How To Build a News Agent with GPT-OSS, Hugging Face Inference & Gradio

def fetch_google_news_rss(num: int = 10) -> List[Dict[str, Any]]:
    """Fetch general news from Google News RSS feed."""
    try:
        url = "https://news.google.com/rss"
        r = requests.get(url, timeout=30)
        r.raise_for_status()
        
        # Parse RSS XML
        root = ET.fromstring(r.content)
        items = root.findall('.//item')
        
        results = []
        for item in items[:num]:
            title = item.find('title')
            link = item.find('link')
            pub_date = item.find('pubDate')
            source = item.find('source')
            
            results.append({
                "title": title.text if title is not None else "No title",
                "link": link.text if link is not None else "",
                "pub_date": pub_date.text if pub_date is not None else "No date",
                "source": source.text if source is not None else "Google News"
            })
        
        return results
    except Exception as e:
        return {"ok": False, "error": repr(e)}

def serper_news_search(query: str, num: int = 5) -> List[Dict[str, Any]]:
    """Fetch news for a specific topic or query."""
    url = "https://google.serper.dev/news"
    headers = {"X-API-KEY": SERPER_API_KEY, "Content-Type": "application/json"}
    payload = {"q": query, "gl": "us", "hl": "en", "tbs": "qdr:d"}
    r = requests.post(url, headers=headers, json=payload, timeout=30)
    r.raise_for_status()
    data = r.json()
    results = []
    for item in data.get("news", [])[:num]:
        results.append({
            "title": item.get("title"),
            "link": item.get("link"),
            "snippet": item.get("snippet"),
            "date": item.get("date"),  # ISO8601 when available
            "source": item.get("source")
        })
    return results

def serper_site_search(query: str, site: str, num: int = 5) -> List[Dict[str, Any]]:
    """Site restricted web search."""
    url = "https://google.serper.dev/search"
    headers = {"X-API-KEY": SERPER_API_KEY, "Content-Type": "application/json"}
    payload = {"q": f"site:{site} {query}", "gl": "us", "hl": "en"}
    r = requests.post(url, headers=headers, json=payload, timeout=30)
    r.raise_for_status()
    data = r.json()
    results = []
    for item in data.get("organic", [])[:num]:
        results.append({
            "title": item.get("title"),
            "link": item.get("link"),
            "snippet": item.get("snippet"),
            "favicons": item.get("favicons", {})
        })
    return results

def fetch_article(url: str, max_chars: int = 12000) -> Dict[str, Any]:
    """Fetch and extract clean article text with trafilatura."""
    try:
        downloaded = trafilatura.fetch_url(url, timeout=30)
        text = trafilatura.extract(downloaded, include_comments=False) if downloaded else None
        if not text:
            return {"ok": False, "error": "could_not_extract"}
        text = text.strip()
        if len(text) > max_chars:
            text = text[:max_chars] + " ..."
        return {"ok": True, "text": text}
    except Exception as e:
        return {"ok": False, "error": repr(e)}

# OpenAI-style tool specs for function calling
TOOLS = [
    {
        "type": "function",
        "function": {
            "name": "fetch_google_news_rss",
            "description": "Fetch general top headlines from Google News RSS feed. Use this when you want to see what's happening in the world today without a specific topic focus.",
            "parameters": {
                "type": "object",
                "properties": {
                    "num": {"type": "integer", "minimum": 1, "maximum": 20, "description": "Number of news items to fetch"}
                },
                "required": []
            }
        }
    },
    {
        "type": "function",
        "function": {
            "name": "serper_news_search",
            "description": "Search Google News for articles about a specific topic or query. Use this when you need news about particular subjects, companies, or events.",
            "parameters": {
                "type": "object",
                "properties": {
                    "query": {"type": "string"},
                    "num": {"type": "integer", "minimum": 1, "maximum": 20}
                },
                "required": ["query"]
            }
        }
    },
    {
        "type": "function",
        "function": {
            "name": "serper_site_search",
            "description": "Search a specific news domain for relevant articles.",
            "parameters": {
                "type": "object",
                "properties": {
                    "query": {"type": "string"},
                    "site": {"type": "string", "description": "Domain like ft.com or nytimes.com"},
                    "num": {"type": "integer", "minimum": 1, "maximum": 10}
                },
                "required": ["query", "site"]
            }
        }
    },
    {
        "type": "function",
        "function": {
            "name": "fetch_article",
            "description": "Download and extract the main text of an article from a URL. ONLY use this when the user asks specific questions about article content, details, or wants to analyze/quote from particular articles. Do NOT use this for general news summaries or overviews.",
            "parameters": {
                "type": "object",
                "properties": {
                    "url": {"type": "string"},
                    "max_chars": {"type": "integer", "minimum": 1000, "maximum": 60000}
                },
                "required": ["url"]
            }
        }
    }
]

