fix:hybrid retriever: reuse query embedding across retries

datamind/capabilities/kb/providers/hybrid_retriever.py

@@ -145,16 +145,20 @@ async def _retrieve(self, query, *, top_k, filters, version):
             bm25, ids, texts, metas = await self._ensure_lexical_at_version(version)
         k_inner = top_k * self._cm
         stage = "kb.embedding"
+        # Reuse a completed embedding across vector-search retries within this
+        # retrieval only. Failed embedding attempts leave the cache empty.
+        query_vector: list[float] | None = None
 
         async def vector_search():
-            nonlocal stage
-            stage = "kb.embedding"
-            vec = await self._embed.embed_query(query)
+            nonlocal stage, query_vector
+            if query_vector is None:
+                stage = "kb.embedding"
+                query_vector = await self._embed.embed_query(query)
             stage = "kb.vector_query"
             kwargs = {"top_k": k_inner, "where": filters}
             if version is not None:
                 kwargs["version"] = version
-            return await self._store.query(vec, **kwargs)
+            return await self._store.query(query_vector, **kwargs)
 
         def rank_lexical():
             if bm25 is not None and ids:

datamind/tests/test_read_recovery.py

@@ -58,7 +58,7 @@ async def test_transient_failure_falls_back_after_bounded_retries(stage):
         hits = await retriever.aretrieve("needle", filters={"tenant": "a"})
     assert [h.id for h in hits] == ["new"]
     assert hits[0].metadata["degraded"] is True
-    assert embed.calls == 2
+    assert embed.calls == (2 if stage == "embedding" else 1)
     summary = recovery_summary(ctx.extra["recovery_events"])
     assert summary["retries"] == 1 and summary["fallbacks"] == 1
     assert summary["events"][0]["component"] == f"kb.{'embedding' if stage == 'embedding' else 'vector_query'}"