From 14b465f076e532aae655248fb3e25daba09849e5 Mon Sep 17 00:00:00 2001
From: Daniel Chalef <131175+danielchalef@users.noreply.github.com>
Date: Fri, 26 Sep 2025 14:15:54 -0700
Subject: [PATCH] document llm dedupe guardrails

---
 graphiti_core/utils/maintenance/node_operations.py | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/graphiti_core/utils/maintenance/node_operations.py b/graphiti_core/utils/maintenance/node_operations.py
index 122dbc00..2d87d31b 100644
--- a/graphiti_core/utils/maintenance/node_operations.py
+++ b/graphiti_core/utils/maintenance/node_operations.py
@@ -241,7 +241,11 @@ async def _resolve_with_llm(
     previous_episodes: list[EpisodicNode] | None,
     entity_types: dict[str, type[BaseModel]] | None,
 ) -> None:
-    """Escalate unresolved nodes to the dedupe prompt so the LLM can select or reject duplicates."""
+    """Escalate unresolved nodes to the dedupe prompt so the LLM can select or reject duplicates.
+
+    The guardrails below defensively ignore malformed or duplicate LLM responses so the
+    ingestion workflow remains deterministic even when the model misbehaves.
+    """
     if not state.unresolved_indices:
         return