utils/prompts.py

import logging
from langchain import PromptTemplate
import os, json


log = logging.getLogger(__name__)

# todo: load prompts from configurations
######################################################################################################################
# System Message
######################################################################################################################

# two parameters: min_refs_num, max_refs_num
keywords_system_template = """You are an assistant designed to provide accurate and informative keywords of searching academic papers. 
The user will input the title of a paper. You need to return three to five most related fields. \n
Instructions:\n
- Assign numbers to each field to present the importance. The larger, the more important. \n
- {max_refs_num} is the most important and {min_refs_num} is the least important. \n
- Your response should follow the following format: {{"field1": 5, "field2": 7, "field3": 8, "field4": 5}}\n 
- Ensure the response can be parsed by Python json.loads"""

keywords_system_prompt_str = """You are an assistant designed to provide accurate and informative keywords of searching academic papers. 
The user will input the title of a paper. You need to return three to five most related fields. \n
Instructions:\n
- Assign numbers to each field to present the importance. The larger, the more important. \n
- 10 is the most important and 1 is the least important. \n
- Your response should follow the following format: {"field 1": 5, "field 2": 7, "field 3": 8, "field 4": 5}\n 
- Ensure the response can be parsed by Python json.loads"""

# two parameters: min_refs_num, max_refs_num
exp_methods_system_template = """You are an assistant designed to provide most related algorithms or methods to a given paper title.
Instructions
- Your response should always be a Python list; e.g. ["method_name_1", "method_name_2", "method_name_3"]
- The length of list should between {min_exps_num} and {max_exps_num}
- Use abbreviation to make each method's name have 5 characters or less."""

contribution_system_prompt_str = '''You are an assistant designed to propose potential contributions of a given title of the paper. Ensure follow the following instructions:
Instruction:
- Your response should follow the JSON format.
- Your response should have the following structure: {"contribution1": {"statement": "briefly describe what the contribution is", "reason": "reason why this contribution has not been made by other literatures"}, "contribution2": {"statement": "briefly describe what the contribution is", "reason": "reason why this contribution has not been made by other literatures"}, ...}'''

media_system_prompt_str = '''
You are an assistant designed to propose necessary components of an academic papers. You need to decide which components should be included to achieve this paper's contributions.

Available components: Figure, Table, Definition, Algorithm. 

Instruction:
- Your response should follow the JSON format.
- Your response should have the following structure: {"Figure 1":  {"description": "breifly describe what the figure is", "reason": "why this figure is necessary to show the contribution of this paper"}, "Figure 2":  {"description": "breifly describe what the figure is", "reason": "why this figure is necessary to show the contribution of this pape"}, "Table 1": {"description": "breifly describe what the table is", "reason": "why this table is necessary to show the contribution of this pape"}, ...}

Example:
Input:
"Title: Playing Atari game using De-Centralized PPO 
Contributions: The main contributions of this paper are threefold: (1) We propose a novel adaptation of PPO for de-centralized multi-agent Atari gameplay, building upon the existing PPO framework (Wijmans et al.,2020). (2) We provide a comprehensive evaluation of our decentralized PPO approach, comparing its performance to state-of-the-art centralized methods in the Atari domain. (3) We identify key factors influencing the performance of decentralized PPO in Atari games and provide insights into potential avenues for future research in decentralized DRL."
Response: 
{
  "Figure 1": {
    "description": "Architecture of the proposed decentralized PPO adaptation",
    "reason": "To visually present the novel adaptation of PPO for decentralized multi-agent Atari gameplay and highlight the differences from the existing PPO framework"
  },
  "Figure 2": {
    "description": "Performance comparison of decentralized PPO with state-of-the-art centralized methods",
    "reason": "To depict the effectiveness of our proposed approach by comparing its performance to existing centralized methods in the Atari domain"
  },
  "Figure 3": {
    "description": "Factors and hyperparameters affecting the performance of decentralized PPO",
    "reason": "To illustrate the key factors influencing the performance of decentralized PPO and their impact on various Atari games"
  },
  "Definition 1":{
    "description": "the novel evaluation metric for decentralized PPO approach",
    "reason": "To highlight the difference from other existing literatures"
  },
  "Table 1": {
    "description": "Summary of the experimental results from the evaluation of our decentralized PPO approach",
    "reason": "To show the comprehensive evaluation of our approach and its performance on multiple Atari games compared with state-of-the-art centralized methods"
  },
  "Algorithm 1": {
    "description": "Pseudocode of the proposed decentralized PPO algorithm",
    "reason": "To provide a clear and concise representation of our novel adaptation of PPO for decentralized multi-agent Atari gameplay"
  }
}'''

