fix(graphrag): drop entities whose name normalizes to nothing

``canonical_name`` answers "" for a punctuation-only name, which callers are
meant to read as "no entity" -- ``_resolve_endpoint`` already does. Entity
extraction did not, and nodes merge on that key, so every such entity in a
source collapsed onto one shared node that belonged to none of them.

docsgpt/graphrag/extraction.py

@@ -424,7 +424,7 @@ def extract_graph_for_source(
 
 
 def _build_entities(raw_entities: Any) -> List[Dict[str, Any]]:
-    """Normalize the LLM's entity dicts (drop nameless ones)."""
+    """Normalize the LLM's entity dicts (drop the ones with no usable name)."""
     entities = []
     for e in raw_entities:
         if not isinstance(e, dict):
@@ -432,10 +432,16 @@ def _build_entities(raw_entities: Any) -> List[Dict[str, Any]]:
         name = str(e.get("name", "")).strip()
         if not name:
             continue
+        normalized_name = normalize_entity_name(name)
+        if not normalized_name:
+            # A punctuation-only name normalizes to nothing, and nodes merge on
+            # that key: keeping it collapses every such entity onto one shared
+            # node. The relationship side already drops them.
+            continue
         entities.append(
             {
                 "name": name,
-                "normalized_name": normalize_entity_name(name),
+                "normalized_name": normalized_name,
                 "type": str(e.get("type") or "") or None,
                 "description": str(e.get("description") or "") or None,
             }

tests/graphrag/test_extraction.py

@@ -1085,6 +1085,27 @@ class TestSummaryCountFailure:
         assert store.count_nodes.call_args.kwargs.get("strict") is True
 
 
+@pytest.mark.unit
+class TestEntityNormalization:
+    """An entity whose name normalizes to nothing is not an entity.
+
+    ``canonical_name`` answers "" for a punctuation-only name, and nodes are
+    merged on that key, so keeping them collapses every such entity onto one
+    shared node. ``_resolve_endpoint`` already drops them on the relationship
+    side.
+    """
+
+    @pytest.mark.parametrize("name", ["!!!", "--", "?", "  ***  "])
+    def test_a_name_that_normalizes_to_nothing_is_dropped(self, name):
+        assert extraction_module._build_entities([{"name": name}]) == []
+
+    def test_real_names_survive(self):
+        built = extraction_module._build_entities(
+            [{"name": "Quill Store"}, {"name": "!!!"}, {"name": "Alder"}]
+        )
+        assert [e["normalized_name"] for e in built] == ["quill store", "alder"]
+
+
 @pytest.mark.unit
 class TestParsing:
     def test_parses_embedded_json(self):