
+     * Candidate count before evidence deduplication.
+     */
+    private Integer evidenceCandidateCount;
+
+    /**
+     * Evidence block count after post-processing.
+     */
+    private Integer evidenceBlockCount;
+
     /**
      * 归一化质量等级：PRECISE / HIGHLY_RELEVANT / REFERENCE
      */

src/main/java/com/superbiz/agent/service/ToolInvocationRecorder.java

@@ -3,6 +3,7 @@ package com.superbiz.agent.service;
 import com.fasterxml.jackson.core.JsonProcessingException;
 import com.fasterxml.jackson.databind.ObjectMapper;
 import com.superbiz.agent.domain.entity.ToolInvocation;
+import com.superbiz.agent.dto.EvidenceBlock;
 import com.superbiz.agent.dto.LookupResult;
 import com.superbiz.agent.repository.ToolInvocationRepository;
 import com.superbiz.agent.dto.KnowledgeEntry;
@@ -128,6 +129,15 @@ public class ToolInvocationRecorder {
         if (record.l1Scores() != null && !record.l1Scores().isEmpty()) {
             details.put("l1_scores", record.l1Scores());
         }
+        if (record.evidenceCandidateCount() != null) {
+            details.put("evidence_candidate_count", record.evidenceCandidateCount());
+        }
+        if (record.evidenceBlockCount() != null) {
+            details.put("evidence_block_count", record.evidenceBlockCount());
+        }
+        if (record.evidenceBlocks() != null && !record.evidenceBlocks().isEmpty()) {
+            details.put("evidence_blocks", record.evidenceBlocks());
+        }
         if (record.relevanceLevel() != null) {
             details.put("relevance_level", record.relevanceLevel());
         }
@@ -209,7 +219,10 @@ public class ToolInvocationRecorder {
             List<String> l0Entities,
             Double l1TopScore,
             Double l1TopSimilarity,
-            List<Double> l1Scores
+            List<Double> l1Scores,
+            Integer evidenceCandidateCount,
+            Integer evidenceBlockCount,
+            List<Map<String, Object>> evidenceBlocks
     ) {
         public static LookupKnowledgeRecord from(String query,
                                                  KnowledgeIndexService.L0Hint l0Hint,
@@ -290,7 +303,33 @@ public class ToolInvocationRecorder {
                     .l1TopScore(hasL1 ? (double) l1Results.get(0).getScore() : null)
                     .l1TopSimilarity(hasL1 ? l1TopSimilarity : null)
                     .l1Scores(l1Scores)
+                    .evidenceCandidateCount(result != null ? result.getEvidenceCandidateCount() : null)
+                    .evidenceBlockCount(result != null ? result.getEvidenceBlockCount() : null)
+                    .evidenceBlocks(result != null ? summarizeEvidenceBlocks(result.getEvidenceBlocks()) : List.of())
                     .build();
         }
+
+        private static List<Map<String, Object>> summarizeEvidenceBlocks(List<EvidenceBlock> blocks) {
+            if (blocks == null || blocks.isEmpty()) {
+                return List.of();
+            }
+            List<Map<String, Object>> summaries = new ArrayList<>();
+            for (int i = 0; i < Math.min(5, blocks.size()); i++) {
+                EvidenceBlock block = blocks.get(i);
+                Map<String, Object> summary = new LinkedHashMap<>();
+                summary.put("source", block.getSource());
+                summary.put("title", block.getTitle());
+                summary.put("breadcrumb", block.getBreadcrumb());
+                summary.put("retrieval_layer", block.getRetrievalLayer());
+                summary.put("score", block.getScore());
+                summary.put("hit_reasons", block.getHitReasons());
+                String content = block.getContent();
+                if (content != null) {
+                    summary.put("content_preview", content.length() <= 180 ? content : content.substring(0, 180) + "...");
+                }
+                summaries.add(summary);
+            }
+            return summaries;
+        }
     }
 }

src/main/java/com/superbiz/agent/tool/LookupKnowledgeTool.java

@@ -13,8 +13,11 @@ import org.springframework.beans.factory.annotation.Autowired;
 import org.springframework.beans.factory.annotation.Value;
 import org.springframework.stereotype.Component;
 
+import java.util.ArrayList;
+import java.util.LinkedHashMap;
 import java.util.List;
 import java.util.Locale;
+import java.util.Map;
 import java.util.stream.Collectors;
 
