Add cohere v2/rerank support (#8421) (#8605)

* Add cohere v2/rerank support (#8421)

* Support v2 endpoint cohere rerank

* Add tests and docs

* Make v1 default if old params used

* Update docs

* Update docs pt 2

* Update tests

* Add e2e test

* Clean up code

* Use inheritence for new config

* Fix linting issues (#8608)

* Fix cohere v2 failing test + linting (#8672)

* Fix test and unused imports

* Fix tests

* fix: fix linting errors

* test: handle tgai instability

* fix: skip service unavailable err

* test: print logs for unstable test

* test: skip unreliable tests

---------

Co-authored-by: vibhavbhat <vibhavb00@gmail.com>

docs/my-website/docs/providers/cohere.md

@@ -108,7 +108,7 @@ response = embedding(
 
 ### Usage
 
-
+LiteLLM supports the v1 and v2 clients for Cohere rerank. By default, the `rerank` endpoint uses the v2 client, but you can specify the v1 client by explicitly calling `v1/rerank`
 
 <Tabs>
 <TabItem value="sdk" label="LiteLLM SDK Usage">

docs/my-website/docs/rerank.md

@@ -111,7 +111,7 @@ curl http://0.0.0.0:4000/rerank \
 
 | Provider    | Link to Usage      |
 |-------------|--------------------|
-| Cohere      |   [Usage](#quick-start)                 |
+| Cohere (v1 + v2 clients)      |   [Usage](#quick-start)                 |
 | Together AI|   [Usage](../docs/providers/togetherai)                 |  
 | Azure AI|   [Usage](../docs/providers/azure_ai)                 |  
 | Jina AI|   [Usage](../docs/providers/jina_ai)                 |  

litellm/__init__.py

@@ -824,6 +824,7 @@ from .llms.predibase.chat.transformation import PredibaseConfig
 from .llms.replicate.chat.transformation import ReplicateConfig
 from .llms.cohere.completion.transformation import CohereTextConfig as CohereConfig
 from .llms.cohere.rerank.transformation import CohereRerankConfig
+from .llms.cohere.rerank_v2.transformation import CohereRerankV2Config
 from .llms.azure_ai.rerank.transformation import AzureAIRerankConfig
 from .llms.infinity.rerank.transformation import InfinityRerankConfig
 from .llms.jina_ai.rerank.transformation import JinaAIRerankConfig

litellm/cost_calculator.py

@@ -855,7 +855,10 @@ def rerank_cost(
 
     try:
         config = ProviderConfigManager.get_provider_rerank_config(
-            model=model, provider=LlmProviders(custom_llm_provider)
+            model=model,
+            api_base=None,
+            present_version_params=[],
+            provider=LlmProviders(custom_llm_provider),
         )
 
         try:

litellm/llms/azure_ai/rerank/transformation.py

@@ -17,7 +17,6 @@ class AzureAIRerankConfig(CohereRerankConfig):
     """
     Azure AI Rerank - Follows the same Spec as Cohere Rerank
     """
-
     def get_complete_url(self, api_base: Optional[str], model: str) -> str:
         if api_base is None:
             raise ValueError(

litellm/llms/base_llm/rerank/transformation.py

@@ -77,6 +77,7 @@ class BaseRerankConfig(ABC):
         rank_fields: Optional[List[str]] = None,
         return_documents: Optional[bool] = True,
         max_chunks_per_doc: Optional[int] = None,
+        max_tokens_per_doc: Optional[int] = None,
     ) -> OptionalRerankParams:
         pass
 

litellm/llms/cohere/rerank/transformation.py

@@ -52,6 +52,7 @@ class CohereRerankConfig(BaseRerankConfig):
         rank_fields: Optional[List[str]] = None,
         return_documents: Optional[bool] = True,
         max_chunks_per_doc: Optional[int] = None,
+        max_tokens_per_doc: Optional[int] = None,
     ) -> OptionalRerankParams:
         """
         Map Cohere rerank params
@@ -147,4 +148,4 @@ class CohereRerankConfig(BaseRerankConfig):
     def get_error_class(
         self, error_message: str, status_code: int, headers: Union[dict, httpx.Headers]
     ) -> BaseLLMException:
-        return CohereError(message=error_message, status_code=status_code)
+        return CohereError(message=error_message, status_code=status_code)

litellm/llms/cohere/rerank_v2/transformation.py

@@ -0,0 +1,80 @@
+from typing import Any, Dict, List, Optional, Union
+
+from litellm.llms.cohere.rerank.transformation import CohereRerankConfig
+from litellm.types.rerank import OptionalRerankParams, RerankRequest
+
+class CohereRerankV2Config(CohereRerankConfig):
+    """
+    Reference: https://docs.cohere.com/v2/reference/rerank
+    """
+
+    def __init__(self) -> None:
+        pass
+
+    def get_complete_url(self, api_base: Optional[str], model: str) -> str:
+        if api_base:
+            # Remove trailing slashes and ensure clean base URL
+            api_base = api_base.rstrip("/")
+            if not api_base.endswith("/v2/rerank"):
+                api_base = f"{api_base}/v2/rerank"
+            return api_base
+        return "https://api.cohere.ai/v2/rerank"
+
+    def get_supported_cohere_rerank_params(self, model: str) -> list:
+        return [
+            "query",
+            "documents",
+            "top_n",
+            "max_tokens_per_doc",
+            "rank_fields",
+            "return_documents",
+        ]
+
+    def map_cohere_rerank_params(
+        self,
+        non_default_params: Optional[dict],
+        model: str,
+        drop_params: bool,
+        query: str,
+        documents: List[Union[str, Dict[str, Any]]],
+        custom_llm_provider: Optional[str] = None,
+        top_n: Optional[int] = None,
+        rank_fields: Optional[List[str]] = None,
+        return_documents: Optional[bool] = True,
+        max_chunks_per_doc: Optional[int] = None,
+        max_tokens_per_doc: Optional[int] = None,
+    ) -> OptionalRerankParams:
+        """
+        Map Cohere rerank params
+
+        No mapping required - returns all supported params
+        """
+        return OptionalRerankParams(
+            query=query,
+            documents=documents,
+            top_n=top_n,
+            rank_fields=rank_fields,
+            return_documents=return_documents,
+            max_tokens_per_doc=max_tokens_per_doc,
+        )
+
+    def transform_rerank_request(
+        self,
+        model: str,
+        optional_rerank_params: OptionalRerankParams,
+        headers: dict,
+    ) -> dict:
+        if "query" not in optional_rerank_params:
+            raise ValueError("query is required for Cohere rerank")
+        if "documents" not in optional_rerank_params:
+            raise ValueError("documents is required for Cohere rerank")
+        rerank_request = RerankRequest(
+            model=model,
+            query=optional_rerank_params["query"],
+            documents=optional_rerank_params["documents"],
+            top_n=optional_rerank_params.get("top_n", None),
+            rank_fields=optional_rerank_params.get("rank_fields", None),
+            return_documents=optional_rerank_params.get("return_documents", None),
+            max_tokens_per_doc=optional_rerank_params.get("max_tokens_per_doc", None),
+        )
+        return rerank_request.model_dump(exclude_none=True)

litellm/llms/custom_httpx/llm_http_handler.py

@@ -710,6 +710,7 @@ class BaseLLMHTTPHandler:
         model: str,
         custom_llm_provider: str,
         logging_obj: LiteLLMLoggingObj,
+        provider_config: BaseRerankConfig,
         optional_rerank_params: OptionalRerankParams,
         timeout: Optional[Union[float, httpx.Timeout]],
         model_response: RerankResponse,
@@ -719,10 +720,7 @@ class BaseLLMHTTPHandler:
         api_base: Optional[str] = None,
         client: Optional[Union[HTTPHandler, AsyncHTTPHandler]] = None,
     ) -> RerankResponse:
-
-        provider_config = ProviderConfigManager.get_provider_rerank_config(
-            model=model, provider=litellm.LlmProviders(custom_llm_provider)
-        )
+        
         # get config from model, custom llm provider
         headers = provider_config.validate_environment(
             api_key=api_key,