FUNCTION_MAP = {
    "fetch_google_news_rss": fetch_google_news_rss,
    "serper_news_search": serper_news_search,
    "serper_site_search": serper_site_search,
    "fetch_article": fetch_article,
}

def call_model(messages: List[Dict[str, str]], tools=TOOLS, temperature: float = 0.3, model: str = DEFAULT_MODEL):
    """One step with tool calling support."""
    try:
        return client.chat.completions.create(
            model=model,
            temperature=temperature,
            messages=messages,
            tools=tools,
            tool_choice="auto"
        )
    except Exception as e:
        print(f"Error calling model: {e}")
        raise

def run_agent(user_prompt: str, site_limit: Optional[str] = None, model: str = DEFAULT_MODEL) -> str:
    """
    High level prompt for a news agent.
    It may search, read links, then synthesize and cite URLs.
    """
    system = {
        "role": "system",
        "content": (
            "You are a careful news agent. Follow these steps:\n"
            "1. For general news requests: Use fetch_google_news_rss to get top headlines\n"
            "2. For specific topic requests: Use serper_news_search with the topic\n"
            "3. ONLY use fetch_article when the user asks specific questions about article content, details, or wants to analyze/quote from particular articles\n"
            "4. For general news summaries, provide information based on headlines and snippets without fetching full articles\n"
            "5. STOP calling tools and provide your final answer\n"
            "6. Always include a bullet list of sources with URLs\n"
            "IMPORTANT: After reading articles (if any), you must provide your final answer without calling more tools.\n\n"
            "TOOL SELECTION GUIDE:\n"
            "- fetch_google_news_rss: Use for 'what's happening today' or 'top news' requests\n"
            "- serper_news_search: Use for specific topics like 'AI chips', 'Nvidia', 'climate change'\n"
            "- serper_site_search: Use when restricted to specific news sources\n"
            "- fetch_article: ONLY use when user asks about specific article content, details, or wants to analyze particular articles\n"
            "PRIORITY: For general news requests, provide summaries based on headlines and snippets. Only fetch full articles when specifically needed for detailed analysis.\n"
        ),
    }

    messages: List[Dict[str, str]] = [system, {"role": "user", "content": user_prompt}]
    if site_limit:
        messages.append({"role": "user", "content": f"Restrict searches to {site_limit} when appropriate."})

    for step in range(6):  # small safety cap
        try:
            resp = call_model(messages, model=model)
            msg = resp.choices[0].message

            # If the model wants to call tools
            if getattr(msg, "tool_calls", None) and msg.tool_calls:
                # Add the assistant message with tool calls to the conversation
                assistant_message = {
                    "role": "assistant",
                    "content": msg.content or "",
                    "tool_calls": [
                        {
                            "id": tool_call.id,
                            "type": "function",
                            "function": {
                                "name": tool_call.function.name,
                                "arguments": tool_call.function.arguments
                            }
                        }
                        for tool_call in msg.tool_calls
                    ]
                }
                messages.append(assistant_message)
                
                # Process each tool call
                for tool_call in msg.tool_calls:
                    name = tool_call.function.name
                    args = {}
                    try:
                        args = json.loads(tool_call.function.arguments or "{}")
                    except json.JSONDecodeError:
                        args = {}

                    fn = FUNCTION_MAP.get(name)
                    if not fn:
                        messages.append({
                            "role": "tool",
                            "tool_call_id": tool_call.id,
                            "name": name,
                            "content": json.dumps({"ok": False, "error": "unknown_tool"})
                        })
                        continue

                    try:
                        result = fn(**args)
                    except TypeError as e:
                        result = {"ok": False, "error": f"bad_args: {e}"}
                    except Exception as e:
                        result = {"ok": False, "error": repr(e)}

                    tool_response = {
                        "role": "tool",
                        "tool_call_id": tool_call.id,
                        "name": name,
                        "content": json.dumps(result),
                    }
                    messages.append(tool_response)
                
