From 251412cb72f5c05fe6544cb5d4567ff3d5705d72 Mon Sep 17 00:00:00 2001
From: Inkithai <108919130+Inkithai@users.noreply.github.com>
Date: Sun, 9 Aug 2026 04:01:44 +0000
Subject: [PATCH] fix(vector_store): serialize numpy ndarray embeddings for
 Supabase jsonb upsert

Co-authored-by: arena-agent <297053741+arena-agent@users.noreply.github.com>
---
 backend/tests/test_qa_self_healing.py | 35 +++++++++++++++++++++++++++
 backend/vector_store.py               | 15 +++++++++---
 2 files changed, 47 insertions(+), 3 deletions(-)

diff --git a/backend/tests/test_qa_self_healing.py b/backend/tests/test_qa_self_healing.py
index aa5a879..3a5f788 100644
--- a/backend/tests/test_qa_self_healing.py
+++ b/backend/tests/test_qa_self_healing.py
@@ -207,6 +207,41 @@ def select(self, *cols, **kw):
         importlib.reload(vector_store)  # restore chroma default for other tests
 
 
+def test_supabase_upsert_handles_numpy_arrays():
+    """Ensure numpy ndarray embeddings do not cause TypeError when upserting to Supabase."""
+    os.environ["VECTOR_STORE"] = "supabase"
+    import importlib
+    importlib.reload(vector_store)
+    
+    import numpy as np
+    
+    class FakeSupabaseTable:
+        def __init__(self):
+            self.rows = {}
+        def upsert(self, row, on_conflict=None):
+            self.rows[row["id"]] = row
+            return self
+        def execute(self):
+            return types.SimpleNamespace(data=list(self.rows.values()), count=len(self.rows))
+
+    fake_table = FakeSupabaseTable()
+    
+    try:
+        with mock.patch("db._get_client", lambda: mock.Mock(table=lambda name: fake_table)):
+            emb_array = np.array([0.1, 0.2, 0.3])
+            vector_store.upsert(
+                patient_key="anon_np",
+                ids=["id1"],
+                embeddings=[emb_array],
+                documents=["some doc"],
+                metadatas=[{"source": "test"}]
+            )
+        assert fake_table.rows["id1"]["embedding"] == [0.1, 0.2, 0.3]
+    finally:
+        os.environ.pop("VECTOR_STORE", None)
+        importlib.reload(vector_store)
+
+
 if __name__ == "__main__":
     fns = [v for k, v in sorted(globals().items()) if k.startswith("test_")]
     for fn in fns:
diff --git a/backend/vector_store.py b/backend/vector_store.py
index 7a7c664..275d0d4 100644
--- a/backend/vector_store.py
+++ b/backend/vector_store.py
@@ -151,16 +151,21 @@ def _cosine(a: List[float], b: List[float]) -> float:
         return -1.0
     return dot / (na * nb)
 
-def _supabase_upsert(patient_key: str, ids: List[str], embeddings: List[List[float]], documents: List[str], metadatas: List[Dict[str, Any]]):
+def _supabase_upsert(patient_key: str, ids: List[str], embeddings: List[Any], documents: List[str], metadatas: List[Dict[str, Any]]):
     client = _supabase_client()
     # Upsert per chunk. Supabase doesn't have bulk upsert with jsonb vector easily,
     # so we do per-row upsert. Chunk counts are small (20-30), so fine.
     for i, cid in enumerate(ids):
+        emb = embeddings[i]
+        if hasattr(emb, "tolist"):
+            emb = emb.tolist()
+        elif not isinstance(emb, list):
+            emb = list(emb)
         row = {
             "id": cid,
             "patient_key": patient_key,
             "text": documents[i],
-            "embedding": embeddings[i],  # stored as jsonb array
+            "embedding": emb,  # stored as jsonb array
             "metadata": metadatas[i],
         }
         # postgrest upsert on conflict (id)
@@ -176,7 +181,11 @@ def _supabase_upsert(patient_key: str, ids: List[str], embeddings: List[List[flo
                 raise
     logger.info("Supabase upsert %d chunks for %s", len(ids), patient_key)
 
-def _supabase_query(patient_key: str, query_embedding: List[float], n_results: int) -> Tuple[List[str], List[str], List[Dict[str, Any]]]:
+def _supabase_query(patient_key: str, query_embedding: Any, n_results: int) -> Tuple[List[str], List[str], List[Dict[str, Any]]]:
+    if hasattr(query_embedding, "tolist"):
+        query_embedding = query_embedding.tolist()
+    elif not isinstance(query_embedding, list):
+        query_embedding = list(query_embedding)
     client = _supabase_client()
     try:
         res = client.table("chunks").select("id, text, metadata, embedding").eq("patient_key", patient_key).execute()