Smart Spatial System

Ask a geospatial question in plain language and get map layers, tables, reports and files back.

s3geo.com — project site, plugin catalog and case-study index.

Used in research: five case studies have been built on s3geo so far (Vienna, Tehran, İstanbul, Urmia, Cairo). Most compare an LLM-planned analysis against one written by hand on real OpenStreetMap data; the Cairo study is a pure-plugin raster pipeline (satellite NDVI change, no LLM arm) that found 8 defects fixed in 0.5.7. The İstanbul study is included in full in examples/istanbul_health_access/. See Case studies and papers.

Smart Spatial System is a plugin-based GeoAI backend with a React workbench. A question such as "rank these candidate properties by distance to metro stations, malls and main roads" is turned into a structured QuerySpec, planned as a DAG of spatial operations, executed by plugins against uploaded files or PostGIS, and returned as map-ready outputs with a full execution trace.

Real, unedited terminal output from examples/s3geo_accessibility_demo.py — the same deterministic ranking as accessibility_analysis.py, called entirely through import s3geo (no server, no LLM key needed for this rule-based path). Run it yourself after installing below.

It is the application built on top of geochat-platform: plugins are written with geochat_sdk and executed through geochat_kernel.

Status: published and usable, still refactoring internally. Logic is moving out of orchestrator/ into the layered smart_spatial_system/ package (see docs/ARCHITECTURE_TARGET.md); the orchestrator/*_service.py modules are compatibility shims during that move. The public surface - the CLI, the HTTP API and the documented entry points below - is stable.

Why this, not a general-purpose LLM agent with a GIS tool belt

The natural alternative is: give an LLM function-calling access to some GIS functions and let it decide what to call. This project deliberately doesn't stop there, for reasons that turned out to matter in practice, not just in theory - see smart-spatial-vienna-accessibility's case study and CHANGELOG.md's 0.2.1-0.2.4 entries for five real bugs this surfaced and fixed:

• A deterministic path exists alongside the LLM path, built from the same operation catalog (OP_CATALOG). The same analysis can be run by rule and by LLM, so "is the LLM's plan reliable" is a measurable question (Plan Agreement Rate, Rank Stability - see docs/CASE_STUDIES.md), not a leap of faith.
• Every LLM-generated plan is validated before execution, not just before syntax errors: missing input roles, an unchained multi-factor scoring step, an asymmetric CRS reprojection, an untyped scoring factor - each is a distinct, previously-silent failure mode this system now catches and raises a specific error for, before spending a single second executing.
• Every plugin carries a full execution trace - which operations ran, in what order, with what parameters - so a wrong answer is debuggable, not a black box.

How a query runs

natural-language question
  → QuerySpec            LLM (OpenAI-compatible) or rule-based, with PostGIS semantic context
  → DeterministicPlanner + OP_CATALOG
  → DagPlan → DagExecutor
  → CapabilityRegistry   weighted router, learns from user feedback
  → geochat_sdk plugins  vector, raster, PostGIS, reporting, export
  → outputs              map layers · tables · documents (PDF/HTML) · files · trace

Design decisions are recorded as ADRs in docs/: single kernel pipeline, artifact-based responses, a multilingual semantic layer (English and Persian questions, language-neutral concepts inside), and a service-oriented modular backend.

What is in the box

• 37 plugins registered by default (orchestrator/plugin_modules.py::DEFAULT_SAFE_PLUGIN_MODULES): buffer, ring buffer, spatial join, intersection, predicates, dissolve, nearest neighbour, distance, area and perimeter, centroids, CRS transform, geometry validation, attribute statistics, zonal statistics, band math, NDVI and spectral indices, slope/aspect, raster clip/reclassify/threshold/statistics, raster-to-vector, WMS/WFS fetcher, PostGIS connector, feature scoring/enrichment/risk enrichment, real-estate scoring and spatial enrichment, query filter, vector loader, report builder, PDF renderer and data export. Raster uploads load through local_raster_loader, and geocoding_resolver ships but is not registered by default.
• Data sources: raster and vector uploads, CSV tables, WMS, WFS, PostGIS and remote URLs, grouped into projects.
• Workflows: multi-amenity accessibility scoring (rule-based, reproducible), real-estate site ranking with a generated PDF report, and NDVI analysis.
• Learning router: capability weights adjust from user feedback, with reviewable weight proposals.
• Workbench: React + Leaflet UI for queries, step-by-step progress, map layers, inspection, plugin settings and outputs.

