GeoRescue

Ask GeoRescue

These are recorded tool-call sessions, not a live chatbot. scripts/mcp_demo.py starts the GeoRescue MCP server over stdio and calls its tools exactly as an AI agent would; the outputs below are the real responses from the owner-name-redacted results store, and each answer is filled in by a fixed template from those outputs (no AI model wrote it). All results are automated, experimental data quality test results — proposals, not findings. Recorded 2026-10-26T16:12:51+00:00.

Tools: inspect_dataset · repair_dataset · list_tables · describe_table · run_query · list_metrics · get_metric · get_case · find_issues · explain_issue · compare_runs

Session 1 · recorded

Q: Why is Parcel A flagged?

get_case ✓

arguments

{
 "case_id": "0015420010__0015315300"
}

output

{
 "case": {
  "case_id": "0015420010__0015315300",
  "parcel_a": "001542 0010",
  "parcel_b": "001531 5300",
  "check": "contained_parcel",
  "overlap_sqft": 107875.71,
  "proposal": "cut 001531 5300 from 001542 0010",
  "direction": "cut smaller from larger",
  "confidence": "medium",
  "rationale": "smaller parcel lies ~100% inside larger; stacked/un-carved parent polygon",
  "status": "proposed",
  "orb_refs": {
   "001542 0010": [
    "19870-00841",
    "17966-00196",
    "16276-02117",
    "12781-02349",
    "11954-01118",
    "10417-00883"
   ],
   "001531 5300": [
    "00056-00042",
    "11504-02030"
   ]
  },
  "basis": "deed-checked",
  "run_id": "20261026T160815+0000"
 },
 "deed_check": {
  "case_id": "0015420010__0015315300",
  "parcel_a": "001542 0010",
  "parcel_b": "001531 5300",
  "verdict": "Owner of 001542 0010 GIS polygon over-includes Owner of 001531 5300 001531 5300; cut Owner of 001531 5300 from Owner of 001542 0010 (contained_parcel: cut smaller from larger)",
  "confidence": "high",
  "confidence_score": 0.85,
  "orb": "ORB 19870-841",
  "evidence": [
   "ORB 19870-841 closes 1:353363 (misclosure 0.007 ft), COGO 7.963 ac vs stated 7.83",
   "Record fits GIS-minus-Owner of 001531 5300 (8.11 ac) with IoU 0.981, Hausdorff 6.1 ft, zero rotation/shift",
   "Owner of 001531 5300 inside record tract: 0%; GIS Owner of 001542 0010 10.59 = 8.11 + Owner of 001531 5300 2.48 ac",
   "Deed's 60 ft ingress/egress strip is an easement ('Together with'), not fee - consistent with excluding the northern strip",
   "Parent 17966-196 (3.41 ac, folio 001542-0100) closes 1:294k and lies within Tract 8"
  ],
  "caveat": "no monument/PLSS tie; placement is best-fit to GIS (it lands at 0,0 offset, which is itself corroborating)",
  "closure_precision": "1:353363",
  "closure_ratio": 353363.0,
  "misclosure_ft": 0.007,
  "n_calls": 6.0,
  "perimeter_ft": 2393.44,
  "acres_record_cogo": 7.963,
  "acres_record_stated": 7.83,
  "fit_iou": 0.734,
  "status": "proposed (test result)"
 },
 "issues": [
  {
   "feature_id": "001542 0010",
   "check": "contained_parcel",
   "severity": "error",
   "message": "001531 5300 is 100.0% inside 001542 0010 (107875.71 units^2); proposal: cut smaller from larger|contained_row=154045|other_row=154045",
   "metric": 1.0000000000000002
  }
 ],
 "other_cases_for_parcel": 0,
 "note": "Automated, experimental data quality test result \u2014 a proposal, not a survey, boundary determination or legal advice."
}

Answer (template from tool outputs): Case 0015420010__0015315300 pairs parcel 001542 0010 (Parcel A) with 001531 5300 (Parcel B); check: contained_parcel. First-pass proposal: cut 001531 5300 from 001542 0010 (confidence medium). Deed check (ORB 19870-841): proposed verdict — Owner of 001542 0010 GIS polygon over-includes Owner of 001531 5300 001531 5300; cut Owner of 001531 5300 from Owner of 001542 0010 (contained_parcel: cut smaller from larger) (confidence high); traverse closes 1:353363 (misclosure 0.007 ft), best-fit IoU 0.734. 1 issue(s) on these parcels. Proposal only.

