Skip to content

CHANGELOG.md

Latest commit

 

History

History
288 lines (240 loc) · 27.5 KB

CHANGELOG.md

File metadata and controls

288 lines (240 loc) · 27.5 KB

Changelog

All notable changes to the GIS Data Agent project.

[v4.0.0] - 2026-03-02 (Frontend Integration, CI & Post-Launch Enhancement)

Feature — React-Integrated Self-Registration (F1)

  • In-app registration: Login page now includes register mode toggle (username, email, password, confirm password, display name fields). Calls POST /auth/register directly.
  • Email support: register_user() accepts optional email parameter with regex validation. email column added to agent_app_users table.
  • Migration: 015_add_email_column.sql adds email column.

Feature — Token Usage Dashboard (F2)

  • UsageView component in DataPanel: 5th tab ("用量") showing daily count progress bar, monthly token summary, and pipeline breakdown visualization.
  • Pipeline breakdown API: get_pipeline_breakdown() in token_tracker.py returns [{pipeline_type, count, tokens}] grouped by pipeline for current month.
  • Auto-refresh: Dashboard refreshes every 30 seconds.

Feature — Natural Language Layer Control (F3)

  • control_map_layer() tool: 5 actions — show, hide, style, remove, list. Supports color/opacity customization.
  • Pipeline: Tool result → app.py detects layer_control key → injects metadata → ChatPanel onLayerControl → App state → MapPanel useEffect handler.
  • MapPanel integration: Reacts to control commands by toggling layer visibility, updating styles, or removing layers from the Leaflet map.

Feature — Collaborative Map Annotations (F4)

  • map_annotations.py (new): CRUD functions for annotations with PostgreSQL storage (agent_map_annotations table).
  • 4 REST endpoints: GET/POST /api/annotations, PUT/DELETE /api/annotations/{id}.
  • MapPanel UI: Annotation toggle button (pencil icon), click-to-add with popup form (title + comment), marker rendering with resolve/delete actions.
  • Team scope: Users see own annotations + team annotations. Delete requires ownership verification.
  • Migration: 016_create_map_annotations.sql.

Feature — Tianditu Compliant Basemap (F5)

  • GET /api/config/basemaps: Returns {gaode_enabled, tianditu_enabled, tianditu_token} based on environment configuration.
  • Dynamic basemap loading: MapPanel fetches basemap config on mount. Tianditu Vec + Img layers loaded conditionally when token is available.
  • Static basemaps: Gaode (高德) added as permanent option alongside CartoDB and OpenStreetMap.

Feature — Toolset Optimization — Skill Bundles (F6)

  • skill_bundles.py (new): SkillBundle dataclass with 5 named bundles:
    • SPATIAL_ANALYSIS: Exploration + GeoProcessing + Location + RemoteSensing + SpatialStatistics + Analysis
    • DATA_QUALITY: Exploration + Database + File + SemanticLayer + DataLake
    • VISUALIZATION: Visualization tools (9 tools including layer control)
    • DATABASE: Database + DataLake
    • COLLABORATION: Memory + Admin + Team + DataLake
  • Intent matching: get_bundles_for_intent() and build_toolsets_for_intent() for dynamic toolset assembly.
  • Filter presets: AUDIT_TOOLS, TRANSFORM_TOOLS, DB_READ, GENERAL_VIZ, SEMANTIC_READONLY.

Feature — User Account Self-Deletion (F7)

  • delete_user_account() in auth.py: Password verification → cascade delete (token_usage, memories, share_links, team_members, audit_log, annotations) → remove upload directory → delete user account.
  • Admin protection: Admins cannot delete themselves via this method.
  • DELETE /api/user/account: REST endpoint with password confirmation in request body.
  • UserSettings.tsx (new): Modal component with user info display and danger zone for account deletion.
  • User menu: Header avatar click reveals dropdown menu with "账户设置" (settings) and "退出登录" (logout) options.

