run_claude.py

import anthropic
import base64
import pandas as pd
import requests
import os
import numpy as np
from openai import OpenAI
import io
import tiktoken
import PyPDF2

import prompts

from typing import List, Literal
from pydantic import BaseModel

import time

import gradio as gr

ANTHROPIC_API_KEY = os.environ.get("ANTHROPIC_API_KEY")
OPENAI_API_KEY = os.environ.get("OPENAI_API_KEY")


def ask_claude(
    query: str, 
    pdf_path: str = None, 
    use_cache: bool = False, 
    system: str = None,
    max_tokens: int = 1024,
    model: str = "claude-3-5-sonnet-20241022"
) -> str:
    """
    Unified function to query Claude API with various options.
    
    Args:
        query: Question/prompt for Claude
        pdf_path: Optional path to PDF file (local or URL)
        use_cache: Whether to enable prompt caching
        system: Optional system prompt
        max_tokens: Maximum tokens in response (default 1024)
        model: Claude model to use (default claude-3-5-sonnet)
        
    Returns:
        Claude's response as a string
    """
    client = anthropic.Anthropic(api_key=ANTHROPIC_API_KEY)
    
    # Handle PDF if provided
    content = query
    betas = []
    if pdf_path:
        # Get PDF content
        if pdf_path.startswith(('http://', 'https://')):
            response = requests.get(pdf_path)
            binary_data = response.content
        else:
            with open(pdf_path, "rb") as pdf_file:
                binary_data = pdf_file.read()
        
        pdf_data = base64.standard_b64encode(binary_data).decode("utf-8")
        content = [
            {
                "type": "document",
                "source": {
                    "type": "base64",
                    "media_type": "application/pdf",
                    "data": pdf_data
                }
            },
            {
                "type": "text",
                "text": query
            }
        ]
        betas.append("pdfs-2024-09-25")
    
    # Add prompt caching if requested
    if use_cache:
        betas.append("prompt-caching-2024-07-31")
    
    # Prepare API call kwargs
    kwargs = {
        "model": model,
        "max_tokens": max_tokens,
        "messages": [{"role": "user", "content": content}]
    }
    
    # Add optional parameters if provided
    if system:
        kwargs["system"] = system
    if betas:
        kwargs["betas"] = betas
        
    message = client.beta.messages.create(**kwargs)
    return message.content[0].text

class Point(BaseModel):
    content: str
    importance: Literal["critical", "minor"]


class Review(BaseModel):
    contributions: str
    strengths: List[Point]
    weaknesses: List[Point]
    requested_changes: List[Point]
    impact_concerns: str


importance_mapping = {"critical": 2, "minor": 1}

client = OpenAI(api_key=OPENAI_API_KEY)

model_name = "gpt-4o-2024-08-06"

def format_gpt(prompt):
    chat_completion = client.beta.chat.completions.parse(
        messages=[
            {
                "role": "user",
                "content": prompt,
            }
        ],
        model='gpt-4o',
        response_format=Review,
    )
    return chat_completion.choices[0].message.parsed.model_dump()

def parse_final(parsed, max_strengths=3, max_weaknesses=5, max_requested_changes=5):
    new_parsed = {}
    new_parsed["contributions"] = parsed["contributions"]
    new_parsed["impact_concerns"] = parsed["impact_concerns"]
    new_parsed["strengths"] = "\n".join(
        [f'- {point["content"]}' for point in parsed["strengths"][:max_strengths]]
    )
    new_parsed["weaknesses"] = "\n".join(
        [f'- {point["content"]}' for point in parsed["weaknesses"][:max_weaknesses]]
    )
    request_changes_sorted = sorted(
        parsed["requested_changes"],
        key=lambda x: importance_mapping[x["importance"]],
        reverse=True,
    )
    new_parsed["requested_changes"] = "\n".join(
        [
            f"- {point['content']}"
            for point in request_changes_sorted[:max_requested_changes]
        ]
    )
    return new_parsed

def process(file_content, progress=gr.Progress()):
    # Create a list to store log messages
    log_messages = []
    
    def log(msg):
        print(msg)
        log_messages.append(msg)
        return "\n".join(log_messages)

    if not os.path.exists("cache"):
        os.makedirs("cache")
    pdf_path = f"cache/{time.time()}.pdf"
    with open(pdf_path, "wb") as f:
        f.write(file_content)

    progress(0, desc="Starting review process...")
    log("Starting review process...")

