src/tools/llm_tools.py

import json
import string

import wikipedia
from langchain import PromptTemplate
from langchain.vectorstores import Chroma
from langchain.text_splitter import CharacterTextSplitter

from src.tools.llms import openai_llm
from src.tools.wiki import Wiki




def get_wikilist(task: {}) -> str:
    """
    get the titles of wiki pages interesting for solving the given task
    """

    llm = openai_llm
    template = (f"\n"
                f"    Your task consists in finding the list of wikipedia page titles which provide useful content "
                f"    for a paragraph whose description is delimited by triple backticks: ```{task['description']}```\n"
                f"    \n"
                f"    The paragraph belongs at the top level of the hierarchy to a document"
                f"    whose description is delimited by triple backticks: ``` {task['doc_description']}```\n"
                f"    Make sure that the paragraph relates the top level of the document\n"
                f"    \n"
                f"    The paragraph belongs to a higher paragraph in the hierarchy \\n"
                f"    whose description is delimited by triple backticks: ``` {task['above']}```\n"
                f"    Make sure that the paragraph relates with the paragraph in the hierarchy of the document\n"
                f"        \n"
                f"    The paragraphs comes after previous paragraphs \\n"
                f"    whose description is delimited by triple backticks: ``` {task['before']}```\n"
                f"    Make sure that the paragraph relates with previous paragraph without any repetition\n"
                f"    \n"
                f"    The paragraphs comes before next paragraphs \\n"
                f"    whose description is delimited by triple backticks: ``` {task['after']}```\n"
                f"    \n"
                f"    Format your response as a JSON list of strings separated by commas.\n"
                f"  \n"
                f"\n"
                f"    ")

    prompt = PromptTemplate(
        input_variables=[],
        template=template
    )

    #wikilist = LLMChain(llm=openai_llm, prompt=prompt).run()
    llm_list = llm(template)
    wikilist = extract_list(llm_list)

    expanded_wikilist = []

    expand_factor = 2

    for wikipage in wikilist:
        expanded_wikilist += wikipedia.search(wikipage, expand_factor)

    wikilist = list(set(expanded_wikilist))

    return wikilist


def extract_list(llm_list: str):
    print(llm_list)

    def filter_(el: str):
        resp = 2 < len(el)
        usable_length = len([c for c in el if c in string.ascii_letters])
        resp = resp and len(el)*3/4 < usable_length
        return resp

    try:
        wikilist = llm_list[1:-1].split('"')
        wikilist = [el for el in wikilist if filter_(el)]
        print(wikilist)
    except:
        wikilist = []
        print('issues with the wikilist')
    return wikilist


def get_public_paragraph(task: {}) -> str:
    """returns the task directly performed by chat GPT"""

    llm = openai_llm
    template = (f"\n"
                f"    Your task consists in generating a paragraph\\n"
                f"    whose description is delimited by triple backticks: ```{task['description']}```\n"
                f"\n"
                f"    The paragraph belongs at the top level of the hierarchy to a document \\n"
                f"    whose description is delimited by triple backticks: ``` {task['doc_description']}```\n"
                f"    Make sure that the paragraph relates the top level of the document\n"
                f"    \n"
                f"    The paragraph belongs to a higher paragraph in the hierarchy \\n"
                f"    whose description is delimited by triple backticks: ``` {task['above']}```\n"
                f"    Make sure that the paragraph relates with the paragraph in the hierarchy of the document\n"
                f"        \n"
                f"    The paragraphs comes after previous paragraphs \\n"
                f"    whose description is delimited by triple backticks: ``` {task['before']}```\n"
                f"    Make sure that the paragraph relates with previous paragraph without any repetition\n"
                f"    \n"
                f"    The paragraphs comes before next paragraphs \\n"
                f"    whose description is delimited by triple backticks: ``` {task['after']}```\n"
                f"    Make sure that the paragraph prepares the transition to the next paragraph without any repetition\n"
                f"    \n"
                f"  \n"
                f"\n"
                f"    ")

    p = llm(template)

    return p


def create_index(wikilist: [str]):
    """
    useful for creating the index of wikipages
    """
    fetch = Wiki().fetch