Session 2 · recorded

Q: Which overlaps are high confidence?

run_query ✓

arguments

{
 "sql": "SELECT c.case_id, c.\"check\", c.basis, coalesce(v.confidence, c.confidence) AS confidence, c.proposal FROM cases c LEFT JOIN verdicts v USING (case_id) WHERE coalesce(v.confidence, c.confidence) = 'high' ORDER BY 1"
}

output

{
 "columns": [
  "case_id",
  "check",
  "basis",
  "confidence",
  "proposal"
 ],
 "rows": [
  {
   "case_id": "0015420010__0015315300",
   "check": "contained_parcel",
   "basis": "deed-checked",
   "confidence": "high",
   "proposal": "cut 001531 5300 from 001542 0010"
  },
  {
   "case_id": "0130300010__0130300300",
   "check": "contained_parcel",
   "basis": "deed-checked",
   "confidence": "high",
   "proposal": "cut 013030 0300 from 013030 0010"
  }
 ],
 "row_count": 2,
 "truncated": false
}

Answer (template from tool outputs): 2 case file(s) have high confidence: 0015420010__0015315300 (contained_parcel, deed-checked); 0130300010__0130300300 (contained_parcel, deed-checked). Confidence is a heuristic about the data, not a finding.

Session 3 · recorded

Q: What does sliver_polygon mean?

explain_issue ✓

arguments

{
 "check": "sliver_polygon"
}

output

{
 "term": "sliver_polygon",
 "kind": "check",
 "name": "Sliver polygon",
 "what": "A parcel that is extremely small or extremely thin \u2014 a narrow strip rather than a normal lot.",
 "why": "Slivers are frequently leftovers of mismatched lines, though some are real (strips, easements, remnants).",
 "how": "Area below 50 sq ft (Florida East State Plane, US feet) (sliver_max_area) or thinness score below 0.02 (sliver_max_compactness; 4\u03c0\u00b7area \u00f7 perimeter\u00b2, where a circle is 1 and a long thin strip approaches 0).",
 "action": "Proposal only \u2014 reported for review.",
 "causes": "Boundary mismatches between neighbours, remnant strips, right-of-way pieces.",
 "glossary_url": "https://knightchrism.github.io/georescue/glossary.html#sliver_polygon"
}

Answer (template from tool outputs): Sliver polygon: A parcel that is extremely small or extremely thin — a narrow strip rather than a normal lot. Detection: Area below 50 sq ft (Florida East State Plane, US feet) (sliver_max_area) or thinness score below 0.02 (sliver_max_compactness; 4π·area ÷ perimeter², where a circle is 1 and a long thin strip approaches 0). Automatic action: Proposal only — reported for review.

Session 4 · recorded

Q: How many missing owner names are there?

get_metric ✓

arguments

{
 "name": "issues_by_check",
 "group_by": "check"
}

output

{
 "metric": "issues_by_check",
 "label": "Issues by check",
 "unit": "issues",
 "sql": "SELECT \"check\", count(*) AS issues_by_check FROM issues GROUP BY 1 ORDER BY 2 DESC NULLS LAST",
 "columns": [
  "check",
  "issues_by_check"
 ],
 "rows": [
  {
   "check": "gap",
   "issues_by_check": 1090
  },
  {
   "check": "sliver_polygon",
   "issues_by_check": 995
  },
  {
   "check": "attribute_missing",
   "issues_by_check": 271
  },
  {
   "check": "overlap",
   "issues_by_check": 37
  },
  {
   "check": "geometry_invalid",
   "issues_by_check": 14
  },
  {
   "check": "contained_parcel",
   "issues_by_check": 5
  }
 ],
 "row_count": 6,
 "truncated": false,
 "caveats": [
  "Automated, experimental data quality test results on public data; verdicts are proposals, not findings.",
  "Not a survey, boundary determination, title opinion or legal advice."
 ]
}
run_query ✓

arguments

{
 "sql": "SELECT count(*) AS n FROM issues WHERE \"check\" = 'attribute_missing' AND message LIKE 'LNAMEOWNER%'"
}

output

{
 "columns": [
  "n"
 ],
 "rows": [
  {
   "n": 271
  }
 ],
 "row_count": 1,
 "truncated": false
}

Answer (template from tool outputs): 271 attribute_missing issues in the latest run; 271 of them are an empty owner-name field (LNAMEOWNER). Names themselves are never exposed by this server.

