ibm-watsonx-orchestrate-evaluation-framework 1.0.9__py3-none-any.whl → 1.1.0__py3-none-any.whl

{ibm_watsonx_orchestrate_evaluation_framework-1.0.9.dist-info → ibm_watsonx_orchestrate_evaluation_framework-1.1.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ibm-watsonx-orchestrate-evaluation-framework
-Version: 1.0.9
+Version: 1.1.0
 Summary: The WxO evaluation framework
 Author-email: Haode Qi <Haode.Qi@ibm.com>
 License: MIT

{ibm_watsonx_orchestrate_evaluation_framework-1.0.9.dist-info → ibm_watsonx_orchestrate_evaluation_framework-1.1.0.dist-info}/RECORD

@@ -53,7 +53,7 @@ wxo_agentic_evaluation/red_teaming/attack_generator.py,sha256=YQi9xoaFATBNGe_Neb
 wxo_agentic_evaluation/red_teaming/attack_list.py,sha256=edphWARWqDtXFtcHTVbRXngvO0YfG5SgrfPtrBRXuFw,4734
 wxo_agentic_evaluation/red_teaming/attack_runner.py,sha256=qBZY4GK1352NUMyED5LVjjbcvpdCcxG6mDIN1HvxKIc,4340
 wxo_agentic_evaluation/referenceless_eval/__init__.py,sha256=lijXMgQ8nQe-9eIfade2jLfHMlXfYafMZIwXtC9KDZo,106
-wxo_agentic_evaluation/referenceless_eval/referenceless_eval.py,sha256=82jutvZ21GO76rRcjGWux5kAanDtzE728BijC7trSxY,4297
+wxo_agentic_evaluation/referenceless_eval/referenceless_eval.py,sha256=ypEMOeAwaztGkOuDr_2JArSQWwos7XcBTwo8lFs2N5w,4262
 wxo_agentic_evaluation/referenceless_eval/function_calling/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 wxo_agentic_evaluation/referenceless_eval/function_calling/consts.py,sha256=UidTaT9g5IxbcakfQqP_9c5civ1wDqY-PpPUf0uOXJo,915
 wxo_agentic_evaluation/referenceless_eval/function_calling/metrics/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -67,7 +67,7 @@ wxo_agentic_evaluation/referenceless_eval/function_calling/metrics/function_sele
 wxo_agentic_evaluation/referenceless_eval/function_calling/metrics/function_selection/function_selection_metrics_runtime.json,sha256=o4oRur1MiXO2RYzmzj07QOBzX75DyU7T7yd-gFsgFdo,30563
 wxo_agentic_evaluation/referenceless_eval/function_calling/pipeline/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 wxo_agentic_evaluation/referenceless_eval/function_calling/pipeline/adapters.py,sha256=kMMFq4ABX5q6cPnDdublLMVqXu4Ij-x4OlxZyePWIjc,3599
-wxo_agentic_evaluation/referenceless_eval/function_calling/pipeline/pipeline.py,sha256=rKKBtL47gvN_fFy3FVbQTaa0U1bhv8bls0HPZi66EZ8,9279
+wxo_agentic_evaluation/referenceless_eval/function_calling/pipeline/pipeline.py,sha256=44HNEoIt3_jKZczs1qB8WGltCG-vn3ZI5aNhucxSDeM,9272
 wxo_agentic_evaluation/referenceless_eval/function_calling/pipeline/semantic_checker.py,sha256=z_k-qdFoUJqstkPYn9Zmhlp2YTVQKJtoDZCIdKow664,17306
 wxo_agentic_evaluation/referenceless_eval/function_calling/pipeline/static_checker.py,sha256=_Er2KfCkc3HFmOmxZT6eb-e7qF7ukqsf6Si5CJTqPPg,6016
 wxo_agentic_evaluation/referenceless_eval/function_calling/pipeline/transformation_prompts.py,sha256=tW1wc87WIm8BZh2lhdj1RDP6VdRLqZBWSMmemSttbGs,22034
@@ -80,17 +80,18 @@ wxo_agentic_evaluation/referenceless_eval/metrics/prompt.py,sha256=CGQ5LvhQrmxAy
 wxo_agentic_evaluation/referenceless_eval/metrics/utils.py,sha256=jurmc4KFFKH4hwnvor2xg97H91b-xJc3cUKYaU2I8uM,1370
 wxo_agentic_evaluation/referenceless_eval/prompt/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 wxo_agentic_evaluation/referenceless_eval/prompt/runner.py,sha256=FFmcSWXQnLmylpYyj8LZuPwb6nqwQp-jj6Mv9g8zby0,5052
-wxo_agentic_evaluation/service_provider/__init__.py,sha256=EaY4jjKp58M3W8N3b3a8PNC2S81xA7YV2_QkTIy9DfI,1600
-wxo_agentic_evaluation/service_provider/model_proxy_provider.py,sha256=Y36Ryv4nPG8RdVP_zsQsRlEWv8F_hGi7-wOppWPQTwc,4026
+wxo_agentic_evaluation/service_provider/__init__.py,sha256=yNQ-urOIdjANbpCzVAhkPHNcpBY6hndDJgPZM1C2qeo,2107
+wxo_agentic_evaluation/service_provider/model_proxy_provider.py,sha256=EW1JIiIWoKaTTC-fqKURSsbdyo-dbVWYVrXY8-gEmvc,4081
 wxo_agentic_evaluation/service_provider/ollama_provider.py,sha256=HMHQVUGFbLSQI1dhysAn70ozJl90yRg-CbNd4vsz-Dc,1116
 wxo_agentic_evaluation/service_provider/provider.py,sha256=MsnRzLYAaQiU6y6xf6eId7kn6-CetQuNZl00EP-Nl28,417
-wxo_agentic_evaluation/service_provider/watsonx_provider.py,sha256=C_uezsx4FHqkvtCIQGIii82nu9_vSOKP70uAVXWj5fM,10619
+wxo_agentic_evaluation/service_provider/referenceless_provider_wrapper.py,sha256=aJrCz8uco6HOQwNCSjEKviwnhlyLTNAGpLtsOAegQ70,5200
+wxo_agentic_evaluation/service_provider/watsonx_provider.py,sha256=ugXCXwrfi_XC2d9FPa96ccMKGQbTd1ElDw8RNR8TDB8,6544
 wxo_agentic_evaluation/utils/__init__.py,sha256=ItryTgc1jVc32rB3XktTFaYGA_A6bRIDZ1Pts_JGmv8,144
 wxo_agentic_evaluation/utils/open_ai_tool_extractor.py,sha256=Vyji_edgou2xMLbsGwFG-QI7xRBNvO3-1nbeOc8ZuFo,5646
 wxo_agentic_evaluation/utils/rich_utils.py,sha256=J9lzL4ETQeiAJcXKsUzXh82XdKvlDY7jmcgTQlwmL9s,6252
 wxo_agentic_evaluation/utils/rouge_score.py,sha256=WvcGh6mwF4rWH599J9_lAt3BfaHbAZKtKEJBsC61iKo,692
 wxo_agentic_evaluation/utils/utils.py,sha256=qQR_2W5p0Rk6KSE3-llRyZrWXkO5zG9JW7H1692L4PI,11428
-ibm_watsonx_orchestrate_evaluation_framework-1.0.9.dist-info/METADATA,sha256=OgVKCRnQdk7cYnzrexY8VPPgQW73uHdX2jfI9a7o5IE,16105
-ibm_watsonx_orchestrate_evaluation_framework-1.0.9.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-ibm_watsonx_orchestrate_evaluation_framework-1.0.9.dist-info/top_level.txt,sha256=2okpqtpxyqHoLyb2msio4pzqSg7yPSzwI7ekks96wYE,23
-ibm_watsonx_orchestrate_evaluation_framework-1.0.9.dist-info/RECORD,,
+ibm_watsonx_orchestrate_evaluation_framework-1.1.0.dist-info/METADATA,sha256=4yfSRfaNQUwauYPqvTFAoaVSn_c3i5YbIC7SFK4SnDU,16105
+ibm_watsonx_orchestrate_evaluation_framework-1.1.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+ibm_watsonx_orchestrate_evaluation_framework-1.1.0.dist-info/top_level.txt,sha256=2okpqtpxyqHoLyb2msio4pzqSg7yPSzwI7ekks96wYE,23
+ibm_watsonx_orchestrate_evaluation_framework-1.1.0.dist-info/RECORD,,

wxo_agentic_evaluation/referenceless_eval/function_calling/pipeline/pipeline.py

@@ -20,7 +20,7 @@ from wxo_agentic_evaluation.referenceless_eval.function_calling.pipeline.types i
     ToolSpec,
 )
 from wxo_agentic_evaluation.referenceless_eval.function_calling import metrics
-from wxo_agentic_evaluation.service_provider.watsonx_provider import WatsonXLLMKitWrapper
+from wxo_agentic_evaluation.service_provider.referenceless_provider_wrapper import LLMKitWrapper
 
 def metrics_dir():
     path = importlib.resources.files(metrics)
@@ -57,8 +57,8 @@ class ReflectionPipeline:
 
     def __init__(
         self,
-        metrics_client: WatsonXLLMKitWrapper,
-        codegen_client: Optional[WatsonXLLMKitWrapper] = None,
+        metrics_client: LLMKitWrapper,
+        codegen_client: Optional[LLMKitWrapper] = None,
         general_metrics: Optional[
             Union[Path, List[FunctionCallMetric], List[str]]
         ] = _DEFAULT_GENERAL_RUNTIME,

wxo_agentic_evaluation/referenceless_eval/referenceless_eval.py

@@ -1,5 +1,4 @@
 import json
-import os
 from typing import Any, List, Mapping
 
 import rich
@@ -16,7 +15,7 @@ from wxo_agentic_evaluation.referenceless_eval.function_calling.pipeline.types i
     ToolSpec,
 )
 from wxo_agentic_evaluation.type import Message
-from wxo_agentic_evaluation.service_provider.watsonx_provider import WatsonXLLMKitWrapper
+from wxo_agentic_evaluation.service_provider import get_provider
 
 class ReferencelessEvaluation:
     """
@@ -33,12 +32,12 @@ class ReferencelessEvaluation:
         messages: List[Message],
         model_id: str,
         task_n: str,
-        dataset_name: str,
-    ):
-        self.metrics_client = WatsonXLLMKitWrapper(
+        dataset_name: str,):        
+        
+        self.metrics_client = get_provider(
             model_id=model_id,
-            api_key=os.getenv("WATSONX_APIKEY", ""),
-            space_id=os.getenv("WATSONX_SPACE_ID")
+            params={"min_new_tokens": 0, "decoding_method": "greedy", "max_new_tokens": 4096},
+            referenceless_eval=True
         )
 
         self.pipeline = ReflectionPipeline(
@@ -57,7 +56,6 @@ class ReferencelessEvaluation:
     def _run_pipeline(self, examples: List[Mapping[str, Any]]):
         results = []
         for example in examples:
-            # self.pipeline.sy
             result = self.pipeline.run_sync(
                 conversation=example["context"],
                 inventory=self.apis_specs,

wxo_agentic_evaluation/service_provider/__init__.py

@@ -1,21 +1,30 @@
 from wxo_agentic_evaluation.service_provider.ollama_provider import OllamaProvider
 from wxo_agentic_evaluation.service_provider.watsonx_provider import WatsonXProvider
 from wxo_agentic_evaluation.service_provider.model_proxy_provider import ModelProxyProvider
+from wxo_agentic_evaluation.service_provider.referenceless_provider_wrapper import ModelProxyProviderLLMKitWrapper, WatsonXLLMKitWrapper
 from wxo_agentic_evaluation.arg_configs import ProviderConfig
 
 import os
 
-def _instantiate_provider(config: ProviderConfig, **kwargs):
+def _instantiate_provider(config: ProviderConfig, is_referenceless_eval: bool = False, **kwargs):
     if config.provider == "watsonx":
-        return WatsonXProvider(model_id=config.model_id, **kwargs)
+        if is_referenceless_eval:
+            provider = WatsonXLLMKitWrapper
+        else:
+            provider = WatsonXProvider
+        return provider(model_id=config.model_id, **kwargs)
     elif config.provider == "ollama":
         return OllamaProvider(model_id=config.model_id, **kwargs)
     elif config.provider == "model_proxy":
-        return ModelProxyProvider(model_id=config.model_id, **kwargs)
+        if is_referenceless_eval:
+            provider = ModelProxyProviderLLMKitWrapper
+        else:
+            provider = ModelProxyProvider
+        return provider(model_id=config.model_id, **kwargs)
     else:
         raise RuntimeError(f"target provider is not supported {config.provider}")
     
-def get_provider(config: ProviderConfig = None, model_id: str = None, **kwargs):
+def get_provider(config: ProviderConfig = None, model_id: str = None, referenceless_eval: bool = False, **kwargs):
     if config:
         return _instantiate_provider(config, **kwargs)
 
@@ -24,11 +33,11 @@ def get_provider(config: ProviderConfig = None, model_id: str = None, **kwargs):
 
     if "WATSONX_APIKEY" in os.environ and "WATSONX_SPACE_ID" in os.environ:
         config = ProviderConfig(provider="watsonx", model_id=model_id)
-        return _instantiate_provider(config, **kwargs)
+        return _instantiate_provider(config, referenceless_eval, **kwargs)
 
     if "WO_API_KEY" in os.environ and "WO_INSTANCE" in os.environ:
         config = ProviderConfig(provider="model_proxy", model_id=model_id)
-        return _instantiate_provider(config, **kwargs)
+        return _instantiate_provider(config, referenceless_eval, **kwargs)
 
     raise RuntimeError(
         "No provider found. Please either provide a config or set the required environment variables."

wxo_agentic_evaluation/service_provider/model_proxy_provider.py

@@ -38,9 +38,10 @@ class ModelProxyProvider(Provider):
         self.api_key = api_key
         self.is_ibm_cloud = is_ibm_cloud_url(instance_url)
         self.auth_url = AUTH_ENDPOINT_IBM_CLOUD if self.is_ibm_cloud else AUTH_ENDPOINT_AWS
-        self.url = instance_url + "/ml/v1/text/generation?version=2024-05-01"
-
-        self.embedding_url = instance_url + "/ml/v1/text/embeddings"
+    
+        self.instance_url = instance_url
+        self.url = self.instance_url + "/ml/v1/text/generation?version=2024-05-01"
+        self.embedding_url = self.instance_url + "/ml/v1/text/embeddings"
 
         self.lock = Lock()
         self.token, self.refresh_time = self.get_token()