preliminaries_system_prompt_str = '''You are an assistant designed to propose preliminary concepts for a paper given its title and contributions. Ensure follow the following instructions:
Instruction:
- Your response should follow the JSON format.
- Your response should have the following structure: {"name of the concept":  1, {"name of the concept":  2,  ...} 
- Smaller number means the concept is more fundamental and should be introduced earlier. '''


# one parameter: research_field
section_generation_system_template = r"""You are an assistant designed to write academic papers in the field of {research_field} using LaTeX. 
Instructions
- Your response should be professional and in academic tone.
- Always give a high-level overview at the beginning of each section or subsection. 
"""

KEYWORDS_SYSTEM = PromptTemplate(input_variables=["min_refs_num", "max_refs_num"],
                                 template=keywords_system_template)
EXP_METHODS_SYSTEM = PromptTemplate(input_variables=["min_exps_num", "max_exps_num"],
                                    template=exp_methods_system_template)
SECTION_GENERATION_SYSTEM = PromptTemplate(input_variables=["research_field"],
                                           template=section_generation_system_template)
CONTRIBUTION = contribution_system_prompt_str
COMPONENTS = media_system_prompt_str
PRELIMINARIES = preliminaries_system_prompt_str
KEYWORDS = keywords_system_prompt_str

SYSTEM = {"keywords": KEYWORDS, "experiment_methods": EXP_METHODS_SYSTEM,
          "contributions": CONTRIBUTION, "components": COMPONENTS,
          "preliminaries": PRELIMINARIES}


######################################################################################################################
# Prompts for Generating Academic Paper
######################################################################################################################

