diff --git a/pageindex/filesystem/core.py b/pageindex/filesystem/core.py
index eec0ade..c9a5d70 100644
--- a/pageindex/filesystem/core.py
+++ b/pageindex/filesystem/core.py
@@ -249,33 +249,29 @@ class PageIndexFileSystem:
         """Attach semantic retrieval to already-built projection indexes.
 
         Register-time generation owns building the index files. Opening an
-        existing workspace should still expose semantic browse, without forcing
-        a re-register step.
+        existing workspace should still expose semantic retrieval when the
+        configured embedding dimensions match the existing index.
         """
         if self.semantic_retrieval_backend is not None:
             return bool(self.semantic_retrieval_channels())
         index_config = self._existing_projection_index_config()
         if index_config is None:
             return False
-        metadata = dict(index_config.get("metadata") or {})
-        embedding_provider = str(
-            metadata.get("embedding_provider")
-            or self.summary_projection_embedding_provider
-        )
-        embedding_model = str(
-            metadata.get("embedding_model")
-            or self.summary_projection_embedding_model
-        )
-        embedding_dimensions = int(
-            metadata.get("embedding_dimensions")
-            or index_config.get("dimension")
-            or self.summary_projection_embedding_dimensions
-        )
+        existing_dimension = int(index_config.get("dimension") or 0)
+        if existing_dimension != self.summary_projection_embedding_dimensions:
+            raise RuntimeError(
+                "summary projection index dimension mismatch: "
+                f"{index_config.get('db_path') or self.summary_projection_index_dir} "
+                f"was built with dimension {existing_dimension}, but configured "
+                "summary_projection_embedding_dimensions is "
+                f"{self.summary_projection_embedding_dimensions}. Rebuild the "
+                "projection index or use a matching embedding configuration."
+            )
         self.configure_hybrid_projection_retrieval(
             self.summary_projection_index_dir,
-            embedding_provider=embedding_provider,
-            embedding_model=embedding_model,
-            embedding_dimensions=embedding_dimensions,
+            embedding_provider=self.summary_projection_embedding_provider,
+            embedding_model=self.summary_projection_embedding_model,
+            embedding_dimensions=self.summary_projection_embedding_dimensions,
             embedding_timeout=self.summary_projection_embedding_timeout,
         )
         return bool(self.semantic_retrieval_channels())
diff --git a/tests/test_pageindex_filesystem_scope.py b/tests/test_pageindex_filesystem_scope.py
index 9edf647..46b1161 100644
--- a/tests/test_pageindex_filesystem_scope.py
+++ b/tests/test_pageindex_filesystem_scope.py
@@ -786,7 +786,75 @@ def test_grep_source_file_requires_terms_on_same_line(tmp_path):
     assert "alpha evidence" in matched["data"]["data"][0]["text"]
 
 
-def test_existing_summary_projection_index_configures_retrieval_backend(tmp_path, monkeypatch):
+def test_existing_summary_projection_index_uses_current_config_when_dimensions_match(
+    tmp_path, monkeypatch
+):
+    from pageindex.filesystem import PageIndexFileSystem
+    from pageindex.filesystem.semantic_index import SemanticIndexRecord, SQLiteVecSemanticIndex
+
+    workspace = tmp_path / "workspace"
+    index_dir = workspace / "artifacts" / "projection_indexes"
+    summary_index = SQLiteVecSemanticIndex(index_dir / "summary_only_vector.sqlite")
+    summary_index.reset(
+        dimension=3,
+        metadata={
+            "channel": "summary",
+            "embedding_provider": "stale-provider",
+            "embedding_model": "stale-embedding",
+            "embedding_dimensions": 3,
+        },
+    )
+    summary_index.upsert_many(
+        [
+            SemanticIndexRecord(
+                file_ref="file_a",
+                external_id="doc_a",
+                source_type="documents",
+                source_path="documents/a.pdf",
+                title="A",
+                text="summary",
+                vector=[1.0, 0.0, 0.0],
+            )
+        ]
+    )
+    filesystem = PageIndexFileSystem(
+        workspace,
+        summary_projection_embedding_provider="current-provider",
+        summary_projection_embedding_model="current-embedding",
+        summary_projection_embedding_dimensions=3,
+        summary_projection_embedding_timeout=12,
+    )
+    calls = []
+
+    def fake_configure(index_dir_arg, **kwargs):
+        calls.append((index_dir_arg, kwargs))
+        filesystem.semantic_retrieval_backend = SummaryBackend("doc_a")
+        return filesystem.semantic_retrieval_backend
+
+    monkeypatch.setattr(
+        filesystem,
+        "configure_hybrid_projection_retrieval",
+        fake_configure,
+    )
+
+    assert filesystem.configure_existing_projection_retrieval() is True
+    assert calls == [
+        (
+            filesystem.summary_projection_index_dir,
+            {
+                "embedding_provider": "current-provider",
+                "embedding_model": "current-embedding",
+                "embedding_dimensions": 3,
+                "embedding_timeout": 12,
+            },
+        )
+    ]
+    assert filesystem.semantic_retrieval_channels() == ("summary",)
+
+
+def test_existing_summary_projection_index_dimension_mismatch_rejects_retrieval(
+    tmp_path, monkeypatch
+):
     from pageindex.filesystem import PageIndexFileSystem
     from pageindex.filesystem.semantic_index import SemanticIndexRecord, SQLiteVecSemanticIndex
 
@@ -816,32 +884,24 @@ def test_existing_summary_projection_index_configures_retrieval_backend(tmp_path
         ]
     )
     filesystem = PageIndexFileSystem(workspace)
-    calls = []
 
-    def fake_configure(index_dir_arg, **kwargs):
-        calls.append((index_dir_arg, kwargs))
-        filesystem.semantic_retrieval_backend = SummaryBackend("doc_a")
-        return filesystem.semantic_retrieval_backend
+    def fail_configure(*args, **kwargs):
+        raise AssertionError("retrieval backend should not be configured on dimension mismatch")
 
     monkeypatch.setattr(
         filesystem,
         "configure_hybrid_projection_retrieval",
-        fake_configure,
+        fail_configure,
     )
 
-    assert filesystem.configure_existing_projection_retrieval() is True
-    assert calls == [
-        (
-            filesystem.summary_projection_index_dir,
-            {
-                "embedding_provider": "openai",
-                "embedding_model": "test-embedding",
-                "embedding_dimensions": 3,
-                "embedding_timeout": 60,
-            },
-        )
-    ]
-    assert filesystem.semantic_retrieval_channels() == ("summary",)
+    with pytest.raises(
+        RuntimeError,
+        match=(
+            "summary projection index dimension mismatch: .*"
+            "dimension 3.*summary_projection_embedding_dimensions is 1024.*Rebuild"
+        ),
+    ):
+        filesystem.configure_existing_projection_retrieval()
 
 
 def test_default_semantic_search_uses_summary_projection_when_only_summary_available(tmp_path):