diff --git a/backend/open_webui/migrations/versions/b9f3c1d7a2e4_add_file_pending_knowledge_index.py b/backend/open_webui/migrations/versions/b9f3c1d7a2e4_add_file_pending_knowledge_index.py
new file mode 100644
index 0000000000..6d83564d76
--- /dev/null
+++ b/backend/open_webui/migrations/versions/b9f3c1d7a2e4_add_file_pending_knowledge_index.py
@@ -0,0 +1,55 @@
+"""Add partial index for pending knowledge files
+
+Adds a partial expression index supporting
+Files.get_pending_files_for_knowledge, whose query filters on
+file.meta->data->knowledge_id and file.data->status. Without it, every
+call (driven by the knowledge-base pending-files polling) performs a full
+sequential scan of the file table, which is expensive because file.data
+stores full extracted document content. The index keeps the lookup cost
+proportional to the number of pending files rather than total content size.
+
+Revision ID: b9f3c1d7a2e4
+Revises: 461111b60977
+Create Date: 2026-06-04 12:35:00.000000
+
+"""
+
+import sqlalchemy as sa
+from alembic import op
+
+revision = 'b9f3c1d7a2e4'
+down_revision = '461111b60977'
+branch_labels = None
+depends_on = None
+
+INDEX_NAME = 'idx_file_pending_knowledge'
+
+
+def upgrade():
+    conn = op.get_bind()
+    dialect = conn.dialect.name
+    existing_indexes = {idx['name'] for idx in sa.inspect(conn).get_indexes('file')}
+
+    if INDEX_NAME in existing_indexes:
+        return
+
+    if dialect == 'postgresql':
+        op.execute(
+            f'CREATE INDEX {INDEX_NAME} ON file '
+            "((meta -> 'data' ->> 'knowledge_id')) "
+            "WHERE (data ->> 'status') IN ('pending', 'processing')"
+        )
+    elif dialect == 'sqlite':
+        op.execute(
+            f'CREATE INDEX {INDEX_NAME} ON file '
+            "(json_extract(meta, '$.data.knowledge_id')) "
+            "WHERE json_extract(data, '$.status') IN ('pending', 'processing')"
+        )
+
+
+def downgrade():
+    conn = op.get_bind()
+    existing_indexes = {idx['name'] for idx in sa.inspect(conn).get_indexes('file')}
+
+    if INDEX_NAME in existing_indexes:
+        op.drop_index(INDEX_NAME, table_name='file')