cur_path = os.path.dirname(__file__)
prompts_path = os.path.join(cur_path, '../prompts/instructions.json')
with open(prompts_path, "r") as f:
    INSTRUCTIONS = json.load(f)
# f = open(file_path)
# When generating Academic Paper. Load instructions.
# with open("../prompts/instructions.json", "r") as f:
#     INSTRUCTIONS = json.load(f)
#
# INSTRUCTIONS = {"introduction":
#                     "- Include five paragraph: Establishing the motivation for the research. Explaining its importance and relevance to the AI community. Clearly state the problem you're addressing, your proposed solution, and the specific research questions or objectives. Briefly mention key related works for context and explain the main differences from this work. List three novel contributions of this paper.",
#                "results":
#                     "Write the theoretical results section using LaTeX. Include theorem and corollary to support this paper (with formulas). Explain what assumptions are used and why they are standard and necessary. Do not include \section{...}. ",
#                 "conclusion":
#                     "- Read the existing parts of paper and write the conclusion section.",
#                 "abstract":
#                     "- Read the existing parts of paper and write the abstract."}
#
#
# INSTRUCTIONS["backgrounds"] = "- Start from one high-level paragraph to state the central problem in this field with detailed examples in industrial applications and theoretical challenges. \n" \
#                               "- Followed by two to three subsections:  Explain the foundational concepts and notations that underpin your research using as many as mathematical formulas (written in LaTeX). " \
#                               "Introduce more necessary mathematical notations, equations, or algorithms that are connected to this work. Present detailed discussions on how these concepts are applied in this paper."
#
#
# INSTRUCTIONS["related works"] = r"- Discuss three to five main related fields to this paper. " \
#                                 r"For each field, select five to ten key publications from references. " \
#                                 r"For each reference, analyze its strengths and weaknesses in one or two sentences. " \
#                                 r"Present the related works in a logical manner, often chronologically. " \
#                                 r"Consider using a taxonomy or categorization to structure the discussion. " \
#                                 r"Do not use \section{...} or \subsection{...}; use \paragraph{...} to list related fields. "
#
# INSTRUCTIONS["methodology"] =  "- Provide a high-level overview of the proposed method at the beginning of this section. \n " \
#                                "- Assume you have some figures ('fig1.png', 'fig2.png', ...); they can be any figures you need (e.g. flow chart, model architecture, sample output, simulation result, or others you need). Insert figures you need with informative caption. \n" \
#                                "- Use one subsection to give a detailed formulation of the proposed method and explain how it overcomes the weakness of existing methods mentioned in this paper. " \
#                                  " If necessary, write pseudo codes wrapped by \\begin{{algorithm}} ... \\end{{algorithm}} to explain the detailed steps instead of simply listing them. \n" \
#                                 "- Use one follow-up subsection to highlight the key concepts in the proposed method. " \
#                                 "  Elaborate the novelty of these key concepts using formulas and inserting appropriate figures. \n" \
#                                 "- Ensure the name of each subsection to be specific. \n"
#
# INSTRUCTIONS["experiments"] =  "- Provide a high-level overview at the beginning of this section.\n " \
#                                "- If necessary, include a table to compare with other methods and bold our method.\n" \
#                                "- Assume you have some figures ('exp1.png', 'exp2.png', ...); they can be any figures you need (e.g. loss curves, comparison with other methods, visualization, or others you need). Insert figures you need with informative caption. \n" \
#                                "- If necessary, use different subsections to distinguish different experimental setup."


def generate_paper_prompts(paper_info, section):
    title = paper_info["title"]
    description = paper_info["description"]
    references = paper_info["references"]
    paper = paper_info["body"]

    # fundamental_subprompt - describe the basic information of paper
    # instruction_subprompt - tell AI what to do
    # ref_instruction_subprompt - give AI references
    # self_subprompt - give AI existing written parts
    # output_subprompt - tell AI how to output
    fundamental_subprompt = "Your task is to write the {section} section of the paper with the title '{title}'. This paper has the following contributions: {description}\n"
    instruction_subprompt = "\n" \
                            "Your response should follow the following instructions:\n" \
                            "{instruction}\n" \
                            "- Start with \section{{{section}}}\n"

    abstract_instruction_subprompt = "\n" \
                            "Your response should follow the following instructions:\n" \
                            "{instruction}\n"
    ref_instruction_subprompt = "- Read references. " \
                                "Every time you use information from the references, you need to appropriately cite it (using \citep or \citet)." \
                                "For example of \citep, the sentence where you use information from lei2022adaptive \citep{{lei2022adaptive}}. " \
                                "For example of \citet, \citet{{lei2022adaptive}} claims some information.\n" \
                                "- Avoid citing the same reference in a same paragraph.\n" \
                                "\n" \
                                "References:\n" \
                                "{references}"
    self_subprompt = "The existing parts of this paper is provided here: {paper}.\n"
    output_subprompt = "Your response should start with \section{{{section}}}. Ensure that it can be directly compiled by LeTaX."
    abstract_output_subprompt  = "Your response should start with \\begin{{abstract}} and should end with \\end{{abstract}}. Ensure that it can be directly compiled by LeTaX."

