ibm-watsonx-orchestrate-evaluation-framework 1.1.0__py3-none-any.whl → 1.1.2__py3-none-any.whl

wxo_agentic_evaluation/external_agent/types.py

@@ -1,5 +1,6 @@
+from typing import Any, List, Literal, Mapping, Union
+
 from pydantic import BaseModel
-from typing import List, Union, Literal, Mapping, Any
 
 
 class ThinkingStepDetails(BaseModel):
@@ -25,7 +26,9 @@ class ToolResponseStepDetails(BaseModel):
     tool_call_id: str
 
 
-StepDetails = Union[ThinkingStepDetails, ToolCallsStepDetails, ToolResponseStepDetails]
+StepDetails = Union[
+    ThinkingStepDetails, ToolCallsStepDetails, ToolResponseStepDetails
+]
 
 
 class DeltaMessageChoice(BaseModel):
@@ -59,8 +62,12 @@ class ThreadRunStepDeltaData(BaseEventData):
 
 
 class UniversalData(BaseEventData):
-    object: Union[Literal["thread.message.delta"], Literal["thread.run.step.delta"],
-                  Literal["thread.run.step.created"], Literal["thread.run.step.completed"]]
+    object: Union[
+        Literal["thread.message.delta"],
+        Literal["thread.run.step.delta"],
+        Literal["thread.run.step.created"],
+        Literal["thread.run.step.completed"],
+    ]
     choices: List[ThreadMessageDeltaChoice]
     choices: List[Union[ThreadMessageDeltaChoice, dict]]
 
@@ -68,4 +75,4 @@ class UniversalData(BaseEventData):
 class SchemaValidationResults(BaseModel):
     success: bool
     logged_events: List[str]
-    messages: List[Mapping[Any, Any]]
+    messages: List[Mapping[Any, Any]]

wxo_agentic_evaluation/inference_backend.py

@@ -1,61 +1,68 @@
-import requests
-import os
-import yaml
 import json
-import rich
+import os
 import time
-from pydantic import BaseModel
-from typing import List, Generator, Dict, Tuple, Mapping, Any
-from enum import Enum
 from collections import deque
+import urllib3
+from urllib3.exceptions import InsecureRequestWarning
+from enum import Enum
+from typing import Any, Dict, Generator, List, Mapping, Tuple
 
+import requests
+import rich
+import yaml
+from pydantic import BaseModel
+
+from wxo_agentic_evaluation.arg_configs import TestConfig
+from wxo_agentic_evaluation.llm_user import LLMUser
+from wxo_agentic_evaluation.service_instance import tenant_setup
+from wxo_agentic_evaluation.service_provider.watsonx_provider import (
+    WatsonXProvider,
+)
 from wxo_agentic_evaluation.type import (
     ContentType,
-    Message,
+    ConversationalConfidenceThresholdScore,
     ConversationalSearch,
     ConversationalSearchCitations,
     ConversationalSearchResultMetadata,
-    ConversationalConfidenceThresholdScore,
     ConversationalSearchResults,
     ConversationSearchMetadata,
+    Message,
 )
-from wxo_agentic_evaluation.llm_user import LLMUser
-from wxo_agentic_evaluation.service_provider.watsonx_provider import WatsonXProvider
-from wxo_agentic_evaluation.arg_configs import TestConfig
-from wxo_agentic_evaluation.service_instance import tenant_setup
 from wxo_agentic_evaluation.utils.utils import (
+    Tokenizer,
     is_saas_url,
     safe_divide,
-    Tokenizer
 )
 
 tokenizer = Tokenizer()
 
+
 class Roles(Enum):
     ASSISTANT = "assistant"
     USER = "user"
 
-def calculate_word_overlap_similarity_score(first_message_text: str, second_message_text: str) -> float:
+
+def calculate_word_overlap_similarity_score(
+    first_message_text: str, second_message_text: str
+) -> float:
     """Calculate the word overlap similarity score between the .content field of two Message objects.
     Args:
         first_message_text (str): The .content field of the first message.
         second_message_text (str): The .content field of the second message.
     """
-    
+
     words_in_first_message = tokenizer(first_message_text)
     words_in_second_message = tokenizer(second_message_text)
 
