From f44b7a01f5b854f47c1594a1ab5f72096f736262 Mon Sep 17 00:00:00 2001
From: Timothy Jaeryang Baek <tim@openwebui.com>
Date: Fri, 17 Apr 2026 13:59:46 +0900
Subject: [PATCH] refac

---
 backend/open_webui/retrieval/utils.py | 25 +++++++++++++++++++------
 1 file changed, 19 insertions(+), 6 deletions(-)

diff --git a/backend/open_webui/retrieval/utils.py b/backend/open_webui/retrieval/utils.py
index 721d7263a4..07c1ae5210 100644
--- a/backend/open_webui/retrieval/utils.py
+++ b/backend/open_webui/retrieval/utils.py
@@ -939,12 +939,16 @@ async def filter_accessible_collections(
 ) -> set[str]:
     """
     Return only the collection names the user is allowed to access.
-    Admins bypass all checks.  For non-admins:
+    Admins bypass all checks.  For non-admins the policy is:
+
       - file-*          → validated via has_access_to_file
       - user-memory-*   → must match user's own memory collection
-      - knowledge-bases → always denied (meta-collection)
-      - known KB ids    → validated via Knowledges.check_access_by_user_id
-      - everything else → allowed (ephemeral collections like web-search-*)
+      - web-search-*    → ephemeral per-query collections, always allowed
+      - knowledge-bases → always denied (system meta-collection)
+      - everything else → if the name matches a knowledge base, validated
+                          via Knowledges.check_access_by_user_id; if no
+                          such KB exists, the name is treated as an
+                          ephemeral/legacy collection and allowed
     """
     if user.role == 'admin':
         return collection_names
@@ -952,6 +956,7 @@ async def filter_accessible_collections(
     validated = set()
     for name in collection_names:
         if name == 'knowledge-bases':
+            # System meta-collection — never exposed to non-admins.
             continue
         elif name.startswith('file-'):
             file_id = name[len('file-'):]
@@ -960,12 +965,20 @@ async def filter_accessible_collections(
         elif name.startswith('user-memory-'):
             if name == f'user-memory-{user.id}':
                 validated.add(name)
+        elif name.startswith('web-search-'):
+            # Ephemeral collections created by process_web_search — safe
+            # to allow because they contain only transient web-search
+            # results scoped to the requesting user's session.
+            validated.add(name)
         else:
-            # May be a knowledge-base ID or an ephemeral collection
+            # May be a knowledge-base ID or a legacy/ephemeral collection.
+            # If it IS a KB, enforce access control.  If no such KB
+            # exists, treat it as a non-sensitive collection (e.g. legacy
+            # model knowledge, process_text SHA256 collections) and allow.
             if await Knowledges.check_access_by_user_id(name, user.id, permission=access_type):
                 validated.add(name)
             elif not await Knowledges.get_knowledge_by_id(name):
-                # Not a KB at all — ephemeral collection (e.g. web-search-*), allow
+                # Not a KB at all — legacy/ephemeral collection, allow
                 validated.add(name)
     return validated