Add files using upload-large-folder tool

src/ctx/core/graph/vector_index.py

@@ -73,9 +73,9 @@ class NumpyFlatVectorIndex:
         min_score: float,
         exclude_node_ids: set[str] | None = None,
     ) -> list[list[Neighbor]]:
+        queries = _normalize_query_vectors(vectors, expected_dim=self.meta.dim)
         if top_k <= 0 or self.vectors.size == 0:
-            return [[] for _ in range(len(vectors))]
-        queries = _normalize(vectors)
+            return [[] for _ in range(len(queries))]
         scores = queries @ self.vectors.T
         if exclude_node_ids:
             for node_id in exclude_node_ids:
@@ -132,9 +132,9 @@ class HnswlibVectorIndex(NumpyFlatVectorIndex):
         min_score: float,
         exclude_node_ids: set[str] | None = None,
     ) -> list[list[Neighbor]]:
+        queries = _normalize_query_vectors(vectors, expected_dim=self.meta.dim)
         if top_k <= 0 or self.vectors.size == 0:
-            return [[] for _ in range(len(vectors))]
-        queries = _normalize(vectors)
+            return [[] for _ in range(len(queries))]
         extra = len(exclude_node_ids or ())
         k = min(len(self.node_ids), top_k + extra)
         labels, distances = self._hnsw_index.knn_query(queries, k=k)
@@ -224,6 +224,7 @@ def load_vector_index(
             node_ids, content_hashes, vectors = _load_numpy_data(cache_dir / _NUMPY_NAME)
             if (
                 meta.node_count != len(node_ids)
+                or vectors.ndim != 2
                 or meta.dim != int(vectors.shape[1])
                 or meta.node_ids_sha256 != _hash_list(node_ids)
                 or meta.content_hashes_sha256 != _hash_list(content_hashes)
@@ -388,5 +389,15 @@ def _normalize(vectors: np.ndarray) -> np.ndarray:
     return (matrix / norms).astype(np.float32)
 
 
+def _normalize_query_vectors(vectors: np.ndarray, *, expected_dim: int) -> np.ndarray:
+    queries = _normalize(vectors)
+    actual_dim = int(queries.shape[1])
+    if actual_dim != expected_dim:
+        raise ValueError(
+            f"query vector dim {actual_dim} does not match index dim {expected_dim}"
+        )
+    return queries
+
+
 def _hash_list(values: list[str]) -> str:
     return hashlib.sha256("\n".join(values).encode("utf-8")).hexdigest()

src/tests/test_vector_index.py

@@ -65,6 +65,23 @@ def test_numpy_flat_query_excludes_node_ids_and_applies_min_score() -> None:
     assert [n.node_id for n in rows[0]] == ["b"]
 
 
+def test_numpy_flat_query_rejects_wrong_dimension() -> None:
+    index = build_vector_index(
+        kind="numpy-flat",
+        model_id="model-a",
+        node_ids=["a", "b"],
+        content_hashes=["ha", "hb"],
+        vectors=_vectors()[:2],
+    )
+
+    with pytest.raises(ValueError, match="query vector dim 3 does not match index dim 2"):
+        index.query(
+            np.asarray([[1.0, 0.0, 0.0]], dtype=np.float32),
+            top_k=1,
+            min_score=0.0,
+        )
+
+
 def test_numpy_flat_round_trip_validates_model_and_fingerprint(tmp_path) -> None:
     index = build_vector_index(
         kind="numpy-flat",
@@ -109,6 +126,32 @@ def test_numpy_flat_round_trip_validates_model_and_fingerprint(tmp_path) -> None
     )
 
 
+def test_load_vector_index_rejects_corrupt_vector_shape(tmp_path) -> None:
+    index = build_vector_index(
+        kind="numpy-flat",
+        model_id="model-a",
+        node_ids=["a", "b"],
+        content_hashes=["ha", "hb"],
+        vectors=_vectors()[:2],
+    )
+    index.save(tmp_path)
+    np.savez_compressed(
+        tmp_path / "vector-index.numpy.npz",
+        node_ids=np.asarray(["a", "b"], dtype="U"),
+        content_hashes=np.asarray(["ha", "hb"], dtype="U"),
+        vecs=np.asarray([1.0, 0.0], dtype=np.float32),
+    )
+
+    assert (
+        load_vector_index(
+            tmp_path,
+            expected_model_id="model-a",
+            expected_content_fingerprint=index.meta.content_fingerprint,
+        )
+        is None
+    )
+
+
 def test_build_vector_index_rejects_duplicate_node_ids() -> None:
     with pytest.raises(ValueError, match="duplicate node_id"):
         build_vector_index(