feat(fusion): v7.0 multi-modal fusion engine enhancements + benchmarks

Four major v7.0 capabilities:
- Vector embedding semantic matching (Gemini text-embedding-004, Tier 2.5)
- LLM-enhanced strategy routing (Gemini 2.0 Flash, strategy="llm_auto")
- Geographic knowledge graph (networkx DiGraph, 7 entity types, 5 relations)
- Distributed/chunked computing (dask + fiona, >500K rows auto-chunked)

Benchmark suite (9 metrics, all passing):
- Semantic matching: P=1.00, R=0.95, F1=0.98
- Strategy selection: 100% accuracy (30/30 scenarios)
- Unit conversion: zero error (8 conversion pairs)
- Compatibility scoring: 100% band accuracy
- Fusion quality: mean 0.96 ± 0.10 (8 strategies)
- Defect detection: 100% (7/7 defect types)
- Data preservation: row 1.00, geometry 1.00
- Coverage: 5/5 modalities, 10/10 strategies, 11/11 type pairs
- Performance: ~30K rows/s at 10K scale

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: Gemini CLI

CLAUDE.md

@@ -4,7 +4,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 
 ## Project Overview
 
-GIS Data Agent (ADK Edition) v4.0 — an AI-powered geospatial analysis platform built on **Google Agent Developer Kit (ADK)**. It uses LLM-based semantic routing to dispatch user requests across three specialized pipelines for data governance, land-use optimization (via Deep Reinforcement Learning), and general spatial intelligence. The frontend is a custom React three-panel SPA served via **Chainlit** with password/OAuth2 authentication.
+GIS Data Agent (ADK Edition) v7.0 — an AI-powered geospatial analysis platform built on **Google Agent Developer Kit (ADK)**. It uses LLM-based semantic routing to dispatch user requests across three specialized pipelines for data governance, land-use optimization (via Deep Reinforcement Learning), and general spatial intelligence. The frontend is a custom React three-panel SPA served via **Chainlit** with password/OAuth2 authentication.
 
 ## Commands
 
@@ -16,7 +16,7 @@ Default login: `admin` / `admin123` (seeded on first run). In-app self-registrat
 