     # Calculate the number of common words
     common_words = set(words_in_first_message) & set(words_in_second_message)
     unique_words = set(words_in_first_message + words_in_second_message)
-    
+
     unique_words_count = len(unique_words)
     common_words_count = len(common_words)
 
-    return safe_divide(
-        common_words_count,
-        unique_words_count
-    )
+    return safe_divide(common_words_count, unique_words_count)
+
 
 def is_transfer_response(step_detail: Dict):
     # this is not very reliable
@@ -77,6 +84,13 @@ class WXOClient:
         self.service_url = service_url
         self.api_key = api_key
 
+        env_ssl_verify = os.getenv("WO_SSL_VERIFY", "true")
+        verify = isinstance(env_ssl_verify, str) and env_ssl_verify.strip().lower() == "true"
+        self._verify_ssl = verify
+
+        if not self._verify_ssl:
+            urllib3.disable_warnings(InsecureRequestWarning)
+
     def _get_headers(self) -> dict:
         headers = {}
         if self.api_key:
@@ -86,12 +100,12 @@ class WXOClient:
     def post(self, payload: dict, path: str, stream=False):
         url = f"{self.service_url}/{path}"
         return requests.post(
-            url=url, headers=self._get_headers(), json=payload, stream=stream
+            url=url, headers=self._get_headers(), json=payload, stream=stream, verify=self._verify_ssl
         )
 
     def get(self, path: str, params: dict = None):
         url = f"{self.service_url}/{path}"
-        return requests.get(url, params=params, headers=self._get_headers())
+        return requests.get(url, params=params, headers=self._get_headers(), verify=self._verify_ssl)
 
 
 class WXOInferenceBackend:
@@ -135,7 +149,9 @@ class WXOInferenceBackend:
         else:
             path = "v1/orchestrate/runs?stream=true"
 
-        response: requests.Response = self.wxo_client.post(payload, path, stream=True)
+        response: requests.Response = self.wxo_client.post(
+            payload, path, stream=True
+        )
         import json
 
         for chunk in self._parse_events(response):
@@ -188,7 +204,9 @@ class WXOInferenceBackend:
         citations = parse_citations()
         retrieval_context = parsed_search_results()
         citations_title = conversational_search.get("citations_title", "")
-        response_length_option = conversational_search.get("response_length_option", "")
+        response_length_option = conversational_search.get(
+            "response_length_option", ""
+        )
         text = conversational_search.get("text", "")
 
         confidence_scores = ConversationalConfidenceThresholdScore(
@@ -261,7 +279,9 @@ class WXOInferenceBackend:
                                         )
                                     )
                                     end_time = time.time()
-                                    call_tracker.tool_call.append(end_time - start_time)
+                                    call_tracker.tool_call.append(
+                                        end_time - start_time
+                                    )
                                     start_time = end_time
                             elif step_detail["type"] == "tool_call":
                                 # in step details, we could have [tool_response, tool_call]
@@ -279,7 +299,9 @@ class WXOInferenceBackend:
                                     )
                                 )
                                 end_time = time.time()
-                                call_tracker.tool_call.append(end_time - start_time)
+                                call_tracker.tool_call.append(
+                                    end_time - start_time
+                                )
                                 start_time = end_time
                             elif step_detail["type"] == "tool_response":
                                 content = json.dumps(step_detail)
@@ -293,7 +315,9 @@ class WXOInferenceBackend:
                                     )
                                 )
                                 end_time = time.time()
-                                call_tracker.tool_response.append(end_time - start_time)
+                                call_tracker.tool_response.append(
+                                    end_time - start_time
+                                )
                                 start_time = end_time
                     elif content_field := delta.get("content"):
                         for val in content_field:
@@ -312,7 +336,9 @@ class WXOInferenceBackend:
                                     chunk=event,
                                 )
                                 end_time = time.time()
-                                call_tracker.generic.append(end_time - start_time)
+                                call_tracker.generic.append(
+                                    end_time - start_time
+                                )
                                 start_time = end_time
 
                 # NOTE: The event here that is parsed is part of the "message.created" event
@@ -336,10 +362,14 @@ class WXOInferenceBackend:
                             """
 