Repository layout

api/                     FastAPI app and routers
orchestrator/            query parsing, planning (QuerySpec, OP_CATALOG, DAG), routing, services
smart_spatial_system/    new layered package (application services; other layers being filled in)
s3geo/                   public entry point: s3geo.query() one-call path, plus the planning
                         pipeline classes it wires up, re-exported for finer-grained control
plugins/                 geochat_sdk capability plugins
config/plugins/          per-plugin YAML config (*.example.yaml are the templates)
templates/reports/       report templates (real-estate report)
scripts/sql/             PostGIS views for the Tehran OSM demo
examples/                runnable examples and their sample data, plus a complete case study
                         (examples/istanbul_health_access/)
frontend/                React + Vite workbench
tests/                   pytest suite (~150 modules)
docs/                    architecture, ADRs, API contracts, phase reports

Runtime data (outputs, uploads, projects) is written to var/ by default, or to SMART_SPATIAL_RUNTIME_DIR, and is not committed.

Install

Requires Python 3.11+.

pip install "smart-spatial-system[raster,pdf]"

Extras are optional and independent: raster (rasterio - NDVI, spectral indices, zonal statistics), pdf (weasyprint - PDF reports), postgis (psycopg), dev (pytest, ruff). Leaving one out does not break the install: the affected plugins are simply not registered, and the rest of the system runs normally.

Run the API:

smart-spatial-api serve --port 8000     # http://127.0.0.1:8000/docs

Set SMART_SPATIAL_API_KEY before exposing this beyond localhost. With it unset every endpoint except / and /health is open, and the server logs a warning saying so at startup. See docs/DEPLOYMENT.md.

Use it

Two ways in, both first-class: import the library, or call the HTTP API. Complete runnable versions of all of these are in examples/.

As a library

Score candidate sites by how close they are to the things that matter, with no server and no LLM involved:

from orchestrator.capability_registry import CapabilityRegistry
from orchestrator.planning.dag_executor import DagExecutor
from orchestrator.planning.planner import DeterministicPlanner
from smart_spatial_system.application.services.query_execution.accessibility_query_spec import (
    AmenitySpec, build_accessibility_initial_inputs, build_accessibility_query_spec,
)

amenities = [
    AmenitySpec(ref="metro", distance_field="distance_to_metro_m",
                max_distance_m=800.0, weight=3.0),
    AmenitySpec(ref="schools", distance_field="distance_to_school_m",
                max_distance_m=1200.0, weight=2.0),
]

query_spec = build_accessibility_query_spec(
    "Rank sites by access to metro and schools", amenities,
    target_crs="EPSG:31256",          # a projected CRS - see the note below
)

plan = DeterministicPlanner().build(query_spec)
registry = CapabilityRegistry.from_plugin_modules(tolerant=True)
result = DagExecutor(lambda name: registry.resolve(name).callable).execute(
    plan,
    initial_inputs=build_accessibility_initial_inputs(
        sites=sites_geojson,
        amenity_layers={"metro": metro_geojson, "schools": schools_geojson},
    ),
)

python examples/accessibility_analysis.py runs exactly this over five candidate sites in Vienna and prints the plan and the ranking:

Plan: 10 operations
  sites_metric      transform_vector_crs
  metro_metric      transform_vector_crs
  sites_with_metro  find_nearest_neighbors
  ...

Site accessibility ranking
  Rank  Name                  Accessibility score  Metro (m)  School (m)  Park (m)
  1     Site 3 - Praterstern  75.4                 23.0       407.0       711.0
  2     Site 4 - Ottakring    60.8                 56.0       684.0       4371.0
  3     Site 2 - Karlsplatz   59.8                 128.0      782.0       630.0

This path is rule-based, not LLM-backed: the operation chain follows mechanically from the amenity list, so identical inputs always produce an identical plan and identical numbers.

Distances need a projected CRS. The spatial operations measure planar distance in whatever units the input CRS uses and never reproject on your behalf, so EPSG:4326 input yields degrees. The generator reprojects every layer first; pass a local projected CRS for your study area (EPSG:31256 for Vienna, the relevant UTM zone elsewhere). EPSG:3857 is a safe global fallback but its metres are inflated by 1/cos(latitude) - about 1.5x at Vienna's latitude.

As a library, LLM-planned in one call

The accessibility_analysis.py path above is rule-based and wires the planning pipeline manually - the right level of control for a reusable workflow, but more ceremony than a one-off question needs. s3geo.query() collapses that same pipeline (LLM client, LLMQuerySpecGenerator, DeterministicPlanner, CapabilityRegistry, DagExecutor) into a single call that plans the operation chain from the question itself:

import s3geo

result = s3geo.query(
    "For every station, find amenity points within 300 meters, "
    "reproject to a metric CRS first.",
    layers={"stations": stations_geojson, "amenity": amenity_geojson},
)

result.goal          # str - the LLM-identified analysis goal
result.operations    # list[str] - operation names, in order
result.output        # the final DAG output

layers accepts GeoJSON FeatureCollection dicts or geopandas.GeoDataFrame objects interchangeably. This needs an LLM key configured (LLM_API_KEY / AVALAI_API_KEY / OPENAI_API_KEY) - unlike the rule-based path above, the operation chain is planned by the model, not built mechanically from a fixed amenity list. LLMSpecGenerationError and PlanningError propagate unchanged if the model's plan doesn't pass validation or can't be built into a DAG; a plan that builds but fails during execution raises RuntimeError with the executor's own message. It is a thin wrapper only - every class it wires up stays directly usable for more control (custom context, a different LLM client, inspecting the DAG plan before executing it). python examples/s3geo_quickstart.py runs this over the same Vienna sample data as accessibility_analysis.py.

import s3geo alone reaches that finer-grained control, too. Every class query() wires up internally is re-exported as s3geo.<Name> - the exact same object defined in orchestrator.planning / orchestrator.capability_registry, not a copy (tests/test_s3geo.py asserts this with is identity checks). No need to know the pipeline lives in orchestrator.* to reach it:

import s3geo

registry = s3geo.registry()                      # every plugin loaded, tolerant=True by default
binding = registry.resolve("buffer_vector_features")
binding.callable(...)                             # call a specific plugin capability directly

spec = s3geo.LLMQuerySpecGenerator(s3geo.OpenAICompatibleLLMClient()).generate(
    "buffer the sites by 100 meters",
)
plan = s3geo.DeterministicPlanner().build(spec)   # inspect the plan before executing it
result = s3geo.DagExecutor(s3geo.RegistryCapabilityResolver(registry)).execute(
    plan, initial_inputs={"sites": sites_geojson},
)

Also re-exported: StaticLLMClient (for tests/local runs without a real LLM key), and the errors above - s3geo.LLMSpecGenerationError, s3geo.PlanningError, s3geo.DagExecutionError, s3geo.CapabilityResolutionError.

Over HTTP

import json, urllib.request

body = {"query": "Display the sites on the map", "inputs": {"vector": sites_geojson}}
req = urllib.request.Request("http://127.0.0.1:8000/query",
                             data=json.dumps(body).encode(), method="POST")
req.add_header("Content-Type", "application/json")
req.add_header("X-API-Key", "...")          # when the server requires a key
response = json.loads(urllib.request.urlopen(req).read())

for layer in response["layers"]:
    print(layer["name"], layer["summary"]["feature_count"])

inputs is required and must be an object even when empty - it is where the data the question refers to is passed in, keyed by role (vector, raster, or a named layer). python examples/query_via_http.py runs this against a live server.

Questions are understood in English and Persian. A question is answered by the LLM-backed planner when an LLM key is configured, and by the rule-based paths otherwise.