    pages = [(title, fetch(title)) for title in wikilist if type(fetch(title)) != str]
    texts = []
    chunk = 800
    for title, page in pages:
        texts.append(WikiPage(title=title, fulltext=page.page_content))

    doc_splitter = CharacterTextSplitter(
        separator=".",
        chunk_size=chunk,
        chunk_overlap=100,
        length_function=len,
    )

    paragraphs = texts[0].get_paragraphs(chunk=800)

    split_texts = []
    for p in paragraphs:
        split_texts += doc_splitter.split_text(p)

    for split_text in split_texts:
        assert type(split_text) == str
        assert 0 < len(split_text) < 2 * 500

    wiki_index = Chroma.from_texts(split_texts)

    return wiki_index


def get_wiki_paragraph(wiki_index, task: {}) -> str:
    """useful to get a summary in one line from wiki index"""

    task_description = get_public_paragraph(task)
    wiki_paragraphs = semantic_search(wiki_index, task_description)
    text_content = ""
    for p in wiki_paragraphs:
        text_content += p.page_content + "/n/n"

    template = (f"\n"
                f"    Your task consists in generating a paragraph\\n"
                f"    whose description is delimited by triple backticks: ```{task['description']}```\n"
                f"\n"
                f"    The text generation is based in the documents provided in these sections \n"
                f"    delimited by by triple backticks: ``` {text_content}``` \n"
                f"    The paragraph belongs at the top level of the hierarchy to a document \\n"
                f"    whose description is delimited by triple backticks: ``` {task['doc_description']}```\n"
                f"    Make sure that the paragraph relates the top level of the document\n"
                f"    \n"
                f"    The paragraph belongs to a higher paragraph in the hierarchy \\n"
                f"    whose description is delimited by triple backticks: ``` {task['above']}```\n"
                f"    Make sure that the paragraph relates with the paragraph in the hierarchy of the document\n"
                f"        \n"
                f"    The paragraphs comes after previous paragraphs \\n"
                f"    whose description is delimited by triple backticks: ``` {task['before']}```\n"
                f"    Make sure that the paragraph relates with previous paragraph without any repetition\n"
                f"    \n"
                f"    The paragraphs comes before next paragraphs \\n"
                f"    whose description is delimited by triple backticks: ``` {task['after']}```\n"
                f"    Make sure that the paragraph prepares the transition to the next paragraph without any repetition\n"
                f"    \n"
                f"  \n"
                f"\n"
                f"    ")

    llm = openai_llm
    p = llm(template)

    return p


def get_private_paragraph(texts, task: {}) -> str:
    """useful to get a summary in one line from wiki index"""

    text_content = ""
    for t in texts:
        text_content += t + "/n/n"

    template = (f"\n"
                f"    Your task consists in generating a paragraph\\n"
                f"    whose description is delimited by triple backticks: ```{task['description']}```\n"
                f"\n"
                f"    The text generation is based in the documents provided in these sections \n"
                f"    delimited by by triple backticks: ``` {text_content}``` \n"
                f"    The paragraph belongs at the top level of the hierarchy to a document \\n"
                f"    whose description is delimited by triple backticks: ``` {task['doc_description']}```\n"
                f"    Make sure that the paragraph relates the top level of the document\n"
                f"    \n"
                f"    The paragraph belongs to a higher paragraph in the hierarchy \\n"
                f"    whose description is delimited by triple backticks: ``` {task['above']}```\n"
                f"    Make sure that the paragraph relates with the paragraph in the hierarchy of the document\n"
                f"        \n"
                f"    The paragraphs comes after previous paragraphs \\n"
                f"    whose description is delimited by triple backticks: ``` {task['before']}```\n"
                f"    Make sure that the paragraph relates with previous paragraph without any repetition\n"
                f"    \n"
                f"    The paragraphs comes before next paragraphs \\n"
                f"    whose description is delimited by triple backticks: ``` {task['after']}```\n"
                f"    Make sure that the paragraph prepares the transition to the next paragraph without any repetition\n"
                f"    \n"
                f"  \n"
                f"\n"
                f"    ")

    llm = openai_llm
    p = llm(template)

    return p