app.py

import gradio as gr
from apscheduler.schedulers.background import BackgroundScheduler
from huggingface_hub import HfApi
from datasets import Dataset, load_dataset, concatenate_datasets
from datetime import datetime
from huggingface_hub import login
import os

login(os.getenv("HF_TOKEN"))

def upload_newsletter(date, newsletter_content):
    try:
        previous_dataset = load_dataset("m-ric/newsletter", split="train")
        new_dataset = Dataset.from_dict({"date": [date], "content": [newsletter_content]})
        final_dataset = concatenate_datasets([previous_dataset, new_dataset])
    except:
        final_dataset = Dataset.from_dict({"date": [date], "content": [newsletter_content]})
    final_dataset.push_to_hub("m-ric/newsletter", split="train")

def restart_space():
    HfApi(token=os.getenv("HF_TOKEN", None)).restart_space(
        repo_id="m-ric/multiagent_newsletter"
    )
    print(f"Space restarted on {datetime.now()}")

# restart the space every day at 9am
scheduler = BackgroundScheduler()
scheduler.add_job(restart_space, "cron", day_of_week="mon-sun", hour=7, minute=0)
scheduler.start()

### CREATE TOOL ##########
from transformers.agents import DuckDuckGoSearchTool

search_tool = DuckDuckGoSearchTool()

print("Search tool created.")

### CREATE LLM ENGINE ############
# Below is an example of how to build an OpenAI engine:
from openai import OpenAI
from transformers.agents.llm_engine import MessageRole, get_clean_message_list

openai_role_conversions = {
    MessageRole.TOOL_RESPONSE: MessageRole.USER,
}


class OpenAIEngine:
    def __init__(self, model_name="gpt-4o"):
        self.model_name = model_name
        self.client = OpenAI(
            api_key=os.getenv('OPENAI_API_KEY')
        )

    def __call__(self, messages, stop_sequences=[]):
        messages = get_clean_message_list(messages, role_conversions=openai_role_conversions)

        response = self.client.chat.completions.create(
            model=self.model_name,
            messages=messages,
            stop=stop_sequences,
            temperature=0.5,
        )
        return response.choices[0].message.content

# But instead we use HF one, since it's free:
from transformers import ReactCodeAgent, HfApiEngine

agent = ReactCodeAgent(
    llm_engine=HfApiEngine("meta-llama/Meta-Llama-3.1-70B-Instruct"),
    tools=[search_tool],
    max_iterations=10
)

print("Agent initiated")

topics = "Large Language Models, Tech, GPUs, AI"

report = agent.run(
    f"""Give me a complete newsletter in markdown format of the latest developments in these topics: {topics}, with a consistent style and layout with URL.
Each topic should have its markdown-formatted analysis, including a rundown, detailed bullet points,
and a "Why it matters" section. For each topic there should be at least 3 articles with URL
Example Output:
# Top stories in AI today:\\n\\n
- AI takes spotlight in Super Bowl commercials\\n
- Altman seeks TRILLIONS for global AI chip initiative\\n\\n
## AI takes spotlight in Super Bowl commercials\\n\\n
**url:** https://example.com/story1 \\n\\
**The Rundown:** AI made a splash in this year\'s Super Bowl commercials...\\n\\n
**The details:**...\\n\\n
**Why it matters::**...\\n\\n
## Altman seeks TRILLIONS for global AI chip initiative\\n\\n
**The Rundown:** OpenAI CEO Sam Altman is reportedly angling to raise TRILLIONS of dollars...\\n\\n'
**The details:**...\\n\\n
**Why it matters::**...\\n\\n
--- end of example.
Don't worry about formatting too much, just dump results into a huge final string and return it with final_answern the formatting will be handled later.
Also, gather all news at once using a for loop on all topics with google_search, it will make things much quicker.
Make sure that you didn't forget any topic! If no news are found for any specific topic, specify 'No news on this topic were found in the last 24 hours'."""
)



date = datetime.today().strftime('%Y-%m-%d')

upload_newsletter(date, report)

print("Report pushed to dataset!")

with gr.Blocks() as demo:
    gr.Markdown(f"{date}, topics: {topics}\n### ✅ Newsletter generated! Content below 👇")
    gr.Markdown(clean_report)

    gr.Markdown("---\n---\n---\n### FYI, before reformulation, the report was: 👇")
    gr.Markdown(intermediate_report)

demo.launch()