Running from a checkout

For development on the system itself, or to use the React workbench:

git clone https://github.com/arazshah/smart_spatial_system.git
cd smart_spatial_system

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.lock      # pinned; requirements.txt for latest upstream

cp .env.example .env                  # add your LLM key
uvicorn api.main:app --reload         # http://127.0.0.1:8000/docs

Frontend:

cd frontend
cp .env.example .env
npm install && npm run dev            # http://localhost:5173

Docker Compose (backend + frontend, PostGIS optional):

docker compose up --build

PostGIS demo data (Tehran OpenStreetMap): see data/README.md. Full deployment guide - environment variables, authentication, CORS, PostGIS: docs/DEPLOYMENT.md.

API at a glance

Full request and response contracts are in docs/phase5_query_api_contract.md and the other docs/phase5_* files.

Research and case studies

Looking for a study to build on this system, or to see one already done? docs/CASE_STUDIES.md has a catalog of suggested studies (accessibility, risk-aware siting, vegetation change, zoning compliance, and a reproducibility-methodology replication) plus what's not a good fit yet. If you use this software in published work, see CITATION.cff.

Case studies and papers written with s3geo

Five research studies so far have used s3geo on real data. Four compare an LLM-planned analysis (from a plain-language question) against the same analysis written by hand on real OpenStreetMap data; the fifth is a satellite-raster pipeline with no LLM arm. Each study has its own paper or plan. Four of the five have run, and all four found real defects in this package, now fixed and released.

The first four share one design. Arm 1 is a deterministic plan written by hand. Arm 2 is the same question in plain language, planned by LLMQuerySpecGenerator / s3geo.query() and run N times. The arms are compared with Plan Agreement Rate, variance in the chosen parameters, Set and Rank Stability, and agreement with Arm 1. Vienna's paper/comparison_metric.md defines the metric, and the İstanbul study reuses it unchanged. Cairo has no LLM arm: it is a fixed, hand-designed raster pipeline (§4 of its paper is titled "Using s3geo: what worked, what broke") verified against an independent numpy/rasterio re-implementation instead of against an LLM-planned one.

1. Vienna: district accessibility to metro, schools and parks

Repository: smart-spatial-vienna-accessibility

This is the reference reproducibility study. It runs the same analysis two ways, as a deterministic rule-based QuerySpec and as one planned entirely by LLMQuerySpecGenerator. It measures Plan Agreement Rate, parametric variance and Rank Stability across N repeated LLM runs. It focused on LLM planning reliability.

What it changed here: it found and fixed five correctness bugs (0.2.1–0.2.4):

• a silent domain-wrong default
• an undocumented required input role
• multi-factor scoring that was not chained
• an asymmetric CRS mismatch
• a scoring factor that silently defaulted to the wrong type

2. Tehran: land-use diversity gradient around metro stations

Repository: smart-spatial-tehran-tod-gradient. The paper draft is paper/paper.md in that repository.

This study tests whether land-use diversity around Tehran's 122 metro stations changes systematically with distance, a transit-oriented development (TOD) gradient. It ran two ways:

• a deterministic pipeline with plugin calls chosen by hand;
• a pipeline planned entirely from one plain-language question.

Both ran through the same DeterministicPlanner/DagExecutor engine. This study focused on correctness.

What it changed here: running the LLM arm with a real model (gpt-4o-mini) on real data surfaced three defects, fixed in 0.2.5–0.2.9:

• a missing multi-ring buffer primitive, which is now the ring_buffer_analysis plugin;
• the planner choosing a boolean membership filter (filter_points_in_polygon) where a join that keeps zone identity (spatial_join) was needed;
• a spatial_join cardinality default ("first") that silently dropped matches.

The STRtree-indexed spatial_join engine also came from this study. After the fixes, the LLM-planned output is bit-identical to the hand-written pipeline on every metric.

3. İstanbul: mahalle underserved by hospitals and clinics

Folder in this repository: examples/istanbul_health_access/. It was developed in smart-spatial-istanbul-health-access. The paper is paper/paper.md.

This is the most complete study so far. It is the only one included in full here, with notebooks, paper, results and figures. The data is 964 mahalle and 1,020 hospitals/clinics from OSM:

• Arm 1 is written by hand: EPSG:32635, nearest facility, and underserved mahalle at 1,000, 1,500 and 2,000 m.
• Arm 2 is s3geo.query() run N=20 times. It is given no threshold, CRS or operation list.