 /**
@@ -397,12 +400,154 @@ public class LookupKnowledgeTool {
         }
         builder.supplement(supplement);
 
+        EvidencePostprocessResult evidence = buildEvidenceBlocks(l0Matches, l1Results);
+        builder.evidenceBlocks(evidence.blocks());
+        builder.evidenceCandidateCount(evidence.candidateCount());
+        builder.evidenceBlockCount(evidence.blocks().size());
+
         boolean found = (primary != null) || (supplement != null);
         builder.found(found);
 
         return builder.build();
     }
 
+    private EvidencePostprocessResult buildEvidenceBlocks(
+            List<KnowledgeEntry> l0Matches,
+            List<VectorSearchService.SearchResult> l1Results) {
+        Map<String, EvidenceBlock> deduped = new LinkedHashMap<>();
+        int candidateCount = 0;
+
+        if (l0Matches != null) {
+            for (int i = 0; i < l0Matches.size(); i++) {
+                KnowledgeEntry entry = l0Matches.get(i);
+                candidateCount++;
+                EvidenceBlock block = EvidenceBlock.builder()
+                        .source(entry.getFilePath())
+                        .title(entry.getTitle())
+                        .breadcrumb(null)
+                        .retrievalLayer("L0")
+                        .content(buildMetadataOnlySummary(entry))
+                        .score(null)
+                        .hitReasons(buildL0HitReasons(entry, i + 1))
+                        .build();
+                mergeEvidence(deduped, sourceKey(block, "l0-" + i), block);
+            }
+        }
+
+        if (l1Results != null) {
+            for (int i = 0; i < l1Results.size(); i++) {
+                VectorSearchService.SearchResult result = l1Results.get(i);
+                candidateCount++;
+                Map<String, String> metadata = parseMetadata(result.getMetadata());
+                String source = firstNonBlank(
+                        metadata.get("_source"),
+                        metadata.get("docId"),
+                        result.getMetadata(),
+                        result.getId()
+                );
+                EvidenceBlock block = EvidenceBlock.builder()
+                        .source(source)
+                        .title(metadata.get("title"))
+                        .breadcrumb(metadata.get("breadcrumb"))
+                        .retrievalLayer("L1")
+                        .content(truncate(result.getContent(), 800))
+                        .score((double) result.getScore())
+                        .hitReasons(List.of("semantic_rank:" + (i + 1)))
+                        .build();
+                mergeEvidence(deduped, sourceKey(block, "l1-" + i), block);
+            }
+        }
+
+        return new EvidencePostprocessResult(candidateCount, new ArrayList<>(deduped.values()));
+    }
+
+    private void mergeEvidence(Map<String, EvidenceBlock> deduped, String key, EvidenceBlock incoming) {
+        EvidenceBlock existing = deduped.get(key);
+        if (existing == null) {
+            deduped.put(key, incoming);
+            return;
+        }
+
+        List<String> mergedReasons = new ArrayList<>();
+        if (existing.getHitReasons() != null) {
+            mergedReasons.addAll(existing.getHitReasons());
+        }
+        if (incoming.getHitReasons() != null) {
+            for (String reason : incoming.getHitReasons()) {
+                if (!mergedReasons.contains(reason)) {
+                    mergedReasons.add(reason);
+                }
+            }
+        }
+
+        String mergedLayer = existing.getRetrievalLayer();
+        if (incoming.getRetrievalLayer() != null && !incoming.getRetrievalLayer().equals(mergedLayer)) {
+            mergedLayer = "L0+L1";
+        }
+
+        existing.setRetrievalLayer(mergedLayer);
+        existing.setHitReasons(mergedReasons);
+        if (existing.getScore() == null && incoming.getScore() != null) {
+            existing.setScore(incoming.getScore());
+        }
+        if ((existing.getBreadcrumb() == null || existing.getBreadcrumb().isBlank())
+                && incoming.getBreadcrumb() != null) {
+            existing.setBreadcrumb(incoming.getBreadcrumb());
+        }
+    }
+
+    private List<String> buildL0HitReasons(KnowledgeEntry entry, int rank) {
+        List<String> reasons = new ArrayList<>();
+        reasons.add("l0_rank:" + rank);
+        if (entry.getKeywords() != null && !entry.getKeywords().isEmpty()) {
+            reasons.add("l0_keywords:" + String.join(",", entry.getKeywords()));
+        }
+        if (entry.getCategory() != null && !entry.getCategory().isBlank()) {
+            reasons.add("domain:" + entry.getCategory());
+        }
+        return reasons;
+    }
+
+    private String sourceKey(EvidenceBlock block, String fallback) {
+        return firstNonBlank(block.getSource(), block.getTitle(), block.getBreadcrumb(), fallback);
+    }
+
+    private Map<String, String> parseMetadata(String metadata) {
+        if (metadata == null || metadata.isBlank()) {
+            return Map.of();
+        }
+        try {
+            Map<?, ?> raw = objectMapper.readValue(metadata, Map.class);
+            Map<String, String> result = new LinkedHashMap<>();
+            for (Map.Entry<?, ?> entry : raw.entrySet()) {
+                if (entry.getKey() != null && entry.getValue() != null) {
+                    result.put(String.valueOf(entry.getKey()), String.valueOf(entry.getValue()));
+                }
+            }
+            return result;
+        } catch (Exception e) {
+            return Map.of();
+        }
+    }
+
+    private String firstNonBlank(String... values) {
+        for (String value : values) {
+            if (value != null && !value.isBlank()) {
+                return value;
+            }
+        }
+        return null;
+    }
+
+    private String truncate(String text, int maxLength) {
+        if (text == null || text.length() <= maxLength) {
+            return text;
+        }
+        return text.substring(0, maxLength) + "...";
+    }
+
+    private record EvidencePostprocessResult(int candidateCount, List<EvidenceBlock> blocks) {}
+
     private int countMdHeadings(String content) {
         if (content == null) return 0;
         return (int) content.lines()

src/test/java/com/superbiz/agent/service/ToolInvocationRecorderTest.java

@@ -2,6 +2,7 @@ package com.superbiz.agent.service;
 
 import com.fasterxml.jackson.databind.ObjectMapper;
 import com.superbiz.agent.domain.entity.ToolInvocation;
+import com.superbiz.agent.dto.EvidenceBlock;
 import com.superbiz.agent.repository.ToolInvocationRepository;
 import com.superbiz.agent.util.SessionContextHolder;
 import org.junit.jupiter.api.Test;
@@ -78,6 +79,14 @@ class ToolInvocationRecorderTest {
                 .l0MatchedKeywords(List.of("ERR_TIMEOUT"))
                 .l0Domains(List.of("payment"))
                 .l0Entities(List.of("ERR_TIMEOUT"))
+                .evidenceCandidateCount(2)
+                .evidenceBlockCount(1)
+                .evidenceBlocks(List.of(Map.of(
+                        "source", "payment/errors.md",
+                        "title", "payment/errors.md",
+                        "retrieval_layer", "L0+L1",
+                        "hit_reasons", List.of("l0_keywords:ERR_TIMEOUT", "semantic_rank:1")
+                )))
                 .build();
 
         try {
@@ -98,5 +107,45 @@ class ToolInvocationRecorderTest {
         assertTrue(saved.getRetrievalDetails().contains("\"l0_matched_keywords\":[\"ERR_TIMEOUT\"]"));
         assertTrue(saved.getRetrievalDetails().contains("\"l0_domains\":[\"payment\"]"));
         assertTrue(saved.getRetrievalDetails().contains("\"l0_entities\":[\"ERR_TIMEOUT\"]"));
+        assertTrue(saved.getRetrievalDetails().contains("\"evidence_candidate_count\":2"));
+        assertTrue(saved.getRetrievalDetails().contains("\"evidence_block_count\":1"));
+        assertTrue(saved.getRetrievalDetails().contains("\"evidence_blocks\""));
+    }
+
+    @Test
+    void lookupKnowledgeRecordFromSummarizesEvidenceBlocks() {
+        EvidenceBlock block = EvidenceBlock.builder()
+                .source("doc.md")
+                .title("Doc")
+                .breadcrumb("A > B")
+                .retrievalLayer("L1")
+                .score(0.42)
+                .hitReasons(List.of("semantic_rank:1"))
+                .content("x".repeat(220))
+                .build();
+
+        com.superbiz.agent.dto.LookupResult result = com.superbiz.agent.dto.LookupResult.builder()
+                .found(true)
+                .evidenceCandidateCount(3)
+                .evidenceBlockCount(1)
+                .evidenceBlocks(List.of(block))
+                .build();
+
+        ToolInvocationRecorder.LookupKnowledgeRecord record = ToolInvocationRecorder.LookupKnowledgeRecord.from(
+                "query",
+                KnowledgeIndexService.L0Hint.empty(),
+                List.of(),
+                false,
+                result,
+                null,
+                null,
+                10,
+                -1
+        );
+
+        assertEquals(3, record.evidenceCandidateCount());
+        assertEquals(1, record.evidenceBlockCount());
+        assertEquals(1, record.evidenceBlocks().size());
+        assertTrue(String.valueOf(record.evidenceBlocks().get(0).get("content_preview")).endsWith("..."));
     }
 }