Feature — CI Pipeline + Evaluation (F8)

  • .github/workflows/ci.yml (new): 3 jobs:
    • test: Ubuntu + PostGIS service container, pytest with JUnit XML output
    • frontend: Node.js 20, npm ci && npm run build
    • evaluate: ADK agent evaluation on main push (requires GOOGLE_API_KEY secret)
  • JSON eval summary: run_evaluation.py now writes eval_results/eval_summary.json with scores and pass/fail status.
  • Additional eval cases: Governance check + general spatial query test cases added to eval_set.json.

Changed

  • Frontend API: 17 REST endpoints (up from 11), organized into catalog, semantic, pipeline, user, admin, annotation, and config groups.
  • Toolset count: 16 BaseToolset modules (up from 15, added skill_bundles.py).
  • Visualization tools: 9 tools (up from 8, added control_map_layer).
  • DataPanel tabs: 5 tabs (files, CSV, catalog, history, usage — added usage tab).
  • App.tsx: User dropdown menu, UserSettings modal integration, layer control wiring.

Tests

  • Total: 923 tests passing (up from 882).
  • New tests: frontend API (47 tests including delete account, basemap config, annotations), map annotations (11), skill bundles (5), layer control (7).

[v4.0.0-beta] - 2026-03-01 (Data Lake, Lineage & Hierarchy Enhancement)

Feature — Unified Data Lake Integration

  • Data Catalog (data_catalog.py): Unified agent_data_catalog table managing assets across local files, OBS cloud, and PostGIS backends. 8 ADK tools via DataLakeToolset: list_data_assets, describe_data_asset, search_data_assets (fuzzy + substring matching), register_data_asset, tag_data_asset, delete_data_asset, share_data_asset (admin-only), get_data_lineage.
  • Auto-registration: Tool outputs automatically registered in catalog via register_tool_output() in app.py. File uploads registered on ingest.

Feature — Data Lineage Tracking

  • Source tracking: source_assets JSONB field in data catalog now populated automatically from tool arguments. _extract_source_paths() scans tool args for file path keys.
  • get_data_lineage() ADK tool: Trace data provenance chain — ancestors (what was it derived from), descendants (what was derived from it), or both directions. Recursive chain walking up to 10 levels.

Feature — Health Check API

  • health.py (new): K8s-ready health endpoints — /health (liveness), /ready (readiness, 503 if DB down), /api/admin/system-info (admin HMAC auth).
  • Subsystem checks: Database latency, OBS cloud storage, Redis, session service.
  • Startup diagnostics: format_startup_summary() prints ASCII banner with all subsystem statuses and feature flags.
  • K8s probes: app-deployment.yaml updated to use /health and /ready.

Feature — 3-Level Semantic Hierarchy

  • Sub-children: LAND_USE hierarchy extended from 2 levels (parent→child) to 3 levels (parent→child→sub_child). Added sub-categories following China's Third National Land Survey codes:
    • 耕地 → 旱地[0101], 水浇地[0102], 水田[0103]
    • 园地 → 果园[0201], 茶园[0202], 其他园地[0204]
    • 林地 → 有林地[0301], 灌木林地[0302], 其他林地[0307]
    • 草地 → 天然牧草地[0401], 人工牧草地[0402], 其他草地[0403]
  • browse_hierarchy() ADK tool: Agents can explore the full taxonomy tree. Returns structured tree + formatted text display.
  • _match_hierarchy(): Sub-child matching with most-specific-first priority. Returns level: "sub_child" with 4-digit code prefix.
  • expand_hierarchy(): Expanding a child with sub_children returns sub_children for finer granularity.
  • SQL filters: Sub-child matches generate 4-digit LIKE patterns (e.g., dlbm LIKE '0103%').

Feature — Bot Integration (DingTalk + Feishu)

  • DingTalk bot (dingtalk_bot.py): Enterprise callback with HMAC signature verification.
  • Feishu bot (feishu_bot.py): Lark/Feishu with event challenge + AES decryption.
  • Shared infrastructure (bot_base.py): TokenCache, RateLimiter, MessageDedup.