litellm/llms/jina_ai/rerank/transformation.py

@@ -44,6 +44,7 @@ class JinaAIRerankConfig(BaseRerankConfig):
         rank_fields: Optional[List[str]] = None,
         return_documents: Optional[bool] = True,
         max_chunks_per_doc: Optional[int] = None,
+        max_tokens_per_doc: Optional[int] = None,
     ) -> OptionalRerankParams:
         optional_params = {}
         supported_params = self.get_supported_cohere_rerank_params(model)

litellm/proxy/_types.py

@@ -239,6 +239,7 @@ class LiteLLMRoutes(enum.Enum):
         # rerank
         "/rerank",
         "/v1/rerank",
+        "/v2/rerank"
         # realtime
         "/realtime",
         "/v1/realtime",

litellm/proxy/rerank_endpoints/endpoints.py

@@ -11,7 +11,12 @@ from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
 router = APIRouter()
 import asyncio
 
-
+@router.post(
+    "/v2/rerank",
+    dependencies=[Depends(user_api_key_auth)],
+    response_class=ORJSONResponse,
+    tags=["rerank"],
+)
 @router.post(
     "/v1/rerank",
     dependencies=[Depends(user_api_key_auth)],

litellm/rerank_api/main.py

@@ -81,6 +81,7 @@ def rerank(  # noqa: PLR0915
     rank_fields: Optional[List[str]] = None,
     return_documents: Optional[bool] = True,
     max_chunks_per_doc: Optional[int] = None,
+    max_tokens_per_doc: Optional[int] = None,
     **kwargs,
 ) -> Union[RerankResponse, Coroutine[Any, Any, RerankResponse]]:
     """
@@ -97,6 +98,14 @@ def rerank(  # noqa: PLR0915
     try:
         _is_async = kwargs.pop("arerank", False) is True
         optional_params = GenericLiteLLMParams(**kwargs)
+        # Params that are unique to specific versions of the client for the rerank call
+        unique_version_params = {
+            "max_chunks_per_doc": max_chunks_per_doc,
+            "max_tokens_per_doc": max_tokens_per_doc,
+        }
+        present_version_params = [
+            k for k, v in unique_version_params.items() if v is not None
+        ]
 
         model, _custom_llm_provider, dynamic_api_key, dynamic_api_base = (
             litellm.get_llm_provider(
@@ -111,6 +120,8 @@ def rerank(  # noqa: PLR0915
             ProviderConfigManager.get_provider_rerank_config(
                 model=model,
                 provider=litellm.LlmProviders(_custom_llm_provider),
+                api_base=optional_params.api_base,
+                present_version_params=present_version_params,
             )
         )
 
@@ -125,6 +136,7 @@ def rerank(  # noqa: PLR0915
             rank_fields=rank_fields,
             return_documents=return_documents,
             max_chunks_per_doc=max_chunks_per_doc,
+            max_tokens_per_doc=max_tokens_per_doc,
             non_default_params=kwargs,
         )
 
@@ -171,6 +183,7 @@ def rerank(  # noqa: PLR0915
             response = base_llm_http_handler.rerank(
                 model=model,
                 custom_llm_provider=_custom_llm_provider,
+                provider_config=rerank_provider_config,
                 optional_rerank_params=optional_rerank_params,
                 logging_obj=litellm_logging_obj,
                 timeout=optional_params.timeout,
@@ -192,6 +205,7 @@ def rerank(  # noqa: PLR0915
                 model=model,
                 custom_llm_provider=_custom_llm_provider,
                 optional_rerank_params=optional_rerank_params,
+                provider_config=rerank_provider_config,
                 logging_obj=litellm_logging_obj,
                 timeout=optional_params.timeout,
                 api_key=dynamic_api_key or optional_params.api_key,
@@ -220,6 +234,7 @@ def rerank(  # noqa: PLR0915
             response = base_llm_http_handler.rerank(
                 model=model,
                 custom_llm_provider=_custom_llm_provider,
+                provider_config=rerank_provider_config,
                 optional_rerank_params=optional_rerank_params,
                 logging_obj=litellm_logging_obj,
                 timeout=optional_params.timeout,
@@ -275,6 +290,7 @@ def rerank(  # noqa: PLR0915
                 custom_llm_provider=_custom_llm_provider,
                 optional_rerank_params=optional_rerank_params,
                 logging_obj=litellm_logging_obj,
+                provider_config=rerank_provider_config,
                 timeout=optional_params.timeout,
                 api_key=dynamic_api_key or optional_params.api_key,
                 api_base=api_base,