    all_reviews = []
    for i in range(3):
        progress((i + 1) / 3, desc=f"Generating review {i+1}/3")
        log(f"Generating review {i+1}/3...")
        all_reviews.append(ask_claude(prompts.review_prompt, pdf_path=pdf_path))
    all_reviews_string = "\n\n".join([f"Review {i+1}:\n{review}" for i, review in enumerate(all_reviews)])
    
    progress(0.4, desc="Combining reviews...")
    log("Combining reviews...")
    combined_review = ask_claude(prompts.combine_prompt.format(all_reviews_string=all_reviews_string,
                                                             review_format=prompts.review_format), pdf_path=pdf_path)

    progress(0.6, desc="Defending paper...")
    log("Defending paper...")
    rebuttal = ask_claude(prompts.defend_prompt.format(combined_review=combined_review), pdf_path=pdf_path)

    progress(0.8, desc="Revising review...")
    log("Revising review...")
    revised_review = ask_claude(prompts.revise_prompt.format(review_format=prompts.review_format, combined_review=combined_review, defended_paper=rebuttal), pdf_path=pdf_path)
    log("Humanizing review...")
    humanized_review = ask_claude(prompts.human_style.format(review=revised_review), pdf_path=pdf_path)

    progress(0.9, desc="Formatting review...")
    log("Formatting review...")
    formatted_review = parse_final(format_gpt(prompts.formatting_prompt.format(review=humanized_review)))

    log("Finished!")

    contributions, strengths, weaknesses, requested_changes, impact_concerns = (
                formatted_review["contributions"],
                formatted_review["strengths"],
                formatted_review["weaknesses"],
                formatted_review["requested_changes"],
                formatted_review["impact_concerns"],
            )
    contributions = f"# Contributions\n\n{contributions}"
    strengths = f"# Strengths\n\n{strengths}"
    weaknesses = f"# Weaknesses\n\n{weaknesses}"
    requested_changes = f"# Requested Changes\n\n{requested_changes}"
    impact_concerns = f"# Impact Concerns\n\n{impact_concerns}"
    return (
        contributions,
        strengths,
        weaknesses,
        requested_changes,
        impact_concerns,
        "\n".join(log_messages),  # Return the log messages
    )

def gradio_interface():
    with gr.Blocks() as demo:
        gr.Markdown("# TMLR Reviewer")
        gr.Markdown("This tool helps you generate high-quality reviews for the Transactions on Machine Learning Research (TMLR).")
        
        with gr.Row():
            # Left column
            left_column = gr.Column(scale=1)
            with left_column:
                upload_component = gr.File(label="Upload PDF", type="binary")
                submit_btn = gr.Button("Generate Review")
                # Progress log moved below upload section
                progress_log = gr.Textbox(label="Progress Log", interactive=False, lines=10)
            
            # Right column for review outputs
            right_column = gr.Column(scale=2)
            with right_column:
                output_component_contributions = gr.Markdown(label="Contributions")
                output_component_strengths = gr.Markdown(label="Strengths")
                output_component_weaknesses = gr.Markdown(label="Weaknesses")
                output_component_requested_changes = gr.Markdown(label="Requested Changes")
                output_component_impact_concerns = gr.Markdown(label="Impact Concerns")

        submit_btn.click(
            fn=process,
            inputs=upload_component,
            outputs=[
                output_component_contributions,
                output_component_strengths,
                output_component_weaknesses,
                output_component_requested_changes,
                output_component_impact_concerns,
                progress_log,
            ]
        )
        
    demo.queue()
    return demo

if __name__ == "__main__":
    demo = gradio_interface()
    demo.launch(share=False)