 ### Run tests
 ```bash
-# All tests (1180+ tests)
+# All tests (1330+ tests)
 .venv/Scripts/python.exe -m pytest data_agent/ --ignore=data_agent/test_knowledge_agent.py -q
 
 # Single test file
@@ -130,11 +130,12 @@ All endpoints use JWT cookie auth. Routes mounted before Chainlit catch-all via
 | `multimodal.py` | Multimodal input processing — image/PDF classification, Gemini Part builders |
 | `workflow_engine.py` | Multi-step workflow engine — CRUD, execution, webhook push, cron scheduling |
 | `fusion_engine.py` | Multi-modal data fusion — profiling, compatibility, alignment, 10 fusion strategies |
+| `knowledge_graph.py` | Geographic knowledge graph — networkx-based entity-relationship modeling |
 | `evals/agent.py` | Evaluation umbrella agent — wraps 4 pipelines as sub_agents for ADK AgentEvaluator |
 | `run_evaluation.py` | Multi-pipeline ADK evaluation runner with per-metric scoring and charts |
 
-### Toolsets (18 modules in `toolsets/`)
-Exploration, GeoProcessing, Visualization (10 tools incl. `generate_3d_map`, `control_map_layer`), Analysis, Database, SemanticLayer (9 tools), DataLake (8 tools), Streaming (5 tools), Team (8 tools), Location, Memory, Admin, File, RemoteSensing, SpatialStatistics, SkillBundles, McpHub, Fusion (4 tools).
+### Toolsets (19 modules in `toolsets/`)
+Exploration, GeoProcessing, Visualization (10 tools incl. `generate_3d_map`, `control_map_layer`), Analysis, Database, SemanticLayer (9 tools), DataLake (8 tools), Streaming (5 tools), Team (8 tools), Location, Memory, Admin, File, RemoteSensing, SpatialStatistics, SkillBundles, McpHub, Fusion (4 tools), KnowledgeGraph (3 tools).
 
 ### Data Loading (`_load_spatial_data` in `agent.py`)
 Supported formats: CSV, Excel (.xlsx/.xls), Shapefile, GeoJSON, GPKG, KML, KMZ. CSV/Excel auto-detect coordinate columns (lng/lat, lon/lat, longitude/latitude, x/y).

README.md

@@ -1,16 +1,22 @@
 [English](./README_en.md) | **中文**
 
-# GIS Data Agent (ADK Edition) v6.0
+# GIS Data Agent (ADK Edition) v7.0
 
-基于 **Google Agent Developer Kit (ADK)** 构建的 AI 驱动地理空间分析平台。通过自然语言语义路由，自动调度四大专业管道完成空间数据治理、用地优化、多源数据融合和商业智能分析。前端为 React 三面板 SPA，后端集成 30 个 REST API，支持多模态输入、3D 可视化和工作流编排。
+基于 **Google Agent Developer Kit (ADK)** 构建的 AI 驱动地理空间分析平台。通过自然语言语义路由，自动调度四大专业管道完成空间数据治理、用地优化、多源数据融合和商业智能分析。前端为 React 三面板 SPA，后端集成 30 个 REST API，支持多模态输入、3D 可视化、工作流编排和知识图谱推理。
 
 ## 核心能力
 
-### 多源数据智能融合 (v5.5–v6.0)
+### 多源数据智能融合 (v5.5–v7.0)
 - **五阶段流水线**：画像 → 评估 → 对齐 → 融合 → 验证
 - **10 种融合策略**：空间连接、属性连接、分区统计、点采样、波段叠加、矢量叠加、时态融合、点云高度赋值、栅格矢量化、最近邻连接
 - **5 种数据模态**：矢量、栅格、表格、点云 (LAS/LAZ)、实时流
-- **智能语义匹配**：目录驱动等价组 + 分词相似度（camelCase↔snake_case）+ 类型兼容检查 + 单位自动转换
+- **智能语义匹配**：
+  - 四层渐进式匹配：精确 → 等价组 → 单位感知 → 模糊匹配
+  - **v7.0 向量嵌入匹配**：Gemini text-embedding-004 语义相似度（可选启用）
+  - 目录驱动等价组 + 分词相似度 + 类型兼容检查 + 单位自动转换
+- **LLM 增强策略路由 (v7.0)**：Gemini 2.0 Flash 根据用户意图智能推荐融合策略
+- **分布式/核外计算 (v7.0)**：大数据集（>50万行/500MB）自动分块处理，避免 OOM
+- **地理知识图谱 (v7.0)**：networkx 实体关系建模，空间邻接/包含关系检测，N跳邻居查询
 - **栅格自动处理**：CRS 重投影、分辨率重采样、大栅格窗口采样
 - **增强质量验证**：10 项检查（空值率、几何有效性、拓扑验证、KS 分布偏移检测等）
 
@@ -120,7 +126,11 @@ cd frontend && npm install && npm run dev
 | | NL 图层控制 | 自然语言 显示/隐藏/样式/移除 地图图层 |
 | | MCP 工具市场 | 配置驱动的 MCP 服务器连接 + 工具聚合 |
 | **数据融合** | 融合引擎 (MMFE) | 五阶段流水线（画像→评估→对齐→融合→验证），10 种策略，5 种模态 |
-| | 语义匹配 | 目录驱动等价组 + 分词相似度 + 类型兼容检查 + 单位自动转换 |
+| | 语义匹配 | 五层渐进匹配：精确 → 等价组 → 嵌入相似度 → 单位感知 → 模糊 |
+| | 嵌入匹配 (v7.0) | Gemini text-embedding-004 向量语义匹配（可选启用） |
+| | LLM 策略路由 (v7.0) | Gemini 2.0 Flash 意图感知策略推荐 (`strategy="llm_auto"`) |
+| | 知识图谱 (v7.0) | networkx 空间实体关系建模，N跳邻居查询，最短路径 |
+| | 分布式计算 (v7.0) | 大数据集 (>500K行) 自动分块读取和处理 |
 | | 栅格处理 | 自动 CRS 重投影、分辨率重采样、大栅格窗口采样 |
 | | 点云 & 流数据 | LAS/LAZ 高度赋值、CSV/JSON 流数据时态融合（时间窗口+空间聚合） |
 | | 质量验证 | 10 项检查：空值/几何/拓扑/CRS/微面/异常值/KS分布偏移 |
@@ -180,22 +190,24 @@ data_agent/
 ├── workflow_engine.py           # 工作流引擎：CRUD、执行、Webhook、Cron 调度
 ├── multimodal.py                # 多模态输入：图片/PDF 分类、Gemini Part 构建
 ├── mcp_hub.py                   # MCP Hub Manager：配置驱动的 MCP 服务器管理
-├── fusion_engine.py                # 多模态数据融合引擎（MMFE，~1750 行）
+├── fusion_engine.py                # 多模态数据融合引擎（MMFE，~2100 行）
+├── knowledge_graph.py              # 地理知识图谱引擎（networkx，~625 行）
 ├── pipeline_runner.py           # 无头管道执行器 (run_pipeline_headless)
-├── toolsets/                    # 18 个 BaseToolset 模块
+├── toolsets/                    # 19 个 BaseToolset 模块
 │   ├── visualization_tools.py   #   10 个工具：分级设色、热力图、3D、图层控制
 │   ├── fusion_tools.py          #   数据融合工具集（4 个工具）
+│   ├── knowledge_graph_tools.py #   知识图谱工具集（3 个工具）
 │   ├── mcp_hub_toolset.py       #   MCP 工具桥接
 │   ├── skill_bundles.py         #   5 个命名工具集分组
 │   └── ...                      #   探查、地理处理、分析、数据库、语义层等
 ├── prompts/                     # 3 个 YAML 提示词文件
-├── migrations/                  # 18 个 SQL 迁移脚本 (001-018)
+├── migrations/                  # 19 个 SQL 迁移脚本 (001-019)
 ├── locales/                     # 国际化：zh.yaml + en.yaml
 ├── db_engine.py                 # 连接池单例
 ├── health.py                    # K8s 健康检查 API
 ├── observability.py             # 结构化日志 + Prometheus
 ├── i18n.py                      # 国际化：YAML + t() 函数
-├── test_*.py                    # 61 个测试文件 (1290+ 测试)
+├── test_*.py                    # 62 个测试文件 (1330+ 测试)
 └── run_evaluation.py            # Agent 评估运行器
 
 frontend/
@@ -270,7 +282,7 @@ docs/                            # 文档
 ## 运行测试
 