litellm/rerank_api/rerank_utils.py

@@ -15,6 +15,7 @@ def get_optional_rerank_params(
     rank_fields: Optional[List[str]] = None,
     return_documents: Optional[bool] = True,
     max_chunks_per_doc: Optional[int] = None,
+    max_tokens_per_doc: Optional[int] = None,
     non_default_params: Optional[dict] = None,
 ) -> OptionalRerankParams:
     all_non_default_params = non_default_params or {}
@@ -28,6 +29,8 @@ def get_optional_rerank_params(
         all_non_default_params["return_documents"] = return_documents
     if max_chunks_per_doc is not None:
         all_non_default_params["max_chunks_per_doc"] = max_chunks_per_doc
+    if max_tokens_per_doc is not None:
+        all_non_default_params["max_tokens_per_doc"] = max_tokens_per_doc
     return rerank_provider_config.map_cohere_rerank_params(
         model=model,
         drop_params=drop_params,
@@ -38,5 +41,6 @@ def get_optional_rerank_params(
         rank_fields=rank_fields,
         return_documents=return_documents,
         max_chunks_per_doc=max_chunks_per_doc,
+        max_tokens_per_doc=max_tokens_per_doc,
         non_default_params=all_non_default_params,
     )

litellm/types/rerank.py

@@ -18,6 +18,8 @@ class RerankRequest(BaseModel):
     rank_fields: Optional[List[str]] = None
     return_documents: Optional[bool] = None
     max_chunks_per_doc: Optional[int] = None
+    max_tokens_per_doc: Optional[int] = None
+
 
 
 class OptionalRerankParams(TypedDict, total=False):
@@ -27,6 +29,7 @@ class OptionalRerankParams(TypedDict, total=False):
     rank_fields: Optional[List[str]]
     return_documents: Optional[bool]
     max_chunks_per_doc: Optional[int]
+    max_tokens_per_doc: Optional[int]
 
 
 class RerankBilledUnits(TypedDict, total=False):

litellm/utils.py

@@ -6191,9 +6191,14 @@ class ProviderConfigManager:
     def get_provider_rerank_config(
         model: str,
         provider: LlmProviders,
+        api_base: Optional[str],
+        present_version_params: List[str],
     ) -> BaseRerankConfig:
         if litellm.LlmProviders.COHERE == provider:
-            return litellm.CohereRerankConfig()
+            if should_use_cohere_v1_client(api_base, present_version_params):
+                return litellm.CohereRerankConfig()
+            else:
+                return litellm.CohereRerankV2Config()
         elif litellm.LlmProviders.AZURE_AI == provider:
             return litellm.AzureAIRerankConfig()
         elif litellm.LlmProviders.INFINITY == provider:
@@ -6277,6 +6282,12 @@ def get_end_user_id_for_cost_tracking(
         return None
     return end_user_id
 
+def should_use_cohere_v1_client(api_base: Optional[str], present_version_params: List[str]):
+    if not api_base:
+        return False
+    uses_v1_params = ("max_chunks_per_doc" in present_version_params) and ('max_tokens_per_doc' not in present_version_params) 
+    return api_base.endswith("/v1/rerank") or (uses_v1_params and not api_base.endswith("/v2/rerank"))
+
 
 def is_prompt_caching_valid_prompt(
     model: str,

tests/litellm_utils_tests/test_utils.py

@@ -1970,6 +1970,26 @@ def test_get_applied_guardrails(test_case):
     # Assert
     assert sorted(result) == sorted(test_case["expected"])
 