Feature — Real-time Streams

  • Stream engine (stream_engine.py): Redis Streams with in-memory fallback.
  • 5 ADK tools: create_iot_stream, set_geofence_alert, query_trajectory, aggregate_stream_window, list_active_streams.
  • Geofence: Point-in-polygon check for real-time location alerts.

Feature — Team Collaboration

  • Team management (team_manager.py): 8 tools for team CRUD, member management, resource sharing.
  • RLS extended: Team members can access tables/templates owned by fellow team members.

Feature — Remote Sensing

  • 5 raster tools: describe_raster, calculate_ndvi, raster_band_math, classify_raster, visualize_raster.
  • Data download: download_lulc (Sentinel-2 10m, 2017-2024), download_dem (Copernicus 30m).

Tests

  • Total: 836 tests passing (up from 538).
  • New test files: test_health.py (24), test_data_catalog.py expanded (+19 lineage), test_semantic_layer.py expanded (+16 hierarchy), test_stream_engine.py (36), test_team.py (30), test_dingtalk_bot.py, test_feishu_bot.py (55 combined).

[v4.0.0-alpha.2] - 2026-02-27 (Architecture Polish & Semantic Enhancement)

Performance — Connection Pool Singleton

  • db_engine.py (new): Singleton SQLAlchemy engine with connection pooling (pool_size=5, max_overflow=10, pool_recycle=1800s, pool_pre_ping=True). Replaces 50 per-function create_engine() calls across 10 modules.
  • Refactored modules: database_tools.py, memory.py, semantic_layer.py, audit_logger.py, token_tracker.py, sharing.py, template_manager.py, auth.py, app.py, utils.py — all now use from .db_engine import get_engine.
  • reset_engine(): For testing and graceful shutdown.

Performance — Semantic Layer Query Cache

  • TTL cache (5 min) for agent_semantic_sources and agent_semantic_registry DB queries in semantic_layer.py. Eliminates redundant DB lookups on every user message.
  • invalidate_semantic_cache(): Called automatically on write operations (register, annotate, auto-register).

Feature — Domain Hierarchy (LAND_USE)

  • Hierarchy tree in semantic_catalog.yaml: 3 parent categories (农用地, 建设用地, 未利用地) with 9 child land-use types, each with code prefixes and Chinese/English aliases.
  • _match_hierarchy(): Matches "耕地" → child (code_prefix 01, parent 农用地); "农用地" → parent (expands to 耕地+园地+林地+草地).
  • Prompt injection: build_context_prompt() outputs hierarchy info like "地类筛选: 农用地 (包含 耕地[01*], 园地[02*], 林地[03*], 草地[04*])".

Feature — Column Equivalence Mappings

  • Equivalences in semantic_catalog.yaml: dlbm↔dlmc (地类编码↔名称), xzqdm↔xzqmc (行政区代码↔名称), qsdwdm↔qsdwmc (权属代码↔名称).
  • _match_equivalences(): When a column like dlbm is matched, automatically associates its equivalent dlmc.
  • Prompt injection: Outputs "等价列: dlbm ↔ dlmc (地类编码 ↔ 地类名称)".

Documentation

  • README.md rewritten for v4.0-alpha: updated architecture diagram (Mermaid), full feature matrix, Docker quick start, project structure, test guide.
  • 3 demo scenarios (demos/): Retail site selection, land governance audit, population analysis — each with sample data generation and headless pipeline execution.
  • Docker quickstart (demos/docker-quickstart.sh): One-command setup script.