Session 5 · recorded

Q: Show me the deed closure for the deed-checked cases

get_metric ✓

arguments

{
 "name": "deed_closure_ratio",
 "group_by": "case_id"
}

output

{
 "metric": "deed_closure_ratio",
 "label": "Deed traverse closure ratio",
 "unit": "1:N (higher is better)",
 "sql": "SELECT \"case_id\", max(closure_ratio) AS deed_closure_ratio FROM verdicts WHERE closure_ratio IS NOT NULL GROUP BY 1 ORDER BY 2 DESC NULLS LAST",
 "columns": [
  "case_id",
  "deed_closure_ratio"
 ],
 "rows": [
  {
   "case_id": "0015420010__0015315300",
   "deed_closure_ratio": 353363.0
  }
 ],
 "row_count": 1,
 "truncated": false,
 "caveats": [
  "Traverses come from OCR'd deed text and best-fit placement; not field measurements.",
  "Automated, experimental data quality test results on public data; verdicts are proposals, not findings.",
  "Not a survey, boundary determination, title opinion or legal advice."
 ]
}

Answer (template from tool outputs): Deed traverse closure ratios (1:N, higher is better): case 0015420010__0015315300: 1:353,363. From OCR'd deed text; not field measurements.

Session 6 · recorded

Q: What changed since the last run?

compare_runs ✓

arguments

{}

output

{
 "before": {
  "run_id": "20261022T210008+0000",
  "finished": "2026-10-22T21:02:52.034572+00:00",
  "features": 408934,
  "issues_total": 2976
 },
 "after": {
  "run_id": "20261026T160815+0000",
  "finished": "2026-10-26T16:12:46.395051+00:00",
  "features": 408950,
  "issues_total": 2412
 },
 "by_check": {
  "attribute_missing": {
   "before": 837,
   "after": 271,
   "change": -566
  },
  "contained_parcel": {
   "before": 5,
   "after": 5,
   "change": 0
  },
  "gap": {
   "before": 1089,
   "after": 1090,
   "change": 1
  },
  "geometry_invalid": {
   "before": 14,
   "after": 14,
   "change": 0
  },
  "overlap": {
   "before": 36,
   "after": 37,
   "change": 1
  },
  "sliver_polygon": {
   "before": 995,
   "after": 995,
   "change": 0
  }
 },
 "changed": {
  "attribute_missing": {
   "before": 837,
   "after": 271,
   "change": -566
  },
  "gap": {
   "before": 1089,
   "after": 1090,
   "change": 1
  },
  "overlap": {
   "before": 36,
   "after": 37,
   "change": 1
  }
 }
}

Answer (template from tool outputs): Comparing run 20261022T210008+0000 with 20261026T160815+0000: issues 2976 -> 2412. Changed checks: attribute_missing 837->271, gap 1089->1090, overlap 36->37

Session 7 · recorded

Q: (Safety) Can an agent delete data?

run_query ✗ refused

arguments

{
 "sql": "DROP TABLE cases"
}

output

{
 "error": "Error executing tool run_query: UnsafeQueryError: only SELECT / WITH queries are allowed"
}

Answer (template from tool outputs): The server refused: Error executing tool run_query: UnsafeQueryError: only SELECT / WITH queries are allowed (read-only by design).

Run it yourself

  1. Install uv and clone the repo, e.g. to C:\Users\knigh\Documents\GitHub\georescue.
  2. In that folder: uv sync --extra verify, then uv run georescue run -c configs/duval_fl_ci.yaml (fetches public COJ parcels; it also builds out/georescue.duckdb) or, if you already have out/latest, just uv run georescue load-db.
  3. Add the server to Claude Desktop (Settings → Developer → Edit Config, claude_desktop_config.json):
    {
 "mcpServers": {
  "georescue": {
   "command": "uv",
   "args": [
    "run",
    "--directory",
    "C:\\Users\\knigh\\Documents\\GitHub\\georescue",
    "--extra",
    "verify",
    "georescue-mcp"
   ],
   "env": {
    "PYTHONPATH": "C:\\Users\\knigh\\Documents\\GitHub\\georescue"
   }
  }
 }
}
  4. or to Cursor (.cursor/mcp.json in the project, or Settings → MCP):
    {
 "mcpServers": {
  "georescue": {
   "command": "uv",
   "args": [
    "run",
    "--directory",
    "C:\\Users\\knigh\\Documents\\GitHub\\georescue",
    "--extra",
    "verify",
    "georescue-mcp"
   ],
   "env": {
    "PYTHONPATH": "C:\\Users\\knigh\\Documents\\GitHub\\georescue"
   }
  }
 }
}
  5. Restart the app and ask e.g. “Why is parcel 001542 0010 flagged?” or “What does sliver_polygon mean?”. Calls are logged to logs/mcp.log.

The store is read-only (SELECT-only guard, row cap, timeout) and owner names are redacted by default. repair_dataset only writes a repaired copy into a new folder. Results are test results, not surveys or legal advice. MCP server pattern adapted from the author's census-pipeline (MIT).