+@pytest.mark.parametrize(
+    "endpoint, params, expected_bool",
+    [
+        ("localhost:4000/v1/rerank", ["max_chunks_per_doc"], True),
+        ("localhost:4000/v2/rerank", ["max_chunks_per_doc"], False),
+        ("localhost:4000", ["max_chunks_per_doc"], True),
+
+        ("localhost:4000/v1/rerank", ["max_tokens_per_doc"], True),
+        ("localhost:4000/v2/rerank", ["max_tokens_per_doc"], False),
+        ("localhost:4000", ["max_tokens_per_doc"], False),
+
+        ("localhost:4000/v1/rerank", ["max_chunks_per_doc", "max_tokens_per_doc"], True),
+        ("localhost:4000/v2/rerank", ["max_chunks_per_doc", "max_tokens_per_doc"], False),
+        ("localhost:4000", ["max_chunks_per_doc", "max_tokens_per_doc"], False),
+
+    ],
+)
+def test_should_use_cohere_v1_client(endpoint, params, expected_bool):
+    assert(litellm.utils.should_use_cohere_v1_client(endpoint, params) == expected_bool)
+
 
 def test_add_openai_metadata():
     from litellm.utils import add_openai_metadata

tests/llm_translation/test_rerank.py

@@ -111,35 +111,41 @@ async def test_basic_rerank(sync_mode):
 
 @pytest.mark.asyncio()
 @pytest.mark.parametrize("sync_mode", [True, False])
+@pytest.mark.skip(reason="Skipping test due to 503 Service Temporarily Unavailable")
 async def test_basic_rerank_together_ai(sync_mode):
-    if sync_mode is True:
-        response = litellm.rerank(
-            model="together_ai/Salesforce/Llama-Rank-V1",
-            query="hello",
-            documents=["hello", "world"],
-            top_n=3,
-        )
+    try:
+        if sync_mode is True:
+            response = litellm.rerank(
+                model="together_ai/Salesforce/Llama-Rank-V1",
+                query="hello",
+                documents=["hello", "world"],
+                top_n=3,
+            )
 
-        print("re rank response: ", response)
+            print("re rank response: ", response)
 
-        assert response.id is not None
-        assert response.results is not None
+            assert response.id is not None
+            assert response.results is not None
 
-        assert_response_shape(response, custom_llm_provider="together_ai")
-    else:
-        response = await litellm.arerank(
-            model="together_ai/Salesforce/Llama-Rank-V1",
-            query="hello",
-            documents=["hello", "world"],
-            top_n=3,
-        )
+            assert_response_shape(response, custom_llm_provider="together_ai")
+        else:
+            response = await litellm.arerank(
+                model="together_ai/Salesforce/Llama-Rank-V1",
+                query="hello",
+                documents=["hello", "world"],
+                top_n=3,
+            )
 
-        print("async re rank response: ", response)
+            print("async re rank response: ", response)
 
-        assert response.id is not None
-        assert response.results is not None
+            assert response.id is not None
+            assert response.results is not None
 
-        assert_response_shape(response, custom_llm_provider="together_ai")
+            assert_response_shape(response, custom_llm_provider="together_ai")
+    except Exception as e:
+        if "Service unavailable" in str(e):
+            pytest.skip("Skipping test due to 503 Service Temporarily Unavailable")
+        raise e
 
 
 @pytest.mark.asyncio()
@@ -184,8 +190,10 @@ async def test_basic_rerank_azure_ai(sync_mode):
 
 
 @pytest.mark.asyncio()
-async def test_rerank_custom_api_base():
+@pytest.mark.parametrize("version", ["v1", "v2"])
+async def test_rerank_custom_api_base(version):
     mock_response = AsyncMock()
+    litellm.cohere_key = "test_api_key"
 
     def return_val():
         return {
@@ -208,6 +216,10 @@ async def test_rerank_custom_api_base():
         "documents": ["hello", "world"],
     }
 
+    api_base = "https://exampleopenaiendpoint-production.up.railway.app/"
+    if version == "v1":
+        api_base += "v1/rerank"
+
     with patch(
         "litellm.llms.custom_httpx.http_handler.AsyncHTTPHandler.post",
         return_value=mock_response,
@@ -217,7 +229,7 @@ async def test_rerank_custom_api_base():
             query="hello",
             documents=["hello", "world"],
             top_n=3,
-            api_base="https://exampleopenaiendpoint-production.up.railway.app/",
+            api_base=api_base,
         )
 