Tests

  • test_db_engine.py (new, 5 tests): Singleton behavior, no-DB fallback, pool config verification.
  • Semantic layer tests expanded (31 new → 63 total): Hierarchy matching (child/parent), equivalence matching, cache invalidation, prompt output verification.
  • All test mock targets updated: 6 test files (test_rls.py, test_integration.py, test_semantic_layer.py, test_template_manager.py, test_token_tracker.py, test_audit.py, test_memory.py, test_wecom.py) now mock get_engine instead of create_engine/get_db_connection_url.
  • Total: 538 tests passing.

[v4.0.0-alpha] - 2026-02-24 (Secure & Enhanced Platform)

Added — OBS Cloud Storage (S3-Compatible)

  • OBS Storage Module (obs_storage.py): Huawei Cloud OBS integration via standard S3 protocol (boto3). Thread-safe singleton client, user-scoped S3 keys ({user_id}/{filename}). Shapefile bundle upload/download auto-handles sidecar files (.cpg/.dbf/.prj/.shx/.sbn/.sbx/.shp.xml). Graceful degradation to local-only mode when OBS env vars not configured.
  • Cloud Fallback (gis_processors.py): _resolve_path() now falls back to OBS download when file not found locally. Transparent to all 30+ tools — no signature changes needed.
  • Upload Sync (app.py): Uploaded files automatically synced to OBS after local copy. Tool output files synced to OBS after each function_response. Report exports synced to OBS.
  • Cloud-Aware File Management (agent.py): list_user_files() shows merged local+cloud view with [云端]/[仅本地] tags. delete_user_file() deletes from both local and cloud storage.
  • Startup Health Check: ensure_obs_connection() verifies bucket access at startup with status logging.

Added — Session Persistence

  • DatabaseSessionService (app.py): ADK sessions now persist to PostgreSQL via asyncpg. _create_session_service() attempts DatabaseSessionService with graceful fallback to InMemorySessionService when asyncpg is unavailable or DB unreachable. Tables (sessions, events, app_states, user_states) auto-created by ADK on first use. Page refresh / server restart no longer loses conversation history.
  • Async DB URL (database_tools.py): get_async_db_url() converts sync PostgreSQL URL to postgresql+asyncpg:// scheme for async session driver.

Added — Analysis Explainability

  • Tool Descriptions (app.py): TOOL_DESCRIPTIONS dict (40+ entries) maps each tool to a Chinese method name and parameter labels. Displayed in Chainlit Step.input during tool execution.
  • Human-readable Formatting (app.py): _format_tool_explanation() converts raw tool args dict into formatted Chinese text with method name header and labeled parameters. File paths shortened to basename; long values truncated.
  • Smart Result Display (app.py): function_response handler now extracts meaningful output (file paths, status messages) instead of showing generic "执行成功" for all tools.

Added — Dynamic Planner

  • Dynamic Planner (agent.py): LlmAgent-based orchestrator replaces fixed SequentialAgent pipelines. Uses ADK transfer_to_agent for dynamic sub-agent delegation. 5 specialist agents: Explorer (data quality/DB), Processor (GIS ops/geocoding), Analyzer (FFI/DRL), Visualizer (maps/charts), Reporter (reports). Handles composite, conditional, and cross-pipeline requests. Feature flag DYNAMIC_PLANNER (default true); set to false for legacy mode. Safety: RunConfig(max_llm_calls=50).

Added — Plan Mode, Quality Gate, Model Tiering

  • Plan Mode Confirmation (app.py): OPTIMIZATION/GOVERNANCE intents trigger a lightweight analysis plan (via Gemini 2.0 Flash) shown to user with confirm/modify/cancel buttons before execution. Approved plan injected as [分析方案] block into agent prompt. New plan_generation_prompt in prompts.yaml.
  • Quality Gate (agent.py): _quality_gate_check() validates tool output files post-execution. Checks: SHP feature count + CRS, CSV data rows, HTML/PNG minimum size, zero-byte detection. Critical failures trigger auto-retry via _self_correction_after_tool; warnings annotated and passed through.
  • Model Tiering (agent.py): Three-tier model strategy for Planner agents. MODEL_FAST (gemini-2.0-flash) for Explorer/Visualizer, MODEL_STANDARD (gemini-2.5-flash) for Processor/Analyzer/Planner root, MODEL_PREMIUM (gemini-2.5-pro) for Reporter. Legacy pipeline agents unchanged.

