dsh-webgis

A WebGIS plugin that lets LLMs truly "see" the geographic world. Built on DeepSeek Harness (DSH), it embeds a full GIS workbench into an AI conversation: models can load map data, understand maps, and run spatial analysis through dialogue — with the ultimate goal of genuinely understanding the 3D world.

Read the map → Operate the map → Understand the 3D world

Highlights

• 🗺️ Conversational GIS — load data, fly to locations, and run spatial analysis, all driven by natural language as the AI calls into a chain of geo-processing tools
• 🚀 Massive data without lag — from million-row local CSVs to database results in the hundreds of thousands of rows, a tiered rendering pipeline keeps the map fluid (see Massive Data Loading)
• 🧮 A GIS toolbox — 36+ AI-callable geo tools: buffers, overlays, kernel density, Moran's I, OD matrices, hex-bin heatmaps …
• 🖌️ Manual workflows too — interactive drawing of points / lines / polygons, layer management, shp / csv / geojson import & export
• 🔌 Multiple data sources in one place — PostGIS, local files, and online map services

---

Quick Start

1. Open DeepSeek Harness and start a new conversation.
2. Pick GIS mode.
3. Choose your task workspace.
4. Ask the model — for example: "Load xxx and show it on the map."

---

Features

Compute: conversational GIS analysis

Once data is loaded, just tell the AI "show this as a hex density heatmap" or "which points fall inside this polygon?" — behind the scenes sit 36+ geo-processing tools that chain together:

• Construct — buffer, centroid, convex hull, bounding box, dissolve, simplify, explode, smooth, grid, Voronoi, OD matrix (origin–destination lines showing flows)
• Overlay — clip, intersect, difference, union
• Query — filter by attribute, spatial join, select by location
• Transform — reprojection between coordinate systems
• Spatial statistics — kernel-density heatmap, average nearest neighbor (ANN), global Moran's I
• Visualization — switch render modes: raw points / planar heatmap / hex-bin heatmap / arc / trips / wall / radial; adjust color, size, and stroke width
• Attribute editing — batch-update field values, assign sequential numbers, add columns
• Layer management — list, remove, show/hide, inspect layer info and feature stats

Every result appears as a new live layer that the next tool can keep processing, forming a complete analysis chain.

Draw: manual drawing & editing

• The layer panel lists all layers (dataset / analysis results / database results / imported), with show-hide and delete support
• The Create button expands a drawing toolbar on demand: points / lines / polygons, Bézier-curve smoothing, and snapping to vertices of existing layers
• Right-click a layer to inspect its attribute table, export geojson / csv / shp, or delete it
• Import shp (.shp/.zip) / csv / geojson (CSV auto-detects a WKT geometry column, or falls back to lon/lat columns)

Maps & data services

• Basemap switcher — vector basemaps (Carto light / dark / Voyager, OpenFreeMap) plus satellite imagery, one click from the bottom-left corner
• Online overlay services — register WMTS / WMS / XYZ services as raster overlays (WMS supports the {bbox-epsg-3857} template)
• Data-source routing — answer first from the data already on screen, and query the configured database only when needed — the AI never guesses

---

Massive Data Loading

Data from a few hundred rows to millions of rows should never bog the UI down. The core idea is to "count first, render in tiers, materialize on demand" instead of shoveling everything into the browser at once.

Huge local CSVs (DuckDB):

• A CSV is loaded once into a DuckDB in-memory table with indexes — 1.68 million rows in ~2.7 s in our tests, with millisecond-level filtering afterwards
• Large files are first shown as a sampled view plus clustering, while the full table stays in memory for later filtering by SQL / conditions / polygon fences
• Built-in memory management (LRU eviction of least-recently-used tables + a memory limit) keeps the browser from OOM-ing

PostGIS databases:

• Read-only queries; a count is issued first to decide whether and how to load the data
• Non-4326 geometries are automatically wrapped in ST_Transform; EWKB geometries are converted to GeoJSON for display
• A statement_timeout and read-only transactions at the connection layer keep slow or misdirected queries from stalling the conversation

Sampling is always transparent: when a large layer shows only a sample, the layer is labeled "N rows total (sampled)", and any statistics derived from the sample are flagged as such — so the AI never mistakes a sample for the full dataset.

Combined with deck.gl's 3D rendering (hex-bin columns, wall extrusions, OD arcs — all GPU-driven), hundreds of thousands of rows stay smooth to zoom, filter, and analyze.

---

Installation & Configuration

Requirements: a DSH install (with the web profile) and pnpm on your PATH.

dsh plugin --profile web add dsh-webgis
# after install, restart with: dsh web, then start a new conversation and pick "GIS mode"

Optional configuration (Settings → Plugins → WebGIS plugin config, or the plugin config file cordis.patch.yml):

• Vision model — vision backend provider / model / baseURL / apiKey (leave unset to use the built-in free fallback)
• PostGIS database — host / port / database / user / password (the password is stored in DSH's credential store, never written to disk in plaintext)
• DuckDB thresholds and memory limit

Example prompts

• "Load this xxx.shp onto the map"
• "Are these points spatially clustered? Show them as a hex heatmap"
• "Filter the points from dataset A that fall inside region B"
• "Where is this place?" (the AI captures a screenshot and "looks" at it to answer)

---

Roadmap: from 2D maps to the 3D world

• Phase 1 (in progress): 2D map vision understanding — the AI reads map content and answers "what is where" and "where is this place"
• Phase 2 (in progress): conversational operation & analysis — moving from "seeing" to "computing", closing the loop of load → analyze → visualize
• Phase 3 (long-term goal): truly understanding the 3D world — from reading 2D symbols toward real 3D spatial comprehension: recognizing buildings and terrain, judging occlusion and spatial relations, reasoning at real-world scale. Today's 3D visuals (hex columns, wall extrusions, OD arcs) are only the beginning; the aim is to give the AI a genuine "spatial worldview" so it can see, reason, and answer within real 3D scenes.

Long-term vision: not merely "describing a picture" — the model truly inhabits spatial cognition, understanding the three-dimensional world behind the map.

Note: DeepSeek's vision models were released only recently and are still iterating quickly, so the vision-related features have not yet been deeply tested.

---

Author: Frank Wang · Feedback: cywanghn@gmail.com