Deploy 8168b12

app/pipeline/nodes/retrieve.py

@@ -308,17 +308,22 @@ def make_retrieve_node(
                 if sibling_count >= _SIBLING_TOTAL_CAP:
                     break
                 doc_id = seed["metadata"]["doc_id"]
-                seed_idx = seed["metadata"].get("chunk_index", -1)
+                seed_idx = seed["metadata"].get("chunk_index")
+                if seed_idx is None:
+                    seed_idx = -1
+
                 siblings = vector_store.fetch_by_doc_id(doc_id, limit=_SIBLING_FETCH_LIMIT)
                 # RC-2 fix: sort by chunk_index so we can prefer adjacent chunks.
-                siblings.sort(key=lambda c: c["metadata"].get("chunk_index", 0))
+                # Guard against chunk_index being actual None in Qdrant payloads.
+                siblings.sort(key=lambda c: c["metadata"].get("chunk_index") or 0)
+
                 # If seed position is known, prefer adjacent indices (±2) first,
                 # then fall through to remaining siblings in document order.
                 if seed_idx >= 0:
                     adjacent = [s for s in siblings
-                                if abs(s["metadata"].get("chunk_index", -999) - seed_idx) <= 2]
+                                if abs((s["metadata"].get("chunk_index") or -999) - seed_idx) <= 2]
                     rest = [s for s in siblings
-                            if abs(s["metadata"].get("chunk_index", -999) - seed_idx) > 2]
+                            if abs((s["metadata"].get("chunk_index") or -999) - seed_idx) > 2]
                     ordered_siblings = adjacent + rest
                 else:
                     ordered_siblings = siblings