Added — Security & Multi-tenancy

  • User Authentication: Chainlit password login + Google OAuth2 (conditional on env vars).
  • Per-user File Sandbox: All uploads and outputs scoped to uploads/{user_id}/.
  • RBAC: Three roles (admin/analyst/viewer). Viewers blocked from Governance & Optimization pipelines.
  • Database User Context: Injects SET app.current_user before SQL queries (RLS-ready).
  • Auth Module (auth.py): Password hashing (PBKDF2-SHA256), app_users table, auto-seeded admin (admin/admin123).
  • User Context Module (user_context.py): ContextVar-based identity propagation across async tool chains.
  • SQL Migration: migrations/001_create_users.sql for user management table.
  • Row-Level Security (RLS): PostgreSQL RLS policies enforce data isolation at the database level.
    • user_memories table: users can only read/write their own memories; admin sees all.
    • token_usage table: users can only see their own consumption records; admin sees all.
    • table_ownership registry: tracks ownership of dynamically-imported PostGIS tables. Users see own + shared tables; admin sees all.
    • _inject_user_context(conn) helper: injects both app.current_user and app.current_user_role via set_config() before every DML operation across database_tools.py, memory.py, and token_tracker.py.
    • _load_spatial_data() fixed: passes conn (with context) to gpd.read_postgis() instead of engine (which opened a new context-less connection).
    • list_tables() / describe_table() rewritten to check table_ownership registry with RLS auto-filtering.
    • register_table_ownership(): UPSERT into registry after PostGIS imports.
    • share_table(): admin-only tool to mark tables as shared (accessible to all users).
    • ensure_table_ownership_table(): startup initializer with superuser/bypassrls warning.
    • import_shp_to_pg.py: --owner and --shared CLI args; auto-registers ownership after import.
    • SQL Migration: migrations/004_enable_rls.sql — RLS policies, FORCE ROW LEVEL SECURITY, pre-seeds existing spatial tables as admin-shared.

Added — Data Quality & Import

  • Open-Source GIS Tools (gis_processors.py): 5 new tools replacing ArcPy equivalents, built on GeoPandas/Shapely/Pandas:
    • polygon_neighbors: Polygon adjacency analysis — detects neighboring polygons via spatial index, outputs CSV with shared boundary length and node count. Auto-projects to planar CRS for accurate measurement.
    • add_field: Add new attribute field with type (TEXT/FLOAT/INTEGER/DOUBLE) and optional default value.
    • add_join: Attribute table join (left join) by common field. Supports CSV and Shapefile as join source.
    • calculate_field: Field expression calculator using !field! ArcGIS syntax (converted to safe df.eval()). Supports arithmetic and field references.
    • summary_statistics: GroupBy aggregation with ArcGIS-style stats format ("field SUM;field MEAN"). Supports SUM/MEAN/MIN/MAX/COUNT/STD/FIRST/LAST with optional case field grouping.
  • Data Quality Pre-check: describe_geodataframe() expanded from 1 check (CRS) to 7 checks:
    • Null values per column, null/empty geometries, anomalous coords (0,0), out-of-bounds coords, duplicate geometries, mixed geometry types, attribute statistics.
    • Returns severity levels: pass / warning / critical.
  • Excel Import: .xlsx / .xls with auto-detection of coordinate columns (same as CSV).
  • KML/KMZ Import: Direct KML reading via pyogrio; KMZ auto-extraction of contained .kml.
  • ZIP Enhancement: ZIP extraction now searches for .kml, .geojson, .gpkg (fallback after .shp).