src/test/java/com/superbiz/agent/tool/LookupKnowledgeToolTest.java

@@ -14,6 +14,7 @@ import org.mockito.MockitoAnnotations;
 
 import java.util.Collections;
 import java.util.List;
+import java.util.Map;
 
 import static org.junit.jupiter.api.Assertions.*;
 import static org.mockito.ArgumentMatchers.*;
@@ -83,6 +84,12 @@ class LookupKnowledgeToolTest {
         assertTrue(content.contains("Test content"));
         assertNotNull(result.getSupplement()); // L0 唯一命中仍调用 L1
         assertEquals("PRECISE", result.getRelevanceLevel());
+        assertNotNull(result.getEvidenceBlocks());
+        assertEquals(2, result.getEvidenceCandidateCount());
+        assertEquals(2, result.getEvidenceBlockCount());
+        assertEquals("L0", result.getEvidenceBlocks().get(0).getRetrievalLayer());
+        assertTrue(result.getEvidenceBlocks().get(0).getHitReasons().stream()
+                .anyMatch(reason -> reason.contains("ERR_TIMEOUT")));
 
         verify(vectorSearchService).searchSimilarDocuments("ERR_TIMEOUT", 3, null);
     }
@@ -290,6 +297,41 @@ class LookupKnowledgeToolTest {
         verify(vectorSearchService).searchSimilarDocuments("fallback", 3, null);
     }
 
+    @Test
+    void testLookup_deduplicatesEvidenceBlocksBySource() throws Exception {
+        KnowledgeEntry entry = KnowledgeEntry.builder()
+                .filePath("shared.md")
+                .title("Shared Doc")
+                .keywords(List.of("shared"))
+                .summary("Shared summary")
+                .build();
+        VectorSearchService.SearchResult l1Result = new VectorSearchService.SearchResult();
+        l1Result.setContent("Shared semantic content");
+        String metadata = "{\"docId\":\"doc-1\",\"_source\":\"shared.md\",\"title\":\"Shared Doc\"}";
+        l1Result.setMetadata(metadata);
+        l1Result.setScore(0.2f);
+
+        when(knowledgeIndexService.analyzeQuery("shared"))
+                .thenReturn(hint(entry));
+        when(knowledgeIndexService.readDocument("shared.md", 2000))
+                .thenReturn("Shared content");
+        when(vectorSearchService.searchSimilarDocuments("shared", 3, null))
+                .thenReturn(List.of(l1Result));
+        when(objectMapper.readValue(metadata, Map.class))
+                .thenReturn(Map.of("docId", "doc-1", "_source", "shared.md", "title", "Shared Doc"));
+
+        LookupResult result = tool.lookupKnowledge("shared");
+
+        assertTrue(result.isFound());
+        assertEquals(2, result.getEvidenceCandidateCount());
+        assertEquals(1, result.getEvidenceBlockCount());
+        assertEquals("shared.md", result.getEvidenceBlocks().get(0).getSource());
+        assertEquals("L0+L1", result.getEvidenceBlocks().get(0).getRetrievalLayer());
+        assertTrue(result.getEvidenceBlocks().get(0).getHitReasons().contains("semantic_rank:1"));
+        assertTrue(result.getEvidenceBlocks().get(0).getHitReasons().stream()
+                .anyMatch(reason -> reason.startsWith("l0_keywords:")));
+    }
+
     private KnowledgeIndexService.L0Hint hint(KnowledgeEntry... entries) {
         List<KnowledgeEntry> matches = List.of(entries);
         List<String> keywords = matches.stream()