[Upstream #145] 知识图谱中存在重复实体节点

[ivanzud]

Upstream Issue Mirror

• Upstream issue: 知识图谱中存在重复实体节点 666ghj/MiroFish#145
• Upstream repository: 666ghj/MiroFish
• State: open
• Author: Stayfoool
• Labels: none
• Last updated: 2026-03-11T14:53:58Z

Summary

问题描述 在使用 MiroFish 构建知识图谱时，Zep 会将同一现实实体识别为多个不同节点。 例如输入包含"特朗普"相关内容的文本后，图谱中会同时出现"特朗普"和 "美国总统特朗普"两个独立节点，它们各自有独立的边和关系。 这会导致： - 图谱中同一实体的信息被分散到多个节点上 - 后续的模拟推演基于不完整的实体关系进行，影响准确性 - 图谱可视化时出现冗余节点，影响可读性 ## 复现步骤 1. 准备一段包含同一人物/组织不同称呼的背景文本 2. 通过前端正常流程构建知识图谱 3. 查看生成的图谱，可以看到同一实体被拆分为多个节点 ## 截图 <img width="675" height="399" alt="Image" src="https://github.com/user-attachments/assets/593f4188-e766-46b3-9b88-25486…

Local Coverage

• Status: partial
• Summary: Repo-native partial mitigations are now landed locally across simulation inputs, backend graph/report/search/statistics/detail surfaces, raw graph introspection, node-edge introspection, textual tool output, both shipped graph renderers, and the visible frontend graph counters/logs: ZepEntityReader.filter_defined_entities() collapses obvious same-entity alias variants before simulation/profile generation, ZepEntityReader.get_entity_with_context() now merges alias-linked relations and related nodes for the entity-detail API, backend/app/services/graph_builder.py now collapses the same conservative alias pairs when serving /api/graph/data/<graph_id> and remaps duplicate edges to the retained node UUID, backend/app/services/zep_tools.py now collapses those aliases when building typed entity lists, raw node/edge introspection payloads, Panorama output, InsightForge entity/relationship summaries, QuickSearch/general search results, graph statistics, node-edge lookups, entity summaries including relations attached only to alias UUIDs, and NodeInfo.to_text() output, while preserving merged alias_names metadata so callers and downstream prompts can still see which labels were folded together. frontend/src/views/processGraphData.js and the shared frontend/src/components/GraphPanel.vue renderer now both collapse them while rendering graph data, the Process plus GraphPanel node detail drawers expose the folded non-canonical aliases via frontend/src/components/graphAliasDetails.js, and frontend/src/components/graphPanelData.js now drives deduplicated Step 1 / Process counters plus MainView refresh logs so title-prefixed duplicates such as 美国总统特朗普 vs 特朗普 no longer appear twice in the graph or its visible counts. Full graph-level persisted deduplication still remains tracked under beads issue mirofish-975 because upstream PR feat: add entity deduplication after graph building 666ghj/MiroFish#141 is not safe to cherry-pick wholesale.
• Local refs: .beads/issues.jsonl, docs/upstream-triage.md, origin/mirror/upstream-pr-141, backend/app/services/zep_entity_reader.py, backend/tests/test_zep_entity_reader.py, backend/app/services/graph_builder.py, backend/tests/test_graph_builder.py, backend/app/services/zep_tools.py, backend/tests/test_zep_tools_dedup.py, backend/tests/test_zep_tools_i18n.py, frontend/src/components/GraphPanel.vue, frontend/src/components/graphAliasDetails.js, frontend/src/components/graphPanelData.js, frontend/src/components/Step1GraphBuild.vue, frontend/tests/graphAliasDetails.test.mjs, frontend/tests/graphPanelData.test.mjs, frontend/src/views/MainView.vue, frontend/src/views/Process.vue, frontend/src/views/processGraphData.js, frontend/tests/processGraphData.test.mjs
• Notes: Mirrored into fork issue tracking on March 11, 2026 after enabling issues on ivanzud/MiroFish. 2026-03-11: Added a third repo-native partial mitigation in backend report/search tooling so conservative alias pairs are collapsed before Step 4/analysis surfaces render entity summaries and relationship chains. 2026-03-11: Added a fourth repo-native partial mitigation in backend graph-data responses so /api/graph/data/<graph_id> collapses the same obvious alias pairs and remaps duplicate edges without mutating stored Zep nodes. 2026-03-11: Added a fifth repo-native partial mitigation in QuickSearch/general search results so search_graph() and the local-search fallback collapse the same obvious alias pairs, deduplicate summary facts, and remap duplicate edges before report tooling or API callers consume the result payload. 2026-03-11: Added a seventh repo-native partial mitigation in backend graph statistics and entity-summary helpers so alias duplicates no longer inflate stats counts and alias queries resolve to the canonical merged entity payload. 2026-03-11: Added an eighth repo-native partial mitigation in backend entity summaries so alias-linked relations are collected from the full edge set before canonical remapping, which prevents summaries from dropping edges attached only to an alias UUID. 2026-03-11: Added a ninth repo-native partial mitigation in backend entity-detail reads so get_entity_with_context() resolves the requested node against its alias group, includes alias-linked relations, and deduplicates related-node payloads. 2026-03-11: Added a tenth repo-native partial mitigation in zep_tools node-edge lookups so get_node_edges() now resolves the requested node against its alias group and remaps duplicate edges instead of dropping relationships attached only to an alias UUID. 2026-03-11: Added an eleventh repo-native partial mitigation in raw zep_tools graph introspection so get_all_nodes() and get_all_edges() now collapse obvious alias duplicates and remap duplicate edge payloads before report-side callers consume the graph snapshot. 2026-03-11: Added a fourteenth repo-native partial mitigation in textual zep_tools node rendering so NodeInfo.to_text() now includes merged alias names with locale-aware labels, preserving the folded names in downstream report/runtime prompts. 2026-03-11: Added a fifteenth repo-native partial mitigation in the shared frontend graph renderer so frontend/src/components/GraphPanel.vue now reuses the conservative display-only alias-collapse mapper and no longer shows obvious duplicate entities outside the Process view. 2026-03-11: Added a sixteenth repo-native partial mitigation in the Process and shared GraphPanel node detail drawers so frontend/src/components/graphAliasDetails.js filters merged alias_names down to the folded non-canonical labels and both detail panels render them explicitly instead of hiding which source names collapsed into the canonical node. 2026-03-11: Added a seventeenth repo-native partial mitigation in frontend graph stats so frontend/src/components/graphPanelData.js now drives Step 1 / Process counters and MainView refresh logs through the same alias-collapse mapping as the visible graph renderer.