import gradio as gr from app.agent.graph import bfsi_agent def process_voice_query(audio_path, customer_id): if not audio_path: return "No audio provided.", None, "" initial_state = { "audio_path": audio_path, "customer_id": customer_id, "messages": [] } try: # Run the full agent pipeline result = bfsi_agent.invoke(initial_state) if result.get("error"): return f"❌ Error: {result['error']}", None, "" transcript = result.get("transcript", "N/A") intent = result.get("intent", "unknown") response_text = result.get("response_text", "N/A") audio_out = result.get("response_audio_path", None) ticket = result.get("ticket_id", "None") # Format the debug/observability log for the UI debug_log = f"🎤 **Transcript**: {transcript}\n" debug_log += f"🧠 **Detected Intent**: `{intent.upper()}`\n" debug_log += f"🧑‍💼 **Customer Profile Used**: {customer_id}\n" if ticket != "None": debug_log += f"🎫 **Escalation Ticket Generated**: {ticket}\n" return response_text, audio_out, debug_log except Exception as e: return f"Pipeline Error: {str(e)}", None, "Traceback failed." # Apply a polished, enterprise-style theme suitable for BFSI bfsi_theme = gr.themes.Soft( primary_hue="blue", secondary_hue="slate", neutral_hue="slate" ) with gr.Blocks(theme=bfsi_theme, title="FinVoice: Production BFSI Support") as ui: gr.Markdown( """ # 🏦 FinVoice: Production BFSI Support *Powered by LangGraph, Groq, FAISS, and Whisper* Demonstrating context-aware voice AI that executes real workflows instead of just chatting. """ ) with gr.Row(): with gr.Column(scale=1): customer_selector = gr.Dropdown( choices=["C1024", "UNKNOWN"], value="C1024", label="Simulated Caller ID (Context Injection)" ) audio_input = gr.Audio(type="filepath", label="🗣️ Speak to the Agent (Mock Mic/Upload)") submit_btn = gr.Button("🎙️ Process Query", variant="primary") with gr.Column(scale=2): agent_text_output = gr.Textbox(label="Agent Text Response", interactive=False, lines=4) agent_voice_output = gr.Audio(label="Agent Voice Response", interactive=False) with gr.Accordion("⚙️ Backend Trace / Pipeline Observability", open=True): debug_output = gr.Markdown("Waiting for input...") submit_btn.click( fn=process_voice_query, inputs=[audio_input, customer_selector], outputs=[agent_text_output, agent_voice_output, debug_output] )