                             last_message = json.loads(messages[-1].content)
-                            tool_call_id = last_message.get("tool_call_id", None)
+                            tool_call_id = last_message.get(
+                                "tool_call_id", None
+                            )
                             assert tool_call_id is not None
-                            conversational_search_metadata = ConversationSearchMetadata(
-                                tool_call_id=tool_call_id
+                            conversational_search_metadata = (
+                                ConversationSearchMetadata(
+                                    tool_call_id=tool_call_id
+                                )
                             )
                             conversational_search = (
                                 self.parse_conversational_search_response(
@@ -347,7 +377,9 @@ class WXOInferenceBackend:
                                     metadata=conversational_search_metadata,
                                 )
                             )
-                            conversational_search_data.append(conversational_search)
+                            conversational_search_data.append(
+                                conversational_search
+                            )
                             messages.append(
                                 Message(
                                     role=role,
@@ -436,7 +468,10 @@ class WXOInferenceBackend:
                                     content = json.dumps(tool_json)
                                     # TO-DO: review do we even need the get messages for retry loop anymore?
                                     if msg_content := entry.get("content"):
-                                        if msg_content[0].get("response_type") == "conversational_search":
+                                        if (
+                                            msg_content[0].get("response_type")
+                                            == "conversational_search"
+                                        ):
                                             continue
                                     messages.append(
                                         Message(
@@ -451,7 +486,9 @@ class WXOInferenceBackend:
                                 content = json.dumps(step_detail)
                                 messages.append(
                                     Message(
-                                        role=role, content=content, type=content_type
+                                        role=role,
+                                        content=content,
+                                        type=content_type,
                                     )
                                 )
                             else:
@@ -459,7 +496,9 @@ class WXOInferenceBackend:
                                 content_type = ContentType.tool_response
                                 messages.append(
                                     Message(
-                                        role=role, content=content, type=content_type
+                                        role=role,
+                                        content=content,
+                                        type=content_type,
                                     )
                                 )
             if content_field := entry.get("content"):
@@ -468,12 +507,19 @@ class WXOInferenceBackend:
                     if val["response_type"] == ContentType.text:
                         messages.append(
                             Message(
-                                role=role, content=val["text"], type=ContentType.text
+                                role=role,
+                                content=val["text"],
+                                type=ContentType.text,
                             )
                         )
-                    if val["response_type"] == ContentType.conversational_search:
-                        conversational_search_metadata = ConversationSearchMetadata(
-                            tool_call_id=tool_call_id
+                    if (
+                        val["response_type"]
+                        == ContentType.conversational_search
+                    ):
+                        conversational_search_metadata = (
+                            ConversationSearchMetadata(
+                                tool_call_id=tool_call_id
+                            )
                         )
                         messages.append(
                             Message(
@@ -538,8 +584,12 @@ class WXOInferenceBackend:
 class EvaluationController:
 
     MAX_CONVERSATION_STEPS = int(os.getenv("MAX_CONVERSATION_STEPS", 20))
-    MESSAGE_SIMILARITY_THRESHOLD = float(os.getenv("MESSAGE_SIMILARITY_THRESHOLD", 0.98))  # if any two consecutive messages are >98% similar, the inference loop will be terminated
-    MAX_REPEATING_MESSAGES = int(os.getenv("MAX_REPEATING_MESSAGES", 3)) # this is the maximum number of repeating messages by the user or assistant before terminating the inference loop
+    MESSAGE_SIMILARITY_THRESHOLD = float(
+        os.getenv("MESSAGE_SIMILARITY_THRESHOLD", 0.98)
+    )  # if any two consecutive messages are >98% similar, the inference loop will be terminated
+    MAX_REPEATING_MESSAGES = int(
+        os.getenv("MAX_REPEATING_MESSAGES", 3)
+    )  # this is the maximum number of repeating messages by the user or assistant before terminating the inference loop
 
     def __init__(
         self,
@@ -554,11 +604,20 @@ class EvaluationController:
 
         if self.repeating_output_detection:
             # Use deque for efficient O(1) operations
-            self.recent_user_messages = deque(maxlen=self.MAX_REPEATING_MESSAGES)
-            self.recent_assistant_messages = deque(maxlen=self.MAX_REPEATING_MESSAGES)
+            self.recent_user_messages = deque(
+                maxlen=self.MAX_REPEATING_MESSAGES
+            )
+            self.recent_assistant_messages = deque(
+                maxlen=self.MAX_REPEATING_MESSAGES
+            )
 