 ```bash
-# 全量测试 (1290+ 测试)
+# 全量测试 (1330+ 测试)
 python -m pytest data_agent/ --ignore=data_agent/test_knowledge_agent.py -q
 
 # 单个模块
@@ -305,9 +317,10 @@ GitHub Actions 工作流（`.github/workflows/ci.yml`）在 push 到 `main`/`dev
 | v5.4 | 工作流编排（引擎 + 编辑器 + Cron + Webhook） | 完成 |
 | v5.5 | 多模态数据融合引擎 MMFE（5 模态、10 策略、语义匹配） | 完成 |
 | v5.6 | MGIM 启发增强（模糊匹配、单位转换、数据感知策略、多源编排） | 完成 |
-| v6.0 | 融合引擎不足修复（栅格重投影、点云高度、流数据时态、语义增强、质量验证） | **当前** |
+| v6.0 | 融合引擎不足修复（栅格重投影、点云高度、流数据时态、语义增强、质量验证） | 完成 |
 | v6.1 | 高级分析引擎（时空预测、场景模拟、网络分析） | 计划中 |
-| v7.0 | 实时协作、知识图谱、边缘部署 | 规划中 |
+| v7.0 | 向量嵌入匹配、LLM 策略路由、地理知识图谱、分布式计算 | **当前** |
+| v8.0 | 实时协作、边缘部署、数据连接器生态 | 规划中 |
 
 ## 许可证
 

README_en.md

@@ -1,16 +1,22 @@
 **English** | [中文](./README.md)
 
-# GIS Data Agent (ADK Edition) v6.0
+# GIS Data Agent (ADK Edition) v7.0
 
-An AI-powered geospatial analysis platform that turns natural language into spatial intelligence. Built on **Google Agent Developer Kit (ADK)** with semantic intent routing, four specialized pipelines, a React three-panel frontend, and enterprise-grade security. Features multi-source data fusion, multimodal input, 3D visualization, and workflow orchestration.
+An AI-powered geospatial analysis platform that turns natural language into spatial intelligence. Built on **Google Agent Developer Kit (ADK)** with semantic intent routing, four specialized pipelines, a React three-panel frontend, and enterprise-grade security. Features multi-source data fusion, multimodal input, 3D visualization, workflow orchestration, and geographic knowledge graph.
 
