Detailed summaries and high-quality content creation with genAI

import json
import os
import sys
from langchain_community.chat_models import BedrockChat
from langchain_core.messages import HumanMessage
import boto3

bedrock_runtime = boto3.client('bedrock-runtime')

llm_chat_2 = BedrockChat(
model_id="anthropic.claude-3-sonnet-20240229-v1:0",
model_kwargs={"temperature": 0.1},
client=bedrock_runtime,
region_name="us-west-2"
)

from langchain_community.document_loaders import PyPDFLoader

loader = PyPDFLoader("docs/treasure_island.pdf", extract_images=False)

from langchain_core.prompts import PromptTemplate
from langchain.chains.llm import LLMChain
from langchain.chains.combine_documents.stuff import StuffDocumentsChain

prompt_template = """Identify the main sections and subsections of the following document. Include the page number where each section and subsection starts and ends. Provide the output in the same language as the original document:
<example>
"Section 1: Title of Section 1 (pages 1-5)",
"Section 2: Title of Section 2 (pages 6-10)",
.... rest of sections
</example>
"{text}"
List of sections and subsections:"""

prompt = PromptTemplate.from_template(prompt_template)

# Define LLM chain

llm_chain = LLMChain(llm=llm_chat_2, prompt=prompt)

# Define StuffDocumentsChain
stuff_chain = StuffDocumentsChain(llm_chain=llm_chain, document_variable_name="text")

docs = loader.load()
documento = stuff_chain.run(docs)

from langchain.prompts import PromptTemplate
from langchain.chains import LLMChain
from langchain.output_parsers import PydanticOutputParser
from pydantic import BaseModel, Field
from typing import List
import re

# Defining the desired output format
class DocumentoSecciones(BaseModel):
sections: List[str] = Field(description="List of document sections")

# Creating the parser
parser = PydanticOutputParser(pydantic_object=DocumentoSecciones)

prompt_template = """You will be provided with a list of sections from a document. Create a detailed list that includes all sections and subsections, with their respective page ranges.

Important Instructions:

Prioritize subsections over general sections.
If a general section is completely divided into subsections, omit the general section.
If there are parts of a general section not covered by subsections, include only those parts.
For sections without specific numbering, use 'N/A' as the page range.
If you are unsure of the page where a section ends, indicate that it ends on the starting page of the next section.
{format_instructions}

<sections> {text} </sections>
Make sure the output complies with the specified format and follows the given instructions. Provide the output in the same language as the original document.
"""

prompt = PromptTemplate(
template=prompt_template,
input_variables=["text"],
partial_variables={"format_instructions": parser.get_format_instructions()}
)

# Define LLM chain
llm_chain = LLMChain(llm=llm_chat_2, prompt=prompt)

# Run the chain
resultado = llm_chain.run(documento)

def extract_json(text):
json_match = re.search(r'\{[\s\S]*\}', text)
if json_match:
return json_match.group()
return None

try:
json_str = extract_json(resultado)
if json_str:
parsed_json = json.loads(json_str)
lista_secciones = parsed_json.get('secciones', [])
else:
raise ValueError("Not valid JSON in the response")
except json.JSONDecodeError as e:
print(f"Decoding error JSON: {e}")
print("Original output:", resultado)
lista_secciones = []
except Exception as e:
print(f"Processing error: {e}")
print("Original output:", resultado)
lista_secciones = []

from concurrent.futures import ThreadPoolExecutor, as_completed

def summarize_section(seccion, docs, llm_chat):
prompt_template = f"""You are an assistant specialized in summarizing documents. Your goal is to help users quickly understand the most important points of a document, including key numerical information, main outcomes, general content etc.

You will be provided with a complete document and you will have to perform a detailed summary of only the section indicated below. For this summary, you can use the context information from the entire document, but the final result must be only the summary of the specific section indicated.

Section to summarize: {seccion}
Document: "{{text}}"

Important Instructions:

Provide the content of the summary directly, without additional introduction or conclusion.
Do not repeat the section title in the summary.
Structure the summary in short and concise paragraphs.
Use bullets to list key points or important numerical data.
Make sure to include all relevant and numerical information from the section.
Maintain a professional and objective tone.
Do not use phrases like "Summary of the section" or similar.
Provide the output in the same language as the original document.

Summary:"""

prompt = PromptTemplate.from_template(prompt_template)

llm_chain = LLMChain(llm=llm_chat_2, prompt=prompt)
stuff_chain = StuffDocumentsChain(llm_chain=llm_chain, document_variable_name="text")

summary_section = stuff_chain.run(docs)
return seccion, summary_section

# Assuming 'loader' and 'llm_chat' are defined earlier in your code
docs = loader.load()

# Use ThreadPoolExecutor to run tasks in parallel
with ThreadPoolExecutor() as executor:
futures = [executor.submit(summarize_section, seccion, docs, llm_chat) for seccion in lista_secciones]

# Collect results in order
results = []
for future in as_completed(futures):
results.append(future.result())

# Sort results to maintain original order
results.sort(key=lambda x: lista_secciones.index(x[0]))

# Combine results with improved formatting
output_final = ""
for seccion, summary in results:
output_final += f"## **{seccion}**\n\n{summary.strip()}\n\n---\n\n"

print(output_final)