diff --git a/packages/kb-retrieval/src/kb_retrieval/embed.py b/packages/kb-retrieval/src/kb_retrieval/embed.py
index ff278e0..f90e459 100644
--- a/packages/kb-retrieval/src/kb_retrieval/embed.py
+++ b/packages/kb-retrieval/src/kb_retrieval/embed.py
@@ -108,6 +108,11 @@ async def embed_batch(
     indexes a vector that doesn't match the `document_chunk.embedding VECTOR(1024)` column.
     Transport failures propagate as `TRANSIENT_EMBED_ERRORS` -- no built-in retry (see
     `embed_batch_resilient`).
+
+    No backend failover: this is the backfill path, and it must NOT fall back to Ollama@PIHA the
+    way `kb-query`'s `embed_router` does for online queries. ~271k chunks at PIHA's ~790 ms/embed
+    CPU is ~60 h on an 8 GB node shared with Home Assistant, Paperless and kb-postgres. See the
+    module docstring for the full argument -- the two paths are separate by decision.
     """
     kwargs = {}
     if timeout_s is not None: