From 9beb5ead4d10854f298bd1b0b281b4c08cf739d1 Mon Sep 17 00:00:00 2001
From: ryan-crabbe-berri <ryan@berri.ai>
Date: Sat, 29 Aug 2026 10:40:39 -0700
Subject: [PATCH] fix(proxy): keep the exact model= DB predicate within the
 type-discipline budget

The where clause now uses the exact name string directly and skips the DB query when the typed search cannot occur in that name, so no new mutable literals are added (LIT002 gate).
---
 litellm/proxy/proxy_server.py                 | 14 +++++++-------
 tests/test_litellm/proxy/test_proxy_server.py | 16 +++++++++++++---
 2 files changed, 20 insertions(+), 10 deletions(-)

diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py
index 6e6f711f468..b398bd7e37e 100644
--- a/litellm/proxy/proxy_server.py
+++ b/litellm/proxy/proxy_server.py
@@ -12785,10 +12785,8 @@ async def _fetch_db_models_for_search(
     filter for `team_public_model_name` instead and keep the DB cost
     bounded by `search`.
     """
-    exact_name_filter: Final[dict[str, Any]] = {} if model_name is None else {"AND": [{"model_name": model_name}]}
     db_where_condition: Final[dict[str, Any]] = {
-        "model_name": {"contains": search_lower, "mode": "insensitive"},
-        **exact_name_filter,
+        "model_name": {"contains": search_lower, "mode": "insensitive"} if model_name is None else model_name
     }
     if db_model_ids_in_router:
         db_where_condition["model_id"] = {"not": {"in": list(db_model_ids_in_router)}}
@@ -12858,9 +12856,10 @@ async def _apply_search_filter_to_models(
             full match set, so the DB fetch is capped at
             ``_SORTED_SEARCH_DB_FETCH_CAP`` instead of one page.
         model_name: Exact ``model_name`` the caller already narrowed
-            ``all_models`` to (``?model=``). The DB query honours it too,
-            otherwise rows from other model groups leak into the result
-            and the count.
+            ``all_models`` to (``?model=``). The DB query matches it
+            exactly instead of the substring, and is skipped when the
+            substring cannot occur in it, otherwise rows from other model
+            groups leak into the result and the count.
 
     Returns:
         Tuple of (filtered_models, total_count). total_count is None if not searching.
@@ -12918,7 +12917,8 @@ async def _apply_search_filter_to_models(
 
     # Query database for additional models with search term
     db_models: list[dict[str, Any]] = []
-    if prisma_client is not None:
+    exact_name_can_match: Final = model_name is None or search_lower in model_name.lower()
+    if prisma_client is not None and exact_name_can_match:
         try:
             db_models, db_models_total_count = await _fetch_db_models_for_search(
                 prisma_client=prisma_client,
diff --git a/tests/test_litellm/proxy/test_proxy_server.py b/tests/test_litellm/proxy/test_proxy_server.py
index fa8d6d8df41..949088ea3ba 100644
--- a/tests/test_litellm/proxy/test_proxy_server.py
+++ b/tests/test_litellm/proxy/test_proxy_server.py
@@ -2149,18 +2149,28 @@ async def test_apply_search_filter_honours_exact_model_name_in_db_query():
         model_name="anthropic-sonnet-5",
     )
     where = prisma_client.db.litellm_proxymodeltable.count.call_args.kwargs["where"]
-    assert where["AND"] == [{"model_name": "anthropic-sonnet-5"}]
-    assert where["model_name"] == {"contains": "sonnet", "mode": "insensitive"}
+    assert where["model_name"] == "anthropic-sonnet-5"
     assert prisma_client.db.litellm_proxymodeltable.find_many.call_args.kwargs["where"] == where
 
     prisma_client.db.litellm_proxymodeltable.count.reset_mock()
+    _, total_count = await _apply_search_filter_to_models(
+        all_models=[],
+        search="opus",
+        prisma_client=prisma_client,
+        proxy_config=proxy_config,
+        model_name="anthropic-sonnet-5",
+    )
+    prisma_client.db.litellm_proxymodeltable.count.assert_not_called()
+    assert total_count == 0
+
     await _apply_search_filter_to_models(
         all_models=[],
         search="sonnet",
         prisma_client=prisma_client,
         proxy_config=proxy_config,
     )
-    assert "AND" not in prisma_client.db.litellm_proxymodeltable.count.call_args.kwargs["where"]
+    where = prisma_client.db.litellm_proxymodeltable.count.call_args.kwargs["where"]
+    assert where["model_name"] == {"contains": "sonnet", "mode": "insensitive"}
 
 
 @pytest.mark.asyncio