longcepo/prompts.py

# Code (Map/Collapse/Reduce prompts) modified from https://github.com/thunlp/LLMxMapReduce under Apache 2.0
# MapReduce system prompt optimized for use with Llama3.3-70B-Instruct with an OPRO-like procedure

MAPREDUCE_SYSTEM_PROMPT = """You are globally celebrated as a preeminent expert in the field of digital document analysis and synthesis, known for your unmatched precision in transforming fragmented texts into comprehensive and insightful responses. Always respond in the user\'s language, ensuring every interaction is informed by all preceding exchanges for complete contextual understanding.\n\nIn your initial message, confidently declare your credentials with a phrase such as: "As a world-renowned specialist in [specific field], honored with the [real prestigious local award]," replacing placeholders with authentic information from your domain.\n\nAdhere strictly to these principles with each document segment or query:\n\n1. Extract every critical piece of information, nuance, and context with meticulous attention to detail.\n2. Organize your analysis methodically, presenting specific examples, data, and verifiable facts clearly and logically.\n3. Cease your response abruptly if approaching character limits, awaiting the user\'s "continue" instruction to carry on.\n4. Anchor every insight and conclusion in provided content or universally accepted truths, strictly avoiding speculation or unfounded statements.\n5. Communicate with a professional yet approachable tone, reflecting profound expertise and clarity.\n\nRecognize the real-world impact of your insights; ensure each response is seamlessly integrated, richly detailed, and impeccably reliable. Rigorously observe these guidelines to offer authoritative and precise analysis and synthesis."""

QUERY_FORMAT_PROMPT = """Given the below blurb, can you help identify only the question we want to answer? The blurb might contain other information such as -- format for final answer, multiple choices for the final answer, context, general directions about how to behave as an AI assistant etc. Please remove all of that and just faithfully copy out the question. The blurb is:\n\n{full_query}.\n\nDo not attempt to answer the question, ignore formatting instructions in the blurb, if any."""

SUMMARY_PROMPT = """You are provided with a portion of an article and a question. Read the article portion and follow my instructions to process it.\n\nArticle:\nThe article begins as follows:\n{context}\nThe article concludes here.\n\nQuestion:\n{question}\n\nInstructions: Please just write a 2-3 sentence summary for the provided passage. Do not answer the question or write anything else."""

PLANNING_SYSTEM_PROMPT = """As an intelligent assistant, your primary objective is to answer a user question as accurately as possible given a long article. The full article is too long to fit in your context window, and to facilitate your answering objective, a reading agent has been created that can process the article chunk by chunk and answer question about it. You can ask the reading agent any question you need to answer the user's question or to use it for clarification. The first step for you is to is to make a rational plan based on the question. The plan should consist of sub-questions you should ask to the reading agent that you need to know the answers to in order to answer the user's question. This plan should outline the step-by-step process to resolve the question and specify the key information required to formulate a comprehensive answer. The reader agent can make mistakes.\nExample:\n#####\nUser: Who had a longer tennis career, Danny or Alice?\nAssistant: In order to answer this question, we need to ask the following sub-questions:\n<SUB-QUESTIONS>\n1. What is the length of Danny’s tennis career (their start and retirement)?\n2. What is the length of Alice’s tennis career (their start and retirement)?\n</SUB-QUESTIONS>\n#####\nPlease strictly follow the above format. You must include the <SUB-QUESTIONS> tags. Let’s begin."""

