llmsummary / main.py
Manaranjan's picture
deploy at 2024-08-13 09:10:42.262523
66b7053 verified
from fasthtml_hf import setup_hf_backup
import io
import os
import traceback
from pydantic_core import from_json
from fasthtml.common import *
from PyPDF2 import PdfReader
from PyPDF2 import PdfReader
from langchain.chains.summarize import load_summarize_chain
from langchain_core.prompts import PromptTemplate
from langchain_openai import ChatOpenAI
from langchain_anthropic import ChatAnthropic
from pydantic import BaseModel, Field, ValidationError
from langchain.output_parsers import PydanticOutputParser
# Initialize the fastHtml application
app, rt = fast_app()
# Define Pydantic models for structured output
# SummaryLine represents a single summary item with its keywords and description
class SummaryLine(BaseModel):
summary_item: str = Field(description = "Actual summary sentence that contains highlighting key data points or information.",
max_length = 200)
keywords: List[str] = Field(description = "A list of exact words or phrases in the summary item that highlights most important data points or key ideas.")
brief_descripton_of_summary: str = Field(description = "This is elaborate description to provide context or background to the summary item.",
min_length = 200,
max_length = 500)
# TopicSummaries represents a collection of summaries for a specific topic
class TopicSummaries(BaseModel):
topic: str = Field(description = "Topics of summary as mentioned in the instructions.")
summaries: List[SummaryLine] = Field(description = "This a list summary for a topic with each one having it's own keywords and context.",
min_items=3,
max_items=5)
# CompleteSummary is the top-level model containing all topic summaries
class CompleteSummary(BaseModel):
summaries_list: List[TopicSummaries]
# Define the template for summarization
# This template provides instructions to the AI model on how to structure the summary
summarize_template = """
Write a concise summary of the case study given in the context. The summary should be based on the following topics.
"""
# Define the specific sections to be included in the summary
summary_sections = """
- Factual: Facts or information that contains numbers, dates, events etc. that are mostly quantitative or qualitative data
- SWOT: Key Strength, weakness, opportunities or threats that are mentioned in the case study
- Decisions and Outcomes: Key decisions taken and it's successful or failed outcomes and reasons
- Ethical and Governance: Key considerations from ethical and governance perspective
"""
# Define the context string for one-pass summarization
# This string provides additional formatting instructions for the summary
context_str = """
<context>
{context_content}
</context>
The response must follow the following schema strictly. There will be penalty for not following the schema.
"""
# Define the template for the reduce step in map-reduce summarization
# This template instructs the model to consolidate multiple summaries into a final summary
refine_str = """The following are set of summaries given in a markdown format:
{previous_summary}
Now add the above summary with more context given below and create final summary, which should contain the following sections.
"""
# Function to get the appropriate language model based on user selection
def getModel(model, key):
if(model == 'OpenAI'):
os.environ['OPENAI_API_KEY'] = key
return ChatOpenAI(temperature=0, # Set to 0 for deterministic output
model="gpt-4o", # Using the GPT-4 Turbo model
max_tokens=4096) # Limit the response length
else:
os.environ['ANTHROPIC_API_KEY'] = key
return ChatAnthropic(model='claude-3-5-sonnet-20240620') # Limit the response length
# Function to highlight specific keywords in the text
def highlight_text(text, key_words):
for word in key_words:
text = text.replace(word, f'<span style="color:red;"><b>{word}</b></span>')
html_text = "<div>" + text + "</div>"
return eval(html2ft(html_text))
# Function to generate an HTML table from the summary object
def generate_table(summaries_obj):
column_names = ['Topic', "Summary"]
table_header = Thead(Tr(*[Th(key) for key in column_names]))
table_rows = []
for topic_summary in summaries_obj.summaries_list:
first_row = True
for summary in topic_summary.summaries:
if(first_row):
table_rows.append(Tr(Td(topic_summary.topic,
rowspan=f"{len(topic_summary.summaries)}",
style = "width: 10%;"),
Td(highlight_text(summary.summary_item, summary.keywords),
style = "width: 60%;"),
Td(Div(Details(Summary( style = "summary::-webkit-details-marker { display: none }; list-style-type: '+'"),
P(summary.brief_descripton_of_summary)),
style ="padding: 0.5em 0.5em 0;"),
style = "width: 30%;")))
first_row = False
else:
table_rows.append(Tr(Td(highlight_text(summary.summary_item, summary.keywords),
style = f"width: 60%; rowspan='{len(topic_summary.summaries)}'"),
Td(Div(Details(Summary( style = "summary::-webkit-details-marker { display: none }; list-style-type: '+'"),
P(summary.brief_descripton_of_summary)),
style ="padding: 0.5em 0.5em 0;"),
style = "width: 30%;")))
return Div(Card(Table(table_header, Tbody(*table_rows))))
# Function to perform one-pass summarization on the given pages
def onepass_summarize(pages, summary_sections, model):
"""
Perform one-pass summarization on the given pages.