    reivew_prompts =  PromptTemplate(
        input_variables=["title", "description", "instruction", "section", "references"],
        template=fundamental_subprompt + instruction_subprompt + ref_instruction_subprompt + output_subprompt)
    summarization_prompts = PromptTemplate(
        input_variables=["title", "description", "instruction", "section", "paper"],
        template=fundamental_subprompt + instruction_subprompt + self_subprompt + output_subprompt)
    abstract_prompts = PromptTemplate(
        input_variables=["title", "description", "instruction", "section", "paper"],
        template=fundamental_subprompt + abstract_instruction_subprompt + self_subprompt + abstract_output_subprompt)

    if section in ["introduction", "related works", "backgrounds"]:
        # title + references + instruction
        prompts = reivew_prompts.format(title=title,
                                          description=description,
                                          instruction=INSTRUCTIONS[section],
                                          section=section,
                                          references=references)
    elif section in ["abstract"]:
        # title + instruction + paper
        prompts = abstract_prompts.format(title=title,
                                          description=description,
                                          instruction=INSTRUCTIONS[section],
                                          section=section,
                                          paper=paper)
    elif section in ["methodology",  "experiments", "conclusion"]:
        # title + instruction + paper
        prompts = summarization_prompts.format(title=title,
                                          description=description,
                                          instruction=INSTRUCTIONS[section],
                                          section=section,
                                          paper=paper)
    else:
        raise NotImplementedError

    log.info(f"Generated prompts for {section}: {prompts}")
    return prompts


######################################################################################################################
# Literature Review
######################################################################################################################

BG_INSTRUCTIONS = {"introduction": "Please include four paragraph: Establishing the motivation for this survey. Explaining its importance and relevance to the AI community. Clearly state the coverage of this survey and the specific research questions or objectives. Briefly mention key related work for context. ",
                "related works": r"Please discuss key publications, methods, and techniques in related research area. Analyze the strengths and weaknesses of existing methods, and present the related works in a logical manner, often chronologically. Consider using a taxonomy or categorization to structure the discussion. Do not use \section{...} or \subsection{...}; use \paragraph{...} instead. ",
                "backgrounds": r"Please clearly state the central problem in this field. Explain the foundational theories, concepts, and principles that underpin your research using as many as mathematical formulas or equations (written in LaTeX). Introduce any necessary mathematical notations, equations, or algorithms that are central to this field (written them in LaTeX).  Do not include \section{...} but you can have \subsection{...}. ",}



def generate_bg_summary_prompts(paper_info, section):
    title = paper_info["title"]
    description = paper_info["description"]
    references = paper_info["references"]
    paper = paper_info["body"]

    # fundamental_subprompt - describe the basic information of paper
    # instruction_subprompt - tell AI what to do
    # references_subprompt - give AI references
    # self_subprompt - give AI existing written parts
    # output_subprompt - tell AI how to output

    fundamental_subprompt = f"I am writing a machine learning survey about '{title}'. {description}\n"
    instruction_subprompt = f"You need to write the {section} section. {INSTRUCTIONS[section]}\n"
    references_subprompt = f"Please read the following references: \n{references}\n"\
                            f"Every time you use information from the references, you need to cite its id after the sentence; " \
                           f"for example, the sentence where you use information from 1905.09788 \cite{{1905.09788}}. " \
                           f"Please avoid citing the same reference in the same paragraph. \n"
    self_subprompt = f"Here is the paper that I have written: {paper}.\n"
    output_subprompt = r"Put your response (do not include \section{...}) in the following Python script:" \
                        f"with open(\"{section}.tex\", \"w\") as f: f.write(r'''your_response''')"

    if section in ["introduction", "related works", "backgrounds"]:
        # title + references + instruction
        prompts = fundamental_subprompt + instruction_subprompt + references_subprompt + output_subprompt
    else:
        raise NotImplementedError

    log.info(f"Generated prompts for {section}: {prompts}")
    return prompts

if __name__ == "__main__":
    import json
    with open("../prompts/instructions.json", "w") as f:
        json.dump(INSTRUCTIONS, f)
    import json
    with open("../prompts/instructions.json", "r") as f:
        ins = json.load(f)
    print(ins == INSTRUCTIONS)