fix: bound nested metadata values so one oversized field cannot scale with chunk count

KEYS_TO_EXCLUDE drops the six analyze-result fields we know are large, but it is a
field-name blacklist: any other nested value passes through, is deep-copied onto every
chunk by the text splitter, and is then stringified by process_metadata for storage.

Azure Document Intelligence returns AnalyzeResult through langchain_community's
AzureAIDocumentIntelligenceLoader, which sets `metadata=result.as_dict()`. Six of its
fields are excluded; `keyValuePairs`, `styles`, `languages` and `documents` are not.
With the default prebuilt-layout model those stay small, but
DOCUMENT_INTELLIGENCE_MODEL is operator-settable and a non-layout model populates
`documents` and `keyValuePairs` with boundingRegions - per-word polygons again.

Bound them by item count instead of by name. len() is O(1) on list and dict, so nothing
is serialized just to be measured. A nested value that survives is stringified by
process_metadata anyway, so it was never queryable structured data downstream - dropping
an oversized one loses nothing usable and removes an unbounded per-chunk cost.

The guard is applied in process_metadata as well, so it holds for all twelve vector
store clients that call it rather than only the loader path.

backend/open_webui/retrieval/vector/utils.py

@@ -6,11 +6,21 @@ from open_webui.utils.misc import sanitize_text_for_db
 
 KEYS_TO_EXCLUDE = ['content', 'pages', 'tables', 'paragraphs', 'sections', 'figures']
 
+# A nested metadata value is deep-copied onto every chunk by the text splitter and then
+# stringified by process_metadata before storage, so a single unbounded one costs memory
+# proportional to the chunk count. KEYS_TO_EXCLUDE only catches the field names we know
+# about; this bounds the ones we do not.
+MAX_NESTED_METADATA_ITEMS = 64
+
+
+def _is_unbounded(value: Any) -> bool:
+    # len() is O(1) for list and dict - never serialize a value just to measure it.
+    return isinstance(value, (list, dict)) and len(value) > MAX_NESTED_METADATA_ITEMS
+
 
 def filter_metadata(metadata: dict[str, any]) -> dict[str, any]:
     # Removes large/redundant fields from metadata dict.
-    metadata = {key: value for key, value in metadata.items() if key not in KEYS_TO_EXCLUDE}
-    return metadata
+    return {key: value for key, value in metadata.items() if key not in KEYS_TO_EXCLUDE and not _is_unbounded(value)}
 
 
 def process_metadata(
@@ -21,7 +31,7 @@ def process_metadata(
     result = {}
     for key, value in metadata.items():
         # Skip large fields
-        if key in KEYS_TO_EXCLUDE:
+        if key in KEYS_TO_EXCLUDE or _is_unbounded(value):
             continue
         if value is None:
             continue