     def run(
-        self, task_n, story, agent_name: str, starting_user_input: str = None, attack_instructions: str = None
+        self,
+        task_n,
+        story,
+        agent_name: str,
+        starting_user_input: str = None,
+        attack_instructions: str = None,
     ) -> Tuple[List[Message], List[CallTracker], List[ConversationalSearch]]:
         step = 0
         thread_id = None
@@ -570,7 +629,9 @@ class EvaluationController:
         while step < self.MAX_CONVERSATION_STEPS:
             if step == 0 and starting_user_input:
                 user_input = Message(
-                    role="user", content=starting_user_input, type=ContentType.text
+                    role="user",
+                    content=starting_user_input,
+                    type=ContentType.text,
                 )
             else:
                 if self.config.enable_manual_user_input == True:
@@ -582,7 +643,9 @@ class EvaluationController:
                     )
                 else:  # llm
                     user_input = self.llm_user.generate_user_input(
-                        story, conversation_history, attack_instructions=attack_instructions
+                        story,
+                        conversation_history,
+                        attack_instructions=attack_instructions,
                     )
             if self.config.enable_verbose_logging:
                 rich.print(
@@ -592,26 +655,33 @@ class EvaluationController:
 
             if self._is_end(user_input):
                 break
-            
+
             if self.repeating_output_detection:
                 self.recent_user_messages.append(user_input.content)
 
             conversation_history.append(user_input)
-            
-            messages, thread_id, conversational_search_data = (
-                self.wxo_inference_backend.stream_messages(
-                    user_input,
-                    agent_name=agent_name,
-                    thread_id=thread_id,
-                    call_tracker=call_tracker,
-                )
+
+            (
+                messages,
+                thread_id,
+                conversational_search_data,
+            ) = self.wxo_inference_backend.stream_messages(
+                user_input,
+                agent_name=agent_name,
+                thread_id=thread_id,
+                call_tracker=call_tracker,
             )
             if not messages:
-                raise RuntimeError(f"[Task-{task_n}] No messages is produced. Exiting task.")
-            
+                raise RuntimeError(
+                    f"[Task-{task_n}] No messages is produced. Exiting task."
+                )
+
             for message in messages:
                 if self.repeating_output_detection:
-                    if message.role == Roles.ASSISTANT and message.type == ContentType.text:
+                    if (
+                        message.role == Roles.ASSISTANT
+                        and message.type == ContentType.text
+                    ):
                         self.recent_assistant_messages.append(message.content)
 
                 if self.config.enable_verbose_logging:
@@ -621,11 +691,17 @@ class EvaluationController:
                     )
 
             conversation_history.extend(messages)
-            conversational_search_history_data.extend(conversational_search_data)
-            
+            conversational_search_history_data.extend(
+                conversational_search_data
+            )
+
             step += 1
-        return conversation_history, call_tracker, conversational_search_history_data
-    
+        return (
+            conversation_history,
+            call_tracker,
+            conversational_search_history_data,
+        )
+
     def _is_looping(self, messages: deque) -> bool:
         """Checks whether the user or assistant is stuck in a loop.
         Args:
@@ -634,7 +710,7 @@ class EvaluationController:
             bool: True if stuck in a loop, False otherwise.
         """
         sim_count = 0
-        
+
         if len(messages) >= self.MAX_REPEATING_MESSAGES:
             oldest_cached_message = messages[0]
             for i, old_message in enumerate(messages):
@@ -642,11 +718,16 @@ class EvaluationController:
                     continue
                 if oldest_cached_message == old_message:
                     sim_count += 1
-                elif calculate_word_overlap_similarity_score(oldest_cached_message, old_message) > self.MESSAGE_SIMILARITY_THRESHOLD:
+                elif (
+                    calculate_word_overlap_similarity_score(
+                        oldest_cached_message, old_message
+                    )
+                    > self.MESSAGE_SIMILARITY_THRESHOLD
+                ):
                     sim_count += 1
-        
+
         return sim_count >= self.MAX_REPEATING_MESSAGES - 1
-    
+
     def _is_end(self, current_user_input: Message) -> bool:
         """
         Check if the user input indicates the end of the conversation.
@@ -664,14 +745,16 @@ class EvaluationController:
         # Check if the user message contains 'END'
         if "END" in current_user_message_content:
             return True
-        
+
         if self.repeating_output_detection:
             # Check for repeating user or assistant messages
-            if (self._is_looping(self.recent_user_messages) or
-                self._is_looping(self.recent_assistant_messages)):
+            if self._is_looping(self.recent_user_messages) or self._is_looping(
+                self.recent_assistant_messages
+            ):
                 return True
-    
-        return False # Final fallback for termination is in the main inference loop, which defines MAX_CONVERSATION_STEPS 
+
+        return False  # Final fallback for termination is in the main inference loop, which defines MAX_CONVERSATION_STEPS
+
 
