smart-spatial-system 0.2.6__tar.gz → 0.2.7__tar.gz

{smart_spatial_system-0.2.6 → smart_spatial_system-0.2.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: smart_spatial_system
-Version: 0.2.6
+Version: 0.2.7
 Summary: Plugin-based GeoAI backend: natural-language questions become QuerySpecs, DAG plans, and map/table/report outputs.
 Author: Araz Shah
 License-Expression: MIT

{smart_spatial_system-0.2.6 → smart_spatial_system-0.2.7}/orchestrator/planning/llm_spec_generator.py

@@ -1951,6 +1951,62 @@ def _validate_score_features_field_chaining(spec: QuerySpec) -> None:
                 )
 
 
+_FILTER_POINTS_IN_POLYGON_IDENTITY_SIGNALS = [
+    "which zone", "which polygon", "which ring", "which band",
+    "which distance band", "belongs to", "falls into",
+    "group by", "group points", "label each", "keep, for every",
+    "for each point", "for every matched",
+]
+
+
+def _validate_filter_points_in_polygon_usage(spec: QuerySpec) -> None:
+    """
+    Catch filter_points_in_polygon used for a query that actually needs to
+    know WHICH polygon/zone/ring matched.
+
+    filter_points_in_polygon returns nothing but a boolean __in_polygon__
+    membership flag - it has no way to carry the matched polygon's
+    identity. A plan built from a query like "for each point, tell me
+    which zone it falls into" plans and executes without error using this
+    op, but the output carries no zone information at all - a silent
+    failure indistinguishable from success anywhere in the pipeline.
+
+    The _domain_guidance() prompt text alone was not enough to prevent
+    this (confirmed against gpt-4o-mini repeating the exact reported
+    scenario even with the guidance in place), so this is enforced the
+    same way score_features field-chaining and CRS symmetry are: as a
+    hard generation-time check that rejects the plan outright rather than
+    only hoping the model reads the prompt correctly.
+
+    Deliberately keyword-based and conservative on raw_query, mirroring
+    the conservative posture of the other _validate_* checks in this
+    module: a query that doesn't match any identity signal is left
+    unvalidated rather than guessed at, so this can only ever reject a
+    plan for a query that explicitly asked "which zone" (or a synonym)
+    and got a boolean-only op in response.
+    """
+    query_lower = spec.raw_query.lower()
+
+    needs_zone_identity = any(
+        signal in query_lower for signal in _FILTER_POINTS_IN_POLYGON_IDENTITY_SIGNALS
+    )
+
+    if not needs_zone_identity:
+        return
+
+    uses_filter = any(op.op == "filter_points_in_polygon" for op in spec.operations)
+    uses_spatial_join = any(op.op == "spatial_join" for op in spec.operations)
+
+    if uses_filter and not uses_spatial_join:
+        raise LLMSpecGenerationError(
+            "Plan uses filter_points_in_polygon, but the query asks to know "
+            "WHICH specific zone/polygon/ring each point belongs to - "
+            "filter_points_in_polygon only returns a boolean membership flag "
+            "and cannot provide this. Use spatial_join with "
+            "include_target_properties=true instead."
+        )
+
+
 # The two vector-bearing input roles for each distance/nearest-neighbor
 # operation, in (this-layer, other-layer) order - used to check that both
 # were reprojected to the same CRS, not just one of them.
@@ -2120,6 +2176,7 @@ class LLMQuerySpecGenerator:
         _validate_score_features_factor_types(spec)
         _validate_score_features_field_chaining(spec)
         _validate_distance_op_crs_symmetry(spec)
+        _validate_filter_points_in_polygon_usage(spec)
         return spec
 
 

{smart_spatial_system-0.2.6 → smart_spatial_system-0.2.7}/pyproject.toml

@@ -12,7 +12,7 @@ build-backend = "setuptools.build_meta"
 # "Open design questions" for why a namespace migration was deliberately
 # not done here.
 name = "smart_spatial_system"
-version = "0.2.6"
+version = "0.2.7"
 description = "Plugin-based GeoAI backend: natural-language questions become QuerySpecs, DAG plans, and map/table/report outputs."
 readme = "README.md"
 requires-python = ">=3.11"

