Merge b3af5f8c0f into ccd783d907

litellm/caching/caching.py

@@ -109,6 +109,8 @@ class Cache:
         default_in_redis_ttl: float | None = None,
         similarity_threshold: float | None = None,
         supported_call_types: list[CachingSupportedCallTypes] | None = list(DEFAULT_CACHING_SUPPORTED_CALL_TYPES),
+        excluded_call_types: list[CachingSupportedCallTypes] | None = None,
+        call_type_ttl: dict[str, float] | None = None,
         # s3 Bucket, boto3 configuration
         azure_account_url: str | None = None,
         azure_blob_container: str | None = None,
@@ -314,9 +316,17 @@ class Cache:
             litellm.logging_callback_manager.add_litellm_success_callback("cache")
         if "cache" not in litellm._async_success_callback:
             litellm.logging_callback_manager.add_litellm_async_success_callback("cache")
-        self.supported_call_types = (
-            supported_call_types  # default to ["completion", "acompletion", "embedding", "aembedding"]
-        )
+        if excluded_call_types and supported_call_types is not None:
+            excluded_set: Final = set(excluded_call_types)
+            self.supported_call_types: list[CachingSupportedCallTypes] | None = [
+                ct for ct in supported_call_types if ct not in excluded_set
+            ]
+        else:
+            self.supported_call_types = (
+                supported_call_types  # default to ["completion", "acompletion", "embedding", "aembedding"]
+            )
+        self.excluded_call_types: list[CachingSupportedCallTypes] | None = excluded_call_types
+        self.call_type_ttl: dict[str, float] | None = call_type_ttl
         self.type = type
         self.namespace = namespace
         self.redis_flush_size = redis_flush_size
@@ -707,6 +717,18 @@ class Cache:
             print_verbose(f"An exception occurred: {traceback.format_exc()}")
             return None
 
+    def _get_ttl_for_call_type(self, kwargs: Mapping[str, object]) -> float | None:
+        if self.call_type_ttl:
+            call_type: Final = kwargs.get("call_type") or ("embedding" if kwargs.get("embedding") is True else None)
+            if isinstance(call_type, str):
+                if call_type in self.call_type_ttl:
+                    return self.call_type_ttl[call_type]
+                if call_type in ("embedding", "aembedding"):
+                    emb_ttl: Final = self.call_type_ttl.get("embedding") or self.call_type_ttl.get("aembedding")
+                    if emb_ttl is not None:
+                        return emb_ttl
+        return self.ttl
+
     def _add_cache_logic(self, result, **kwargs):
         """
         Common implementation across sync + async add_cache functions
@@ -721,8 +743,9 @@ class Cache:
                     result = result.model_dump_json(exclude_unset=True)
 
                 ## DEFAULT TTL ##
-                if self.ttl is not None:
-                    kwargs["ttl"] = self.ttl
+                resolved_ttl: Final = self._get_ttl_for_call_type(kwargs)
+                if resolved_ttl is not None:
+                    kwargs["ttl"] = resolved_ttl
                 ## Get Cache-Controls ##
                 _cache_kwargs: Final = kwargs.get("cache", None)
                 if isinstance(_cache_kwargs, dict):
@@ -911,6 +934,8 @@ class Cache:
             prompt_tokens_details=prompt_tokens_details,
         )
 
+        if "call_type" not in kwargs and "embedding" not in kwargs:
+            kwargs["embedding"] = True
         cache_key, cached_data, kwargs = self._add_cache_logic(
             result=embedding_dict,
             **kwargs,
@@ -938,8 +963,9 @@ class Cache:
                     return
 
                 # set default ttl if not set
-                if self.ttl is not None:
-                    kwargs["ttl"] = self.ttl
+                pipeline_ttl: Final = self._get_ttl_for_call_type({**kwargs, "embedding": True})
+                if pipeline_ttl is not None:
+                    kwargs["ttl"] = pipeline_ttl
 
                 cache_list: Final = []
                 if isinstance(kwargs["input"], list):

tests/unit/caching/test_caching.py

@@ -1566,3 +1566,66 @@ def test_cache_key_separates_model_ids_unless_a_model_group_owns_the_key() -> No
     assert key_a != bare
     assert key_a == key_a_again
     assert grouped_a == grouped_b
+
+
+def test_cache_excluded_call_types_filters_defaults() -> None:
+    cache: Final = Cache(
+        type=LiteLLMCacheType.LOCAL,
+        excluded_call_types=["embedding", "aembedding"],
+    )
+    assert cache.supported_call_types is not None
+    assert "embedding" not in cache.supported_call_types
+    assert "aembedding" not in cache.supported_call_types
+    assert "completion" in cache.supported_call_types
+    assert "acompletion" in cache.supported_call_types
+
+
+@pytest.mark.asyncio
+async def test_cache_call_type_ttl_overrides_default_ttl_for_embeddings() -> None:
+    from litellm.types.utils import EmbeddingResponse
+
+    cache: Final = Cache(
+        type=LiteLLMCacheType.LOCAL,
+        ttl=3600,
+        call_type_ttl={"embedding": 60},
+    )
+    response: Final = EmbeddingResponse(
+        model="text-embedding-3-small",
+        data=[{"embedding": [0.1, 0.2], "index": 0, "object": "embedding"}],
+    )
+    _, _, updated_kwargs = cache.add_embedding_response_to_cache(
+        result=response,
+        input="hello world",
+        kwargs={"model": "text-embedding-3-small"},
+        idx_in_result_data=0,
+    )
+    assert updated_kwargs["ttl"] == 60
+
+    # Explicit per-request cache={"ttl": ...} still overrides call_type_ttl
+    _, _, override_kwargs = cache.add_embedding_response_to_cache(
+        result=response,
+        input="hello world",
+        kwargs={"model": "text-embedding-3-small", "cache": {"ttl": 15}},
+        idx_in_result_data=0,
+    )
+    assert override_kwargs["ttl"] == 15
+
+    # Pipeline bulk write also forwards call_type_ttl
+    mock_backend: Final = AsyncMock()
+    await cache.async_add_cache_pipeline(
+        result=response,
+        dynamic_cache_object=mock_backend,
+        model="text-embedding-3-small",
+        input=["hello world"],
+    )
+    mock_backend.async_set_cache_pipeline.assert_awaited_once()
+    assert mock_backend.async_set_cache_pipeline.call_args.kwargs["ttl"] == 60
+
+    # Non-embedding calls still use default ttl (3600)
+    _, _, completion_kwargs = cache._add_cache_logic(
+        result="cached-completion",
+        model="gpt-4o-mini",
+        messages=[{"role": "user", "content": "hi"}],
+        call_type="completion",
+    )
+    assert completion_kwargs["ttl"] == 3600