Result: at 0.5.6, 20/20 unhinted LLM-planned runs return exactly Arm 1's set of underserved mahalle. Jaccard vs. Arm 1 is 1.0000. All 20 runs use EPSG:32635, which the framework derived from the data's own extent.

What it changed here: every earlier pin, from 0.3.0 to 0.5.5, failed for a specific reason. Each reason was reported in the study's bugs/ or enhancements/ and fixed in 0.4.1–0.5.6:

• raster extras required even for vector queries
• unknown op params passed through silently
• the shape of the where value never shown to the LLM
• explicit None overriding plugin defaults
• a CRS code leaked from a prompt example
• nearest_neighbor sped up about 150× with an STRtree
• a projected CRS derived from the input data
• a max_distance/where validator
• one repair retry

4. Urmia: real-estate suitability ranking

Repository: smart-spatial-urmia-real-estate. The plan is paper/PLAN.md in that repository.

This study reuses this package's shipped real-estate ranking workflow end to end: real_estate_spatial_enrich → real_estate_score → filter_attribute → rank_features → build_report (see orchestrator/planning/op_catalog.py). It runs on real OSM data for Urmia's roads, transit hubs and shopping centers, combined with flood/earthquake/fire risk and allowed-construction zoning. Urmia has no metro, so transit_hubs stands in for its public-transit hubs, the same convention as urmia_real_estate_ranking.py.

Status: scaffolded, not yet run. Both arms are written, and their logic that needs no network or LLM has been verified offline. That repository's paper/PLAN.md lists what is left.

5. Cairo: district-level NDVI change, 2017–2025

Repository: smart-spatial-cairo-vegetation-change. The paper is paper/paper.md in that repository.

This study compares two Sentinel-2 L2A scenes of Greater Cairo (27 Aug 2017, 1 Sep 2025) and ranks 48 districts (qism/markaz, from geoBoundaries via CAPMAS/OCHA) by change in vegetation health and extent. Unlike the other four studies, it has no LLM arm — every analytical step is a fixed, hand-designed chain of s3geo raster and vector plugins:

• NDVI (ndvi_calculator) → ΔNDVI and masks (band_math) → vegetation/ change classes (raster_reclassify) → per-district statistics (zonal_statistics) → a change table (raster_to_vector → centroid_extractor → spatial_join → attribute_statistics).

numpy/rasterio are used only for I/O, regridding and plotting; the result is cross-checked against an independent numpy/rasterio re-computation and agrees to 5 × 10⁻⁵.

Result: mean NDVI across the AOI barely moved (0.1823 → 0.1818), but the vegetated area (NDVI ≥ 0.2) shrank by about 2,000 ha (−6.6%), concentrated on the peri-urban farmland fringe (Waraq, Shubra al-Khayma, Kardasa, Al-Ahram). The ranking is robust to masking water and to using a different metric (change in vegetated share instead of mean ΔNDVI).

What it changed here: running the full pipeline at real resolution surfaced 8 defects in 0.5.6, each with a minimal reproduction (scripts/verify_bugs.py), 7 of them fixed in 0.5.7:

• the local raster loader's output couldn't be consumed by any analysis plugin;
• zonal_statistics(all_touched=True) scanned a zone's bounding box instead of its real polygon;
• raster_to_vector(mode="components") merged neighbouring different classes into one component;
• a dict/affine_transform metadata transform was silently replaced by the pixel-unit default;
• raster_reclassify output metadata carried the input nodata value instead of the output one;
• an O(H²·W) pixel loop shared across every raster plugin (fixed: ~20× faster end to end, 18.5 min → 55 s for the full pipeline);
• a transform-normalization helper crashed on a complete, well-formed dict transform;
• generic top-level installed package names (config, plugins, api, …) — left open, deferred by design.

Re-running the same pipeline on 0.5.7 reproduces every number in the paper exactly, with none of the fixes changing a single published result (the analysis had already worked around every affected option).

Want to write the next one? docs/CASE_STUDIES.md lists suggested studies, and the case-study issue template is where to propose one.

Development

pytest                # full suite
ruff check .          # lint

Contributions are welcome - see CONTRIBUTING.md. Found a security issue? See SECURITY.md rather than opening a public issue.

Author

Araz Shahkarami · araz.me