{smart_spatial_system-0.2.6 → smart_spatial_system-0.2.7}/smart_spatial_system.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: smart_spatial_system
-Version: 0.2.6
+Version: 0.2.7
 Summary: Plugin-based GeoAI backend: natural-language questions become QuerySpecs, DAG plans, and map/table/report outputs.
 Author: Araz Shah
 License-Expression: MIT

{smart_spatial_system-0.2.6 → smart_spatial_system-0.2.7}/tests/test_llm_spec_generator.py

@@ -1459,3 +1459,93 @@ def test_domain_guidance_documents_filter_vs_join_for_zone_identity():
     assert "which zone" in system_prompt.lower()
     assert "spatial_join" in system_prompt
     assert "include_target_properties" in system_prompt
+
+
+# ------------------------------------------------------------------ #
+# Bug 7 (hardened): the prompt guidance alone was not enough - a real
+# second run against gpt-4o-mini repeated the exact reported scenario
+# even with the "0.2.6" guidance text in place. Enforced as a hard
+# generation-time check instead, same as score_features field-chaining
+# and CRS symmetry.
+# ------------------------------------------------------------------ #
+
+def _filter_points_spec_json() -> dict:
+    return {
+        "raw_query": "for each point, tell me which zone it falls into",
+        "goal": "label_points_by_zone",
+        "entities": [
+            {"ref": "points", "kind": "vector"},
+            {"ref": "zones", "kind": "vector"},
+        ],
+        "operations": [
+            {
+                "op": "filter_points_in_polygon",
+                "inputs": {"vector": "points", "polygon": "zones"},
+                "params": {"predicate": "within"},
+                "output": "points_in_zones",
+            }
+        ],
+        "outputs": [
+            {"kind": "vector_layer", "source": "points_in_zones", "config": {}}
+        ],
+    }
+
+
+def _spatial_join_spec_json() -> dict:
+    return {
+        "raw_query": "for each point, tell me which zone it falls into",
+        "goal": "label_points_by_zone",
+        "entities": [
+            {"ref": "points", "kind": "vector"},
+            {"ref": "zones", "kind": "vector"},
+        ],
+        "operations": [
+            {
+                "op": "spatial_join",
+                "inputs": {"source": "points", "target": "zones"},
+                "params": {"predicate": "within", "include_target_properties": True},
+                "output": "points_with_zone",
+            }
+        ],
+        "outputs": [
+            {"kind": "vector_layer", "source": "points_with_zone", "config": {}}
+        ],
+    }
+
+
+def test_generate_raises_when_filter_points_in_polygon_used_for_zone_identity_query():
+    llm_json = _filter_points_spec_json()
+
+    client = StaticLLMClient(json.dumps(llm_json, ensure_ascii=False))
+    generator = LLMQuerySpecGenerator(client)
+
+    with pytest.raises(LLMSpecGenerationError, match="filter_points_in_polygon"):
+        generator.generate("for each point, tell me which zone it falls into")
+
+
+def test_generate_accepts_spatial_join_for_zone_identity_query():
+    llm_json = _spatial_join_spec_json()
+
+    client = StaticLLMClient(json.dumps(llm_json, ensure_ascii=False))
+    generator = LLMQuerySpecGenerator(client)
+
+    spec = generator.generate("for each point, tell me which zone it falls into")
+
+    assert [op.op for op in spec.operations] == ["spatial_join"]
+
+
+def test_generate_does_not_flag_filter_points_in_polygon_for_a_plain_boolean_filter_query():
+    """
+    A query that genuinely only wants a boolean keep/drop filter (no
+    "which zone" signal) must not be rejected just because it happens to
+    use filter_points_in_polygon - that op is exactly correct here.
+    """
+    llm_json = _filter_points_spec_json()
+    llm_json["raw_query"] = "keep only points inside the permitted area"
+
+    client = StaticLLMClient(json.dumps(llm_json, ensure_ascii=False))
+    generator = LLMQuerySpecGenerator(client)
+
+    spec = generator.generate("keep only points inside the permitted area")
+
+    assert [op.op for op in spec.operations] == ["filter_points_in_polygon"]