Added — Visualization

  • Choropleth Map (generate_choropleth): Value-based polygon coloring with 3 classification methods (quantile, equal_interval, natural_breaks), 7 color schemes, floating legend.
  • Bubble Map (generate_bubble_map): Size-scaled circle markers for point data with optional color ramp. Supports same 7 color schemes and 4 basemaps.
  • Basemap Switching: All interactive maps now offer 4 basemaps: CartoDB positron, OpenStreetMap, CartoDB Dark, Gaode (高德).
  • Heatmap / KDE (generate_heatmap): Kernel Density Estimation raster heatmap via scipy.
  • Spatial Distance Query (find_within_distance): Find features within/beyond a given distance.

Added — Geocoding & Public Data

  • Amap (高德) Geocoding: Primary geocoding via Gaode Maps API with Nominatim fallback.
  • Reverse Geocoding (reverse_geocode): Coordinates-to-address conversion via Amap regeocode API. Adds province/city/district columns.
  • POI Nearby Search (search_nearby_poi): Search points of interest (banks, hospitals, restaurants, etc.) within a specified radius around any coordinate. Returns Shapefile with name, type, address, distance.
  • POI Keyword Search (search_poi_by_keyword): Search POIs by keyword within a city or district (e.g., "find all Starbucks in Beijing"). Returns Shapefile with full POI attributes.
  • Admin Boundary (get_admin_boundary): Download administrative district boundary polygons as Shapefile. Supports optional with_sub_districts flag to include child district boundaries (handles municipalities/直辖市 hierarchy automatically).

Added — UX & Safety

  • Real-time Progress Feedback: Pipeline execution now shows hierarchical progress:
    • Pipeline-level: Top-level step showing pipeline name and elapsed time.
    • Agent-level: Stage indicators (e.g., "阶段 2/6: 正在数据质量审计...") tracking SequentialAgent transitions via event.author.
    • Tool-level: Chinese status labels (e.g., "正在批量地理编码..." → "批量地理编码 ✓ (2.3s)") for all 30+ tool functions.
  • Multi-turn Dialogue Context: Users can reference prior results ("上面的结果", "刚才的数据") across conversation turns.
    • After each turn, generated file paths + analysis summary stored in session.
    • Next turn's prompt automatically injected with [上轮分析上下文] block.
    • Intent router uses previous pipeline as fallback when user references prior context.
  • PostGIS Direct Visualization: All 7 visualization tools (generate_choropleth, visualize_interactive_map, etc.) now accept PostGIS table names directly as file_path. _load_spatial_data() auto-detects table names (no extension + alphanumeric pattern) and reads via gpd.read_postgis(). Eliminates the need to export SHP before visualization.
  • File Size Limit: Uploads exceeding 100 MB are rejected with user notification.
  • Upload Data Preview: Immediate markdown preview after file upload showing feature count, CRS, geometry type, bounding box, and first 5 rows.
  • Map Export PNG (export_map_png): High-DPI static map image export with contextily basemap, optional value column coloring and title.
  • Ambiguous Intent Clarification: When the intent router cannot confidently classify a request, it prompts the user to select the desired pipeline via interactive action buttons.
  • User Data Management: list_user_files and delete_user_file tools allow users to view and clean up their uploaded/generated files within their sandbox.
  • Natural Language Map Control: visualize_interactive_map and generate_choropleth now accept optional center_lat, center_lng, zoom parameters. Agents can translate location names to coordinates for map positioning.
  • Tianditu Basemap (天地图): All interactive maps support Tianditu vector + label layers when TIANDITU_TOKEN env var is configured. Basemap code unified via _add_basemap_layers() helper.
  • Driving Distance (calculate_driving_distance): Two-point driving distance/time via Amap Driving Route API, with Haversine straight-line distance as reference.
  • Method Statement: All three Summary agents (GeneralSummary, DataSummary, GovernanceReporter) now mandatorily include a "分析方法" paragraph with method name, parameters, and data source.
  • Self-Correction (after_tool_callback): When a tool returns an error, contextual hints are automatically appended (e.g., "please call describe_table to get real column names"). Retry limit of 3 per tool per invocation prevents infinite loops. Applied to 5 key agents.
  • Spatial Memory System (memory.py): Persistent per-user memory stored in PostgreSQL (user_memories table with JSONB). Supports 4 memory types: region (常用区域), viz_preference (可视化偏好), analysis_result (历史分析), custom. Auto-saves analysis results after each pipeline run. Users can save/recall/list/delete memories via natural language. Visualization preferences automatically injected into agent prompts as defaults.
  • Enhanced Report Generation: Professional Word/PDF reports with page headers/footers, page numbers, pipeline-specific titles (优化/治理/通用), author metadata, date stamps, and styled heading hierarchy. Export button now available for all three pipelines (previously only Optimization/Governance). PDF export via docx2pdf with graceful fallback to Word if conversion unavailable.
  • Token Usage Tracking (token_tracker.py): Per-user LLM token consumption tracking stored in PostgreSQL (token_usage table). Records input/output tokens per pipeline run. Daily analysis limit (configurable via DAILY_ANALYSIS_LIMIT env var, default 20/day, admin exempt). Monthly token limit support (MONTHLY_TOKEN_LIMIT env var, default unlimited). Usage summary queryable via natural language (e.g., "我今天用了多少token"). Tokens accumulated from ADK event.usage_metadata and intent router response.usage_metadata.
  • ArcPy Integration (arcpy_tools.py + arcpy_worker.py): Optional parallel GIS engine via persistent subprocess bridge. 8 tools: arcpy_buffer, arcpy_clip, arcpy_dissolve (ArcPy-unique: field dissolve with multi-stat aggregation), arcpy_project, arcpy_check_geometry (enhanced geometry validation), arcpy_repair_geometry (ArcPy-unique: auto-fix invalid geometries), arcpy_slope, arcpy_zonal_statistics. Enabled via ARCPY_PYTHON_EXE env var pointing to ArcGIS Pro Python. JSON-line IPC protocol over stdin/stdout with persistent worker process (avoids ~3-5s ArcPy cold start). Graceful degradation when ArcPy unavailable.