wxo_agentic_evaluation/service_provider/referenceless_provider_wrapper.py

@@ -0,0 +1,138 @@
+import requests
+from typing import List, Mapping, Union, Optional, Any
+from abc import ABC, abstractmethod
+
+import rich
+
+from wxo_agentic_evaluation.service_provider.model_proxy_provider import ModelProxyProvider
+from wxo_agentic_evaluation.service_provider.watsonx_provider import WatsonXProvider
+
+
+class LLMResponse:
+    """
+    NOTE: Taken from LLM-Eval-Kit
+    Response object that can contain both content and tool calls
+    """
+
+    def __init__(self, content: str, tool_calls: Optional[List[Mapping[str, Any]]] = None):
+        self.content = content
+        self.tool_calls = tool_calls or []
+
+    def __str__(self) -> str:
+        """Return the content of the response as a string."""
+        return self.content
+
+    def __repr__(self) -> str:
+        """Return a string representation of the LLMResponse object."""
+        return f"LLMResponse(content='{self.content}', tool_calls={self.tool_calls})"
+
+class LLMKitWrapper(ABC):
+    """ In the future this wrapper won't be neccesary.
+    Right now the referenceless code requires a `generate()` function for the metrics client.
+    In refactor, rewrite referenceless code so this wrapper is not needed.
+    """
+    @abstractmethod
+    def chat():
+        pass
+
+    def generate(
+            self,
+            prompt: Union[str, List[Mapping[str, str]]],
+            *,
+            schema,
+            retries: int = 3,
+            generation_args: Optional[Any] = None,
+            **kwargs: Any
+        ):
+
+        """
+        In future, implement validation of response like in llmevalkit
+        """
+
+        for attempt in range(1, retries + 1):
+            try:
+                raw_response = self.chat(prompt)
+                response = self._parse_llm_response(raw_response)
+                return response
+            except Exception as e:
+                rich.print(f"[b][r] Generation failed with error '{str(e)}' during `quick-eval` ... Attempt ({attempt} / {retries}))")
+
+    def _parse_llm_response(self, raw: Any) -> Union[str, LLMResponse]:
+        """
+        Extract the generated text and tool calls from a watsonx response.
+
+        - For text generation: raw['results'][0]['generated_text']
+        - For chat:           raw['choices'][0]['message']['content']
+        """
+        content = ""
+        tool_calls = []
+
+        if isinstance(raw, dict) and "choices" in raw:
+            choices = raw["choices"]
+            if isinstance(choices, list) and choices:
+                first = choices[0]
+                msg = first.get("message")
+                if isinstance(msg, dict):
+                    content = msg.get("content", "")
+                    # Extract tool calls if present
+                    if "tool_calls" in msg and msg["tool_calls"]:
+                        tool_calls = []
+                        for tool_call in msg["tool_calls"]:
+                            tool_call_dict = {
+                                "id": tool_call.get("id"),
+                                "type": tool_call.get("type", "function"),
+                                "function": {
+                                    "name": tool_call.get("function", {}).get("name"),
+                                    "arguments": tool_call.get("function", {}).get(
+                                        "arguments"
+                                    ),
+                                },
+                            }
+                            tool_calls.append(tool_call_dict)
+                elif "text" in first:
+                    content = first["text"]
+
+        if not content and not tool_calls:
+            raise ValueError(f"Unexpected watsonx response format: {raw!r}")
+
+        # Return LLMResponse if tool calls exist, otherwise just content
+        if tool_calls:
+            return LLMResponse(content=content, tool_calls=tool_calls)
+
+        return content
+
+class ModelProxyProviderLLMKitWrapper(ModelProxyProvider, LLMKitWrapper):
+    def chat(self, sentence: List[str]):
+        if self.model_id is None:
+            raise Exception("model id must be specified for text generation")
+        chat_url = f"{self.instance_url}/ml/v1/text/chat?version=2023-10-25"
+        self.refresh_token_if_expires()
+        headers = self.get_header()
+        data = {
+            "model_id": self.model_id,
+            "messages": sentence,
+            "parameters": self.params,
+            "space_id": "1",
+            "timeout": self.timeout
+        }
+        resp = requests.post(url=chat_url, headers=headers, json=data)
+        if resp.status_code == 200:
+            return resp.json()
+        else:
+            resp.raise_for_status()
+
+class WatsonXLLMKitWrapper(WatsonXProvider, LLMKitWrapper):
+    def chat(self, sentence: list):
+        chat_url = f"{self.api_endpoint}/ml/v1/text/chat?version=2023-05-02"
+        headers = self.prepare_header()
+        data = {
+            "model_id": self.model_id,
+            "messages": sentence,
+            "parameters": self.params,
+            "space_id": self.space_id
+        }
+        resp = requests.post(url=chat_url, headers=headers, json=data)
+        if resp.status_code == 200:
+            return resp.json()
+        else:
+            resp.raise_for_status()

wxo_agentic_evaluation/service_provider/watsonx_provider.py

@@ -2,12 +2,10 @@ import os
 import requests
 import json
 from types import MappingProxyType
-from typing import List, Mapping, Union, Optional, Any
-from functools import singledispatchmethod
+from typing import List, Mapping, Union
 import dataclasses
 from threading import Lock
 import time
-import rich
 from wxo_agentic_evaluation.service_provider.provider import Provider
 
 ACCESS_URL = "https://iam.cloud.ibm.com/identity/token"
@@ -90,12 +88,7 @@ class WatsonXProvider(Provider):
                   "Content-Type": "application/json"}
         return headers
 
-    @singledispatchmethod
-    def generate(self, sentence):
-        raise ValueError(f"Input must either be a string or a list of dictionaries")
-
-    @generate.register
-    def _(self, sentence: str):
+    def _query(self, sentence: str):
         headers = self.prepare_header()
 
         data = {"model_id": self.model_id, "input": sentence,
@@ -107,22 +100,6 @@ class WatsonXProvider(Provider):
         else:
             resp.raise_for_status()
 