                # After processing tools, add a reminder to synthesize
                if step >= 2:  # After 2+ tool calls, encourage synthesis
                    messages.append({
                        "role": "user",
                        "content": "You now have sufficient information. Please provide your final answer with sources."
                    })
                
                # Continue loop so the model can see tool outputs
                continue

            # If we have a final assistant message without tool calls
            if msg.content:
                return msg.content

            # Fallback tiny sleep then continue
            time.sleep(0.2)
            
        except Exception as e:
            # If there's an error, try to continue or return error message
            if step == 5:  # Last step
                return f"Error occurred during processing: {e}"
            time.sleep(0.5)
            continue

    return "I could not complete the task within the step limit. Try refining your query."

@observe()
def chat_with_agent(message, history, model):
    """Handle chat messages and return agent responses."""
    if not message.strip():
        return history
    
    lf = get_client()
    lf.update_current_trace(
        input={"user_message": message, "model": model, "history_length": len(history)}
    )

    try:
        response = run_agent(message, None, model)

        lf.update_current_trace(
            output={"agent_response": response},
            metadata={
                "model": model,
                "message_length": len(message),
                "response_length": len(response),
                "success": True,
            },
        )

        history.append({"role": "user", "content": message})
        history.append({"role": "assistant", "content": response})
        return history
        
    except Exception as e:
        lf.update_current_trace(
            output={"error": str(e)},
            metadata={"success": False, "error": str(e)},
        )
        error_msg = f"Sorry, I encountered an error: {str(e)}"
        history.append({"role": "user", "content": message})
        history.append({"role": "assistant", "content": error_msg})
        return history

def clear_chat():
    """Clear the chat history."""
    return [], ""

# Create the Gradio interface
with gr.Blocks(
    title="Chat with the News",
    theme=gr.themes.Monochrome()
) as demo:
    
    # Header using Gradio markdown
    gr.Markdown("""
    # 📰 Chat with the News
    
    Your AI-powered news research assistant with real-time search capabilities, based on [GPT-OSS models](https://huggingface.co/collections/openai/gpt-oss-68911959590a1634ba11c7a4) and running on inference providers.
    """)
    
    # Examples section using Gradio markdown
    gr.Markdown("""
    ### 💡 Try these examples:
    
    - **General:** "What are the top news stories today?"
    - **Specific topic:** "What's the latest on artificial intelligence?"
    - **Site-specific:** "What's the latest climate change news on the BBC?"
    """)

    # Model selector
    model_selector = gr.Dropdown(
        choices=AVAILABLE_MODELS,
        value=DEFAULT_MODEL,
        label="🤖 Select Model",
        info="Choose between GPT-OSS 120B and 20B models"
    )

    # Message input
    msg = gr.Textbox(
        label="Ask me about the news",
        placeholder="What would you like to know about today?",
        lines=2
    )
    
    # Buttons in a row
    with gr.Row():
        submit_btn = gr.Button("🚀 Send", variant="primary", size="lg")
        clear_btn = gr.Button("🗑️ Clear Chat", variant="secondary", size="lg")
        
    # Chat interface
    chatbot = gr.Chatbot(
        label="News Agent",
        height=500,
        show_label=False,
        container=True,
        type="messages"
    )
    
    # Event handlers
    submit_btn.click(
        chat_with_agent,
        inputs=[msg, chatbot, model_selector],
        outputs=[chatbot],
        show_progress=True
    )
    
    msg.submit(
        chat_with_agent,
        inputs=[msg, chatbot, model_selector],
        outputs=[chatbot],
        show_progress=True
    )
    
    clear_btn.click(
        clear_chat,
        outputs=[chatbot, msg]
    )
    
    # Instructions using Gradio markdown
    gr.Markdown("""
    ---
    
    ### ℹ️ How it works
    
    This AI agent can search Google News, fetch articles from specific sources, and provide comprehensive news summaries with proper citations. It uses real-time data and can restrict searches to specific news domains when requested.
    
    **Model Selection:**
    - **GPT-OSS 120B**: Larger, more capable model for complex reasoning tasks
    - **GPT-OSS 20B**: Faster, more efficient model for quick responses
    """)

# Launch the app
if __name__ == "__main__":
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=False,
        show_error=True
    )