This function creates a summarization chain using the provided instructions
and model, then applies it to the input pages to generate a summary.
Args:
pages (list): List of pages (documents) to summarize
instructions (str): Custom instructions for summarization
model (ChatOpenAI): Instance of ChatOpenAI model to use for summarization
Returns:
str: Summarized text in markdown format
"""
onepass_summary_template = summarize_template + summary_sections + context_str + "{format_instructions}"
print("Onepass instruction: " + onepass_summary_template)
output_parser = PydanticOutputParser(pydantic_object=CompleteSummary)
format_instructions = output_parser.get_format_instructions()
print("Format instructions: " + format_instructions)
# Create a prompt template combining the instructions and context
prompt = PromptTemplate.from_template(onepass_summary_template)
# Create an LLM chain with the model and prompt
summary_chain = prompt | model | output_parser
print("Getting Summary......")
# Invoke the chain on the input pages and return the summarized text
summaries = summary_chain.invoke({"context_content": pages,
"format_instructions": format_instructions})
return summaries
# Function to generate the configuration form for the web interface
def getConfigForm():
return Card(Form(hx_post="/submit", hx_target="#result", hx_swap_oob="innerHTML", hx_indicator="#indicator")(
Div(
Label(Strong("Model and Prompt Instruction: "), style="color:#3498db; font-size:25px;")
),
Div(
Label(Strong('Model: ')),
Select(Option("OpenAI"), Option("Anthropic"), id="model")
),
Div(
Label(Strong('Secret Key: ')),
Input(id="secret", type="password", placeholder="Key: "),
),
Div(
Label(Strong('Upload File: '), "Upload only pdf file with max size of 1 MB"),
Input(id="file", type = 'file', placeholder="Key: ", accept = ".pdf", max = '1024000'),
),
Div(
Label(Strong('Instruction: ')),
P('Provide the list of topics and their one line description for summarization as shown in example. Summarization will have these sections.',
style = 'font-size: 12px;'),
Textarea(summary_sections, id="instruction",
style="height:250px")
),
Div(
Button("Summarize")
),
Div(
Br(),
A("Developed by Manaranjan Pradhan", href="http://www.manaranjanp.com/",
target="_blank",
style = 'color: red; font-size: 16px;')
)))
# Define the route for the homepage
@app.get('/')
def homepage():
return Titled('Document Summarization', Grid( getConfigForm(),
Div(
Div(Label(Strong('Summarizing the document.... take a deep breath....')),
Progress(), id="indicator", cls="htmx-indicator"),
Div(id="result", style ="font-family:Helvetica; font-size=24pt;")
)
, style="grid-template-columns: 400px 1000px; gap: 50px;"
))
# Define the route for form submission
@app.post('/submit')
async def post(d:dict):
try:
# Check if a file was uploaded
if "file" in d.keys():
pages = await d['file'].read(-1)
pdf_reader = PdfReader(io.BytesIO(pages))
else:
return Div("File not uploaded.", cls = 'alert', )
# Extract text from each page of the PDF
text_content = ""
for page in pdf_reader.pages:
text_content += page.extract_text() + "\n"
# Get the appropriate language model
model = getModel(d['model'], d['secret'])
# Perform one-pass summarization
summaries = onepass_summarize(text_content, d['instruction'], model)
print(f"Summary Obtained: {summaries}")
# Generate and return the HTML table with the summaries
return generate_table(summaries)
except BaseException as e:
print(traceback.format_exc())
return str(e)
setup_hf_backup(app)
# Start the FastAPI server
serve()