-    @generate.register
-    def _(self, sentence: list):
-        chat_url = f"{self.api_endpoint}/ml/v1/text/chat?version=2023-05-02"
-        headers = self.prepare_header()
-        data = {
-            "model_id": self.model_id,
-            "messages": sentence,
-            "parameters": self.params,
-            "space_id": self.space_id
-        }
-        resp = requests.post(url=chat_url, headers=headers, json=data)
-        if resp.status_code == 200:
-            return resp.json()
-        else:
-            resp.raise_for_status()
-
     def _refresh_token(self):
         # if we do not have a token or the current timestamp is 9 minutes away from expire.
         if not self.access_token or time.time() > self.refresh_time:
@@ -134,7 +111,7 @@ class WatsonXProvider(Provider):
         if self.model_id is None:
             raise Exception("model id must be specified for text generation")
         try:
-            response = self.generate(sentence)
+            response = self._query(sentence)
             if (generated_text := response.get("generated_text")):
                 return generated_text
             elif (message := response.get("message")):
@@ -165,90 +142,6 @@ class WatsonXProvider(Provider):
         else:
             resp.raise_for_status()
 
-class LLMResponse:
-    """
-    NOTE: Taken from LLM-Eval-Kit
-    Response object that can contain both content and tool calls
-    """
-
-    def __init__(self, content: str, tool_calls: Optional[List[Mapping[str, Any]]] = None):
-        self.content = content
-        self.tool_calls = tool_calls or []
-
-    def __str__(self) -> str:
-        """Return the content of the response as a string."""
-        return self.content
-
-    def __repr__(self) -> str:
-        """Return a string representation of the LLMResponse object."""
-        return f"LLMResponse(content='{self.content}', tool_calls={self.tool_calls})"
-
-class WatsonXLLMKitWrapper(WatsonXProvider):
-    def generate(
-            self,
-            prompt: Union[str, List[Mapping[str, str]]],
-            *,
-            schema,
-            retries: int = 3,
-            generation_args: Optional[Any] = None,
-            **kwargs: Any
-        ):
-
-        """
-        In future, implement validation of response like in llmevalkit
-        """
-
-        for attempt in range(1, retries + 1):
-            try:
-                raw_response = super().generate(prompt)
-                response = self._parse_llm_response(raw_response)
-                return response
-            except Exception as e:
-                rich.print(f"[b][r] WatsonX generation failed with error '{str(e)}' during `quick-eval` ... Attempt ({attempt} / {retries}))")
-
-    def _parse_llm_response(self, raw: Any) -> Union[str, LLMResponse]:
-        """
-        Extract the generated text and tool calls from a watsonx response.
-
-        - For text generation: raw['results'][0]['generated_text']
-        - For chat:           raw['choices'][0]['message']['content']
-        """
-        content = ""
-        tool_calls = []
-
-        if isinstance(raw, dict) and "choices" in raw:
-            choices = raw["choices"]
-            if isinstance(choices, list) and choices:
-                first = choices[0]
-                msg = first.get("message")
-                if isinstance(msg, dict):
-                    content = msg.get("content", "")
-                    # Extract tool calls if present
-                    if "tool_calls" in msg and msg["tool_calls"]:
-                        tool_calls = []
-                        for tool_call in msg["tool_calls"]:
-                            tool_call_dict = {
-                                "id": tool_call.get("id"),
-                                "type": tool_call.get("type", "function"),
-                                "function": {
-                                    "name": tool_call.get("function", {}).get("name"),
-                                    "arguments": tool_call.get("function", {}).get(
-                                        "arguments"
-                                    ),
-                                },
-                            }
-                            tool_calls.append(tool_call_dict)
-                elif "text" in first:
-                    content = first["text"]
-
-        if not content and not tool_calls:
-            raise ValueError(f"Unexpected watsonx response format: {raw!r}")
-
-        # Return LLMResponse if tool calls exist, otherwise just content
-        if tool_calls:
-            return LLMResponse(content=content, tool_calls=tool_calls)
-
-        return content
 
 if __name__ == "__main__":
     provider = WatsonXProvider(model_id="meta-llama/llama-3-2-90b-vision-instruct")