 ## Core Capabilities
 
-### Multi-Source Data Fusion (v5.5–v6.0)
+### Multi-Source Data Fusion (v5.5–v7.0)
 - **Five-stage pipeline**: Profile → Assess → Align → Fuse → Validate
 - **10 fusion strategies**: spatial join, attribute join, zonal statistics, point sampling, band stack, overlay, temporal fusion, point cloud height assignment, raster vectorize, nearest join
 - **5 data modalities**: vector, raster, tabular, point cloud (LAS/LAZ), real-time stream
-- **Intelligent semantic matching**: catalog-driven equivalence groups + tokenized similarity (camelCase↔snake_case) + type compatibility + automatic unit conversion
+- **Intelligent semantic matching**:
+  - Five-tier progressive matching: exact → equivalence groups → embedding similarity → unit-aware → fuzzy
+  - **v7.0 Vector embedding matching**: Gemini text-embedding-004 cosine similarity (opt-in)
+  - Catalog-driven equivalence groups + tokenized similarity + type compatibility + auto unit conversion
+- **LLM-enhanced strategy routing (v7.0)**: Gemini 2.0 Flash intent-aware strategy recommendation
+- **Distributed/out-of-core computing (v7.0)**: Auto-chunked processing for large datasets (>500K rows / >500MB)
+- **Geographic knowledge graph (v7.0)**: networkx entity-relationship modeling, spatial adjacency/containment detection, N-hop neighbor queries
 - **Raster auto-processing**: CRS reprojection, resolution resampling, windowed sampling for large rasters
 - **Enhanced quality validation**: 10 checks (null rate, geometry validity, topology, KS distribution shift, etc.)
 
@@ -120,7 +126,11 @@ Default login: `admin` / `admin123` (seeded on first run). In-app self-registrat
 | | NL Layer Control | Natural language show/hide/style/remove map layers via `control_map_layer` tool |
 | | MCP Tool Market | Config-driven MCP server connection + tool aggregation |
 | **Data Fusion** | Fusion Engine (MMFE) | Five-stage pipeline (Profile→Assess→Align→Fuse→Validate), 10 strategies, 5 modalities |
-| | Semantic Matching | Catalog-driven equivalence groups + tokenized similarity + type compatibility + auto unit conversion |
+| | Semantic Matching | Five-tier progressive: exact → equivalence groups → embedding similarity → unit-aware → fuzzy |
+| | Embedding Matching (v7.0) | Gemini text-embedding-004 vector semantic matching (opt-in) |
+| | LLM Strategy Routing (v7.0) | Gemini 2.0 Flash intent-aware strategy recommendation (`strategy="llm_auto"`) |
+| | Knowledge Graph (v7.0) | networkx spatial entity-relationship modeling, N-hop queries, shortest path |
+| | Distributed Computing (v7.0) | Auto-chunked processing for large datasets (>500K rows) |
 | | Raster Processing | Auto CRS reprojection, resolution resampling, windowed sampling for large rasters |
 | | Point Cloud & Stream | LAS/LAZ height assignment, CSV/JSON stream temporal fusion (time window + spatial aggregation) |
 | | Quality Validation | 10 checks: null rate, geometry, topology, CRS, micro-polygons, outliers, KS distribution shift |
@@ -180,22 +190,24 @@ data_agent/
 ├── workflow_engine.py           # Workflow engine: CRUD, execution, webhook, cron
 ├── multimodal.py                # Multimodal input: image/PDF classification, Gemini Parts
 ├── mcp_hub.py                   # MCP Hub Manager: config-driven MCP server management
-├── fusion_engine.py                # Multi-modal Data Fusion Engine (MMFE, ~1750 lines)
+├── fusion_engine.py                # Multi-modal Data Fusion Engine (MMFE, ~2100 lines)
+├── knowledge_graph.py              # Geographic Knowledge Graph Engine (networkx, ~625 lines)
 ├── pipeline_runner.py           # Headless pipeline executor (run_pipeline_headless)
-├── toolsets/                    # 18 BaseToolset modules
+├── toolsets/                    # 19 BaseToolset modules
 │   ├── visualization_tools.py   #   10 tools: choropleth, heatmap, 3D, layer control
 │   ├── fusion_tools.py          #   Data fusion toolset (4 tools)
+│   ├── knowledge_graph_tools.py #   Knowledge graph toolset (3 tools)
 │   ├── mcp_hub_toolset.py       #   MCP tool bridge
 │   ├── skill_bundles.py         #   5 named toolset groupings
 │   └── ...                      #   exploration, geo processing, analysis, database, etc.
 ├── prompts/                     # 3 YAML prompt files
-├── migrations/                  # 18 SQL migration scripts (001-018)
+├── migrations/                  # 19 SQL migration scripts (001-019)
 ├── locales/                     # i18n: zh.yaml + en.yaml
 ├── db_engine.py                 # Connection pool singleton
 ├── health.py                    # K8s health check API
 ├── observability.py             # Structured logging + Prometheus
 ├── i18n.py                      # i18n: YAML dict + t() function
-├── test_*.py                    # 61 test files (1290+ tests)
+├── test_*.py                    # 62 test files (1330+ tests)
 └── run_evaluation.py            # Agent evaluation runner
 