Changed

  • Path Consolidation: Removed duplicate _generate_output_path / _resolve_path from agent.py; unified on gis_processors.py versions (user-scoped).
  • File Path Regex: extract_file_paths() now recognizes .xlsx, .xls, .kml, .kmz, .geojson, .gpkg.
  • Prompts: Added choropleth, bubble map, and reverse geocoding guidance to agent instructions.

Fixed

  • Chainlit @cl.oauth_callback made conditional to avoid crash when no OAuth provider configured.

---## [v3.2.0] - 2026-02-23 (Semantic Analyst)

Added

  • Semantic Intent Router: Replaced keyword matching with LLM-based intent classification (Governance vs Optimization vs General).
  • Spatial Semantic Layer: describe_table tool allows agents to inspect DB schema and map natural language queries to actual columns.
  • Business Analysis Suite:
    • perform_clustering: DBSCAN clustering.
    • create_buffer: Buffer zone analysis.
    • overlay_difference: Erase/Difference analysis.
    • summarize_within: Zonal statistics.
  • Visualization: Support for Heatmaps (L.heatLayer) and Clustered Markers (L.circleMarker) in visualize_interactive_map.

Changed

  • Architecture: GeneralPipeline prompt hardened to support complex reasoning chains (e.g., Site Selection).
  • UX: Export button is now hidden for simple General queries to reduce noise.
  • Database: list_tables now identifies Spatial tables.

Fixed

  • Fixed GeneralViz crashing on empty input.
  • Fixed session argument error in app.py.

[v3.1.0] - 2026-02-20

Added

  • Hard Routing: Split monolithic agent into Governance, Optimization, and General pipelines.
  • PostGIS Integration: Full read/write support for PostgreSQL.

[v2.0.0] - 2025-12-15

  • Excel Geocoding.
  • Word Report Generation.