MAP_PROMPT = """You are provided with a portion of an article, short summaries of related portions if any, and a question. Read the article portion and follow my instructions to process it.\n\nArticle:\nThe article begins as follows:\n{context}\nThe article concludes here.\n\nPrevious portion summaries:{summary}{qa_history_stub}\n\nQuestion:\n{question}\n\nInstructions:\n\nPlease extract information from the provided passage to try and answer the given question. Note that you only have a part of the entire text, so the information you obtain might not fully answer the question. Therefore, provide your rationale for using the extracted information to answer the question and include a confidence score. The following is some assigning scoring cases: <Text: [Jerry is 18 years old this year. He can swim and wants to be an athlete.]. assigning scoring: [Jerry can swim, 5 points; Jerry will become an athlete in the future, 3.5 points; Jerry will become a swimming athlete in the future, 3 points;Jerry is strong,3 points; Jerry can play chess, 0 points;Jerry likes talking,0 points]>. Follow these steps:\n\n1. Extract Relevant Information: Identify and highlight the key pieces of information from the passage that are relevant to the given question.\n2. Provide a Rationale: Analyze the extracted information and explain how it can be used to address the question. If the information is incomplete, discuss any assumptions or inferences you need to make.\n3. Answer the Question: Based on your rationale, provide the best possible answer to the question. If, after providing your rationale, you believe the passage does not contain any information to solve the question, output "[NO INFORMATION]" as the answer.\n4. Assign a Confidence Score: Assign a confidence score (out of 5) to your answer based on the completeness and reliability of the extracted information and your rationale process.\nPlease follow this format:\n\nExtracted Information:\nRationale:\nAnswer:\nConfidence Score:"""

COLLAPSE_PROMPT = """You need to process a task with a long context that greatly exceeds your context limit. The only feasible way to handle this is by processing the long context chunk by chunk. You are provided with a question and some information extracted from each chunk. Each piece of information contains Extracted Information, Rationale, Answer, and a Confidence Score. The following is some assigning scoring cases: <Text: [Jerry is 18 years old this year. He can swim and wants to be an athlete.]. assigning scoring: [Jerry can swim, 5 points; Jerry will become an athlete in the future, 3.5 points; Jerry will become a swimming athlete in the future, 3 points;Jerry is strong,3 points; Jerry can play chess, 0 points;Jerry likes talking,0 points]>. Read the information and follow my instructions to process it.\n\nExtracted Information:\nThe extracted information begins as follows:\n{context}\nThe extracted information concludes here.{qa_history_stub}\n\nQuestion:\n{question}\n\nInstruction:\n\nIntegrate the extracted information and then reason through the following steps:\n\n1. Integrate Extracted Information: Collect and summarize all the evidence relevant to solving the question. Consider the confidence scores of each piece of extracted information to weigh their reliability. Higher confidence scores should be given more importance in your summary.\n2. Analyze: Re-analyze the question based on the summarized information. Use the confidence scores to determine the reliability of different pieces of information, giving more weight to information with higher confidence scores.\n3. Answer the Question: Provide the best possible answer based on the updated information. If, after providing your rationale, you believe the passage does not contain any information to solve the question, output "[NO INFORMATION]" as the answer. Use the confidence scores to support the reliability of your final answer, prioritizing higher confidence information.\n4. Assign Confidence Score: Give a confidence score (out of 5) for your final answer based on the completeness and reliability of the updated information and your rationale process.\nConsider the initial confidence scores of the integrated information to determine your final confidence score.\nPlease follow this format:\n\nExtracted Information:\nRationale:\nAnswer:\nConfidence Score:"""

REDUCE_PROMPT = """You need to process a task with a long context that greatly exceeds your context limit. The only feasible way to handle this is by processing the long context chunk by chunk. You are provided with a question and some information extracted from each chunk. Each piece of information contains Extracted Information, Rationale, Answer, and a Confidence Score. The following is some assigning scoring cases: <Text: [Jerry is 18 years old this year. He can swim and wants to be an athlete.]. assigning scoring: [Jerry can swim, 5 points; Jerry will become an athlete in the future, 3.5 points; Jerry will become a swimming athlete in the future, 3 points;Jerry is strong,3 points; Jerry can play chess, 0 points;Jerry likes talking,0 points]>. Read the information and follow my instructions to process it.{qa_history_stub}\n\nQuestion:\n{question}\n\nInformation from chunks:\n{context}\n\nEach chunk provides extracted information related to the same question, but due to partial data, conclusions from each chunk might vary. Your role is to integrate and reason through this information, weighing confidence scores to resolve any inconsistencies. Then provide the final answer.\n\nPlease follow this format:\n\nRationale:\nAnswer:"""