 frontend/
@@ -271,7 +283,7 @@ Custom React SPA replacing Chainlit's default UI:
 ## Running Tests
 
 ```bash
-# All tests (1290+ tests)
+# All tests (1330+ tests)
 python -m pytest data_agent/ --ignore=data_agent/test_knowledge_agent.py -q
 
 # Single module
@@ -306,9 +318,10 @@ GitHub Actions workflow (`.github/workflows/ci.yml`) runs on push to `main`/`dev
 | v5.4 | Workflow Builder (Engine + Editor + Cron + Webhook) | Done |
 | v5.5 | Multi-Modal Data Fusion Engine MMFE (5 modalities, 10 strategies, semantic matching) | Done |
 | v5.6 | MGIM-Inspired Enhancements (fuzzy matching, unit conversion, data-aware scoring, multi-source) | Done |
-| v6.0 | Fusion Engine Improvements (raster reprojection, point cloud, stream temporal, semantic+quality) | **Current** |
+| v6.0 | Fusion Engine Improvements (raster reprojection, point cloud, stream temporal, semantic+quality) | Done |
 | v6.1 | Advanced Analytics Engine (spatiotemporal prediction, scenario simulation, network analysis) | Planned |
-| v7.0 | Real-time Collaboration, Knowledge Graph, Edge Deployment | Future |
+| v7.0 | Vector Embedding Matching, LLM Strategy Routing, Knowledge Graph, Distributed Computing | **Current** |
+| v8.0 | Real-time Collaboration, Edge Deployment, Data Connector Ecosystem | Future |
 
 ## License
 

data_agent/agent.py

@@ -47,6 +47,7 @@
     DataLakeToolset,
     McpHubToolset,
     FusionToolset,
+    KnowledgeGraphToolset,
 )
 
 # ArcPy conditional function lists (for governance agents needing specific subsets)
@@ -297,6 +298,7 @@
         DataLakeToolset(),
         McpHubToolset(pipeline="general"),
         FusionToolset(),
+        KnowledgeGraphToolset(),
     ] + _arcpy_tools,
 )
 
@@ -384,6 +386,7 @@ def _make_planner_processor(name: str, **overrides) -> LlmAgent:
             DatabaseToolset(tool_filter=["import_to_postgis"]),
             McpHubToolset(pipeline="planner"),
             FusionToolset(),
+            KnowledgeGraphToolset(),
         ] + _arcpy_tools,
     )
     defaults.update(overrides)

data_agent/app.py

@@ -121,6 +121,10 @@
     ensure_chainlit_tables()
     from data_agent.workflow_engine import ensure_workflow_tables
     ensure_workflow_tables()
+    from data_agent.fusion_engine import ensure_fusion_tables
+    ensure_fusion_tables()
+    from data_agent.knowledge_graph import ensure_knowledge_graph_tables
+    ensure_knowledge_graph_tables()
 except Exception as _startup_err:
     logger.warning("DB initialization partially failed: %s", _startup_err)
     # Ensure resolve_semantic_context/build_context_prompt are importable even on failure