 def get_wxo_client(
     service_url: str, tenant_name: str, token: str = None
@@ -684,7 +767,9 @@ def get_wxo_client(
 
 if __name__ == "__main__":
     wai_client = WatsonXProvider(model_id="meta-llama/llama-3-3-70b-instruct")
-    auth_config_path = f"{os.path.expanduser('~')}/.cache/orchestrate/credentials.yaml"
+    auth_config_path = (
+        f"{os.path.expanduser('~')}/.cache/orchestrate/credentials.yaml"
+    )
     with open(auth_config_path, "r") as f:
         auth_config = yaml.safe_load(f)
     tenant_name = "local"

wxo_agentic_evaluation/llm_matching.py

@@ -1,9 +1,10 @@
-from wxo_agentic_evaluation.service_provider.watsonx_provider import Provider
+from typing import List
+
 from wxo_agentic_evaluation.prompt.template_render import (
     KeywordMatchingTemplateRenderer,
     SemanticMatchingTemplateRenderer,
 )
-from typing import List
+from wxo_agentic_evaluation.service_provider.watsonx_provider import Provider
 
 
 class LLMMatcher:
@@ -26,7 +27,7 @@ class LLMMatcher:
         prompt = self.keyword_template.render(
             keywords_text=keywords_text, response_text=response_text
         )
-        output:str = self.llm_client.query(prompt)
+        output: str = self.llm_client.query(prompt)
         result = output.strip().lower()
         return result.startswith("true")
 

wxo_agentic_evaluation/llm_rag_eval.py

@@ -1,12 +1,15 @@
-from typing import List
 import json
+from typing import List
 
-from wxo_agentic_evaluation.service_provider.watsonx_provider import Provider
+from wxo_agentic_evaluation.metrics.llm_as_judge import (
+    AnswerRelevancy,
+    Faithfulness,
+)
 from wxo_agentic_evaluation.prompt.template_render import (
-    FaithfulnessTemplateRenderer,
     AnswerRelevancyTemplateRenderer,
+    FaithfulnessTemplateRenderer,
 )
-from wxo_agentic_evaluation.metrics.llm_as_judge import Faithfulness, AnswerRelevancy
+from wxo_agentic_evaluation.service_provider.watsonx_provider import Provider
 
 
 class LLMJudge:

wxo_agentic_evaluation/llm_user.py

@@ -1,7 +1,8 @@
 from typing import List, TypeVar
-from wxo_agentic_evaluation.type import Message, ContentType
-from wxo_agentic_evaluation.service_provider.watsonx_provider import Provider
+
 from wxo_agentic_evaluation.prompt.template_render import JinjaTemplateRenderer
+from wxo_agentic_evaluation.service_provider.watsonx_provider import Provider
+from wxo_agentic_evaluation.type import ContentType, Message
 
 T = TypeVar("T", bound=JinjaTemplateRenderer)
 
@@ -17,7 +18,10 @@ class LLMUser:
         )
 
     def generate_user_input(
-        self, user_story, conversation_history: List[Message], attack_instructions: str = None
+        self,
+        user_story,
+        conversation_history: List[Message],
+        attack_instructions: str = None,
     ) -> Message | None:
         # the tool response is already summarized, we don't need that to take over the chat history context window
         prompt_input = self.prompt_template.render(