         print("async re rank response: ", response)
@@ -230,7 +242,8 @@ async def test_rerank_custom_api_base():
         print("Arguments passed to API=", args_to_api)
         print("url = ", _url)
         assert (
-            _url == "https://exampleopenaiendpoint-production.up.railway.app/v1/rerank"
+            _url
+            == f"https://exampleopenaiendpoint-production.up.railway.app/{version}/rerank"
         )
 
         request_data = json.loads(args_to_api)
@@ -287,6 +300,7 @@ def test_complete_base_url_cohere():
 
     client = HTTPHandler()
     litellm.api_base = "http://localhost:4000"
+    litellm.cohere_key = "test_api_key"
     litellm.set_verbose = True
 
     text = "Hello there!"
@@ -308,7 +322,8 @@ def test_complete_base_url_cohere():
 
         print("mock_post.call_args", mock_post.call_args)
         mock_post.assert_called_once()
-        assert "http://localhost:4000/v1/rerank" in mock_post.call_args.kwargs["url"]
+        # Default to the v2 client when calling the base /rerank
+        assert "http://localhost:4000/v2/rerank" in mock_post.call_args.kwargs["url"]
 
 
 @pytest.mark.asyncio()
@@ -395,6 +410,63 @@ def test_rerank_response_assertions():
     assert_response_shape(r, custom_llm_provider="custom")
 
 
+def test_cohere_rerank_v2_client():
+    from litellm.llms.custom_httpx.http_handler import HTTPHandler
+
+    client = HTTPHandler()
+    litellm.api_base = "http://localhost:4000"
+    litellm.set_verbose = True
+
+    text = "Hello there!"
+    list_texts = ["Hello there!", "How are you?", "How do you do?"]
+
+    rerank_model = "rerank-multilingual-v3.0"
+
+    with patch.object(client, "post") as mock_post:
+        mock_response = MagicMock()
+        mock_response.text = json.dumps(
+            {
+                "id": "cmpl-mockid",
+                "results": [
+                    {"index": 0, "relevance_score": 0.95},
+                    {"index": 1, "relevance_score": 0.75},
+                    {"index": 2, "relevance_score": 0.65},
+                ],
+                "usage": {"prompt_tokens": 100, "total_tokens": 150},
+            }
+        )
+        mock_response.status_code = 200
+        mock_response.headers = {"Content-Type": "application/json"}
+        mock_response.json = lambda: json.loads(mock_response.text)
+
+        mock_post.return_value = mock_response
+
+        response = litellm.rerank(
+            model=rerank_model,
+            query=text,
+            documents=list_texts,
+            custom_llm_provider="cohere",
+            max_tokens_per_doc=3,
+            top_n=2,
+            api_key="fake-api-key",
+            client=client,
+        )
+
+        # Ensure Cohere API is called with the expected params
+        mock_post.assert_called_once()
+        assert mock_post.call_args.kwargs["url"] == "http://localhost:4000/v2/rerank"
+
+        request_data = json.loads(mock_post.call_args.kwargs["data"])
+        assert request_data["model"] == rerank_model
+        assert request_data["query"] == text
+        assert request_data["documents"] == list_texts
+        assert request_data["max_tokens_per_doc"] == 3
+        assert request_data["top_n"] == 2
+
+        # Ensure litellm response is what we expect
+        assert response["results"] == mock_response.json()["results"]
+
+
 @pytest.mark.flaky(retries=3, delay=1)
 def test_rerank_cohere_api():
     response = litellm.rerank(

tests/local_testing/test_embedding.py

@@ -961,7 +961,8 @@ async def test_gemini_embeddings(sync_mode, input):
 
 @pytest.mark.parametrize("sync_mode", [True, False])
 @pytest.mark.asyncio
-@pytest.mark.flaky(retries=3, delay=1)
+@pytest.mark.flaky(retries=6, delay=1)
+@pytest.mark.skip(reason="Skipping test due to flakyness")
 async def test_hf_embedddings_with_optional_params(sync_mode):
     litellm.set_verbose = True
 
@@ -992,8 +993,8 @@ async def test_hf_embedddings_with_optional_params(sync_mode):
                     wait_for_model=True,
                     client=client,
                 )
-        except Exception:
-            pass
+        except Exception as e:
+            print(e)
 
         mock_client.assert_called_once()