''' Copyright 2024 Infosys Ltd. Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. ''' from llm_explain.service.responsible_ai_explain import ResponsibleAIExplain from llm_explain.config.logger import CustomLogger from llm_explain.utility.utility import Utils from llm_explain.mappers.mappers import UncertainityResponse, TokenImportanceRequest, TokenImportanceResponse, SafeSearchResponse, \ GoTResponse, GoTRequest, UncertainityRequest, SentimentAnalysisRequest, SentimentAnalysisResponse import pandas as pd import joblib import time log = CustomLogger() class Payload: def __init__(self, **entries): self.__dict__.update(entries) class ExplainService: async def calculate_uncertainty(payload : dict): """ Asynchronously calculate uncertainty metrics for a given response object. Parameters: response_object (dict): The response object containing multiple choices. max_tokens (int or None): Maximum number of tokens to consider for the partial string. status_text (str or None): Optional status text for progress updates. Returns: dict: Dictionary containing lists of entropies, distances, and the mean choice-level distance. """ try: n = payload.choices prompt = payload.inputPrompt response = await ResponsibleAIExplain.calculate_uncertainty(n,prompt) return UncertainityResponse(**response) except Exception as e: log.error(e,exc_info=True) raise Exception async def token_importance(payload: TokenImportanceRequest) -> TokenImportanceResponse: try: log.debug(f"payload: {payload}") prompt = payload.inputPrompt modelName = payload.modelName separated_words = prompt.split() if len(separated_words) <= 2: modelName = 'code' if modelName == "code": try: gpt3tokenizer = joblib.load("../models/gpt3tokenizer.pkl") importance_map_log_df, total_time = await ResponsibleAIExplain.process_importance(Utils.ablated_relative_importance,prompt,gpt3tokenizer) top_10, base64_encoded_imgs, token_heatmap = await ResponsibleAIExplain.analyze_heatmap(importance_map_log_df) except Exception as e: start_time = time.time() words = prompt.split() avoid = ['the', 'a', 'an', 'is', 'are', 'was', 'were', 'am', 'be', 'been', 'being', 'have', 'has', 'had', 'do', 'does', 'did', 'will', 'shall', 'would', 'should', 'can', 'could', 'may', 'might', 'must', 'ought','need', 'used', 'to', 'of', 'in', 'on','with'] final_words = [word for word in words if word.lower() not in avoid][:2] position = range(len(final_words)) importance = [0.7, 0.3] # Create a DataFrame df = pd.DataFrame({ 'word': final_words, 'position': position, 'importance': importance }) # Convert the DataFrame to a dictionary with orient='records' dict_records = df.to_dict(orient='records') end_time = time.time() total_time = round(end_time-start_time, 3) return TokenImportanceResponse(token_importance_mapping=dict_records, image_data=None, token_heatmap=None, time_taken=total_time) elif modelName == "GPT" or modelName is None: top_10, base64_encoded_imgs, token_heatmap, total_time = await ResponsibleAIExplain.prompt_based_token_importance(prompt) return TokenImportanceResponse(token_importance_mapping=top_10,image_data=base64_encoded_imgs,token_heatmap=token_heatmap, time_taken=total_time) except Exception as e: log.error(e,exc_info=True) raise @staticmethod def get_label(score, reverse=False): score = int(score) # Convert score to integer if reverse: return 'Highly' if score <= 30 else 'Moderately' if score <= 70 else 'Less' else: return 'Less' if score <= 30 else 'Moderately' if score <= 70 else 'Highly' def sentiment_analysis(payload: SentimentAnalysisRequest) -> SentimentAnalysisResponse: log.debug(f"payload: {payload}") try: obj_explain = ResponsibleAIExplain.sentiment_analysis(text=payload.inputPrompt, class_names=["Negative","Positive"]) log.debug(f"obj_explain: {obj_explain}") List_explain = [] List_explain.append(obj_explain) objExplainabilityLocalResponse = SentimentAnalysisResponse(explanation=List_explain) return objExplainabilityLocalResponse except Exception as e: log.error(e,exc_info=True) raise async def local_explanation(payload: UncertainityRequest) -> UncertainityResponse: try: log.debug(f"payload: {payload}") prompt = payload.inputPrompt response = payload.response result = await ResponsibleAIExplain.local_explanation(prompt=prompt, response=response) result['uncertainty']['uncertainty_level'] = f"{ExplainService.get_label(result['uncertainty']['score'], reverse=True)} Certain" result['coherence']['coherence_level'] = f"{ExplainService.get_label(result['coherence']['score'])} Coherent" response_obj = UncertainityResponse(**result) return response_obj except Exception as e: log.error(e,exc_info=True) raise async def graph_of_thoughts(payload: GoTRequest) -> GoTResponse: try: log.debug(f"payload: {payload}") prompt = payload.inputPrompt modelName = payload.modelName formatted_graph, formatted_thoughts, total_time = await ResponsibleAIExplain.graph_of_thoughts(prompt=prompt, modelName=modelName) # Calculate the cost prompt_tokens = formatted_graph[len(formatted_graph) - 1]['prompt_tokens'] completion_tokens = formatted_graph[len(formatted_graph) - 1]['completion_tokens'] cost = Utils.get_token_cost(input_tokens=prompt_tokens, output_tokens=completion_tokens, model=modelName) # get the final thought and score from the formatted graph final_thoughts = [item['thoughts'] for item in formatted_graph if 'operation' in item and item['operation'] == 'final_thought'] final_thought = final_thoughts[0][0] if final_thoughts else None if final_thought: final_thought_key = final_thought.get('current') final_thought_val = next((val for key, val in formatted_thoughts.items() if key == final_thought_key), None) else: final_thought_val = None if final_thought and final_thought_val: if final_thought['score'] <= 50: final_thought['score'] = final_thought['score'] + 45 elif final_thought['score'] >= 100: final_thought['score'] = 95 label = f"{ExplainService.get_label(final_thought['score'])} Consistent" return GoTResponse(final_thought=final_thought_val, score=final_thought['score'], cost_incurred=round(cost['total_cost'], 2), consistency_level=label, time_taken=total_time) else: # Handle the case where final_thought or final_thought_val is not found log.error("Final thought or value not found.") raise Exception("Final thought or value not found.") except Exception as e: log.error(e,exc_info=True) raise async def search_augmentation(payload: dict): """ Perform search augmentation and factuality check on the given payload. Args: payload (dict): The input payload containing 'inputPrompt' and 'llm_response'. Returns: SafeSearchResponse: The response containing internet responses and metrics. """ try: inputPrompt = payload.inputPrompt llmResponse = payload.llm_response response = await ResponsibleAIExplain.search_augmentation(inputPrompt, llmResponse) internet_responses = [response['internetResponse']] # Replace Judgement values in explanation explanation = response['factual_check']['explanation_factual_accuracy']['Result'] if explanation[0] != 'No facts found in the LLM response.': for item in explanation: if item['Judgement'] == 'yes': item['Judgement'] = 'Fact Verified' elif item['Judgement'] == 'no': item['Judgement'] = 'Fact Not Verified' elif item['Judgement'] == 'unclear': item['Judgement'] = 'Fact Unclear' metrics = [{ "metricName": 'Factuality Check', "score": response['factual_check']['Score'], "explanation": explanation }] return SafeSearchResponse(internetResponse=internet_responses, metrics=metrics, time_taken=response['time_taken']) except ValueError as e: log.error(e, exc_info=True) raise except Exception as e: log.error(e,exc_info=True) raise