"
"⚡ In production, this pipeline runs on AMD Instinct™ MI300X with "
"vLLM (PagedAttention) serving Qwen3.5-9B + Qwen3.6-27B models. "
"This demo simulates the agent flow for showcase purposes."
"
"
)
history.append({"role": "assistant", "content": final_msg})
yield history
def handle_user_message(
message: str,
history: list,
) -> Generator:
"""Handle custom user messages with a simulated response."""
if not message.strip():
yield history
return
history = history or []
history.append({"role": "user", "content": message})
yield history
time.sleep(0.5)
# Simulated response for any custom input
response = (
"### 🔀 Router Agent\n\n"
"**Note:** This is a demo environment running on HF Spaces "
"without GPU acceleration.\n\n"
"In the **production deployment** on AMD Instinct™ MI300X, "
"your clinical case would be processed through our full "
"5-agent pipeline:\n\n"
"1. **Router** — Classifies oncological vs. non-oncological\n"
"2. **Clinical Extraction** — Extracts structured entities\n"
"3. **Corrective RAG** — Retrieves from NCCN/ESMO guidelines\n"
"4. **Specialist** — Generates evidence-based recommendation\n"
"5. **Critic (Reflexion)** — Validates safety and completeness\n\n"
"👉 Click **▶ View Demo** to see a complete simulated triage "
"with the endometrial cancer case.\n\n"
"🔗 **Production:** Deploy with `docker compose up` on MI300X hardware.\n"
"📖 **Source:** [GitHub](https://github.com/maximolopezchenlo-lab/OncoAgent)"
)
# Stream it
partial = ""
chunk_size = 12
for i in range(0, len(response), chunk_size):
partial += response[i:i + chunk_size]
display = history.copy()
display.append({"role": "assistant", "content": partial})
yield display
time.sleep(0.01)
history.append({"role": "assistant", "content": response})
yield history
# ── Build the UI ──────────────────────────────────────────────────────
HEADER_HTML: str = """
"""
INFO_HTML: str = """
🏥 Multi-Agent Oncology Triage System
OncoAgent uses a 5-agent LangGraph pipeline
to analyze clinical cases against NCCN/ESMO guidelines
with built-in safety validation and anti-hallucination guardrails.
🔀 Router
→
🔍 Extraction
→
📚 Corrective RAG
→
🧠 Specialist
→
✅ Critic
⚡ Production: Qwen3.5-9B (Tier 1) + Qwen3.6-27B (Tier 2) via vLLM PagedAttention
| 📄 162 NCCN + 16 ESMO guidelines indexed
"""
FOOTER_HTML: str = """
"""
with gr.Blocks(
css=CSS,
head=FONTS_LINK,
title="OncoAgent — Oncology Triage Demo",
theme=gr.themes.Base(),
) as demo:
# Header
gr.HTML(HEADER_HTML)
gr.HTML(INFO_HTML)
# Chat
chatbot = gr.Chatbot(
type="messages",
label="Clinical Triage Chat",
height=520,
show_label=False,
show_copy_button=True,
render_markdown=True,
elem_classes=["card"],
)
# Controls
with gr.Row():
with gr.Column(scale=3):
txt = gr.Textbox(
placeholder="Enter a clinical case or click '▶ View Demo'...",
show_label=False,
lines=2,
max_lines=5,
)
with gr.Column(scale=1, min_width=180):
demo_btn = gr.Button(
"▶ View Demo",
elem_classes=["btn-demo"],
size="lg",
)
with gr.Row():
send_btn = gr.Button("Send", elem_classes=["btn-primary"], size="sm")
clear_btn = gr.Button("🗑 Clear", variant="secondary", size="sm")
# Footer
gr.HTML(FOOTER_HTML)
# ── Event Handlers ────────────────────────────────────────────────
demo_btn.click(
fn=run_demo,
inputs=None,
outputs=chatbot,
)
send_btn.click(
fn=handle_user_message,
inputs=[txt, chatbot],
outputs=chatbot,
).then(lambda: "", outputs=txt)
txt.submit(
fn=handle_user_message,
inputs=[txt, chatbot],
outputs=chatbot,
).then(lambda: "", outputs=txt)
clear_btn.click(lambda: [], outputs=chatbot)
if __name__ == "__main__":
demo.launch(server_name="0.0.0.0", server_port=7860)