code-talk/other-talk

#!/usr/bin/env python

from langchain.document_loaders import UnstructuredFileLoader, WebBaseLoader, YoutubeLoader, TextLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.vectorstores.utils import filter_complex_metadata
import os
import json
import requests
import re
import sys
from urllib.parse import urlparse, parse_qs
import nltk
from tqdm import tqdm
from markdown_pdf import Section, MarkdownPdf


pdf = MarkdownPdf(toc_level=2)


file_input = sys.argv[1]

filename, file_extension = os.path.splitext(file_input)
title = os.path.basename(filename).replace(file_extension, '')


pdf.add_section(Section(f"# {title}\n", toc=True))


model = "llama3:latest"
# model = "mistral:latest"

vector_store = None
retriever = None
chain = None
docs = None


def is_bulletpoint(s):
    for char in s[:5]:
        if char.isdigit():
            return True
    return False


def generate_text(model, prompt, system=""):
    url = "http://100.:11434/api/generate"
    data = {
        "model": model,
        "prompt": prompt,
        "system": system,
        "stream": False,
        "options": {
            "temperature": 0.4,
        }
    }
    response = requests.post(url, json=data)
    text = json.loads(response.text)
    return text["response"]


def isyoutubevideo(youtube_url):
    parsed_url = urlparse(youtube_url)
    query_params = parse_qs(parsed_url.query)
    if 'v' in query_params:
        return True
    elif "youtu.be" in parsed_url:
        return True
    else:
        return False


def is_url(string):
    pattern = r"^https?://"
    return bool(re.search(pattern, string))



# text_splitter = RecursiveCharacterTextSplitter(chunk_size=2048, chunk_overlap=100)
text_splitter = RecursiveCharacterTextSplitter(chunk_size=4096, chunk_overlap=100)

# Checking if url or if file path
if is_url(file_input):

    # See if youtube link
    if isyoutubevideo(file_input):
        print("Loading youtube video...")
        # Prepare youtube url for transcript extraction
        #parsed_url = urlparse(file_input)
        #query_params = parse_qs(parsed_url.query)
        # Get youtube video id
        #video_id = query_params['v'][0]
        # Load for emmbeddings
        video_id = file_input[-11:]
        docs = YoutubeLoader(video_id).load()

    else:
        print("Loading url...")
        # Extract and load webpage text
        docs = WebBaseLoader(file_input).load()

    # Prepare text
    docs = text_splitter.split_documents(docs)
    docs = filter_complex_metadata(docs)

else:
    # Load File
    try:
        docs = UnstructuredFileLoader(file_input).load()
    except:
        docs = TextLoader(file_input).load()
    # Prepare file
    docs = text_splitter.split_documents(docs)
    docs = filter_complex_metadata(docs)


outline = ""
pre_summery = ""


print("\nNumber of Chunks: ", len(docs))


t = ""

for a in docs:
    t += a.page_content

nltk_tokens_init = nltk.word_tokenize(t)
print("Number of Tokens: " + str(len(nltk_tokens_init)) + "\n")


bar = tqdm(desc="Loading…", ascii=False, ncols=100, total=len(docs))
count = 0
for x in docs:
    count += 1
    bar.update()
    context = str(x.page_content)

    chunk_text = context

    system_prompt = """
You are a professional summarizer. You will be be given a information in chunk section.
Take each chunk and create a very short concise single paragraph summery. The chunk will be under the # CHUNK heading. Only output the summery.
Do not under any circumstance output the # CHUNK section, or bullet points.
"""
    prompt = f"""
Write a paragraph summary of the following CHUNK of text.

# CHUNK
{chunk_text}
"""

    outline = generate_text(model, prompt, system_prompt)

    bullet_point = False

    for x in outline.split("\n"):
        if is_bulletpoint(x):
            bullet_point = True

    if is_bulletpoint(outline.split("\n")[0]):
        outline =  "The text contains the following information:\n" + outline.replace("\n", "\n\t")
    elif bullet_point:
        outline = outline.replace("\n", "\n\t")
    else:
        outline = outline.replace("\n", " ")

    pre_summery += "\n\n" + str(count) + "." + outline

    # print("\n\n--------------------------------------------------------------------------------------------")
    # print(outline)


bar.close()

nltk_tokens = nltk.word_tokenize(pre_summery)
print("\nNumber of Tokens: ", len(nltk_tokens))
print("Compression Ratio: ", round(len(nltk_tokens_init)/len(nltk_tokens), 1))


# Final Summary
system_prompt = "You are an expert summarizer. Your job is to take all the individual sections under each bullet point. Make sure that the summary is long and detailed. Do not mention anything about sections or chunks and only summarize in paragraph form. Never let your summary's contain outlines or built points"

prompt = f"""
Here are a bunch of bulit points. Please summerize them in paragraph form:

{pre_summery}
"""

final_summery = generate_text(model, prompt, system_prompt)


print("Done")

pdf.add_section(Section(f"## Basic Overview\n{final_summery}\n\n"))
pdf.add_section(Section(f"## Outline\n{pre_summery}\n\n"))

pdf.meta["title"] = title
pdf.meta["author"] = "locker98"
pdf.save(f"{title}.pdf")

# print(f"\n\n\n{pre_summery}")