""" FastAPI + Gradio server for the Cashflow Multi-Agent RL Environment. """ from fastapi import FastAPI import gradio as gr import pandas as pd import time import json from openenv.core.env_server.http_server import create_app import sys import os print(f"DEBUG: Starting server from CWD: {os.getcwd()}") root_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), '..')) print(f"DEBUG: Root dir added to path: {root_dir}") sys.path.insert(0, root_dir) print(f"DEBUG: sys.path[0]: {sys.path[0]}") try: from models import CashflowmanagerAction, CashflowmanagerObservation from server.cashflowmanager_environment import CashflowmanagerEnvironment print("DEBUG: Successfully imported from root level") except ImportError as e: print(f"DEBUG: Root import failed: {e}") try: from cashflowmanager.models import CashflowmanagerAction, CashflowmanagerObservation from cashflowmanager.server.cashflowmanager_environment import CashflowmanagerEnvironment print("DEBUG: Successfully imported from cashflowmanager package") except ImportError as e2: print(f"DEBUG: Package import failed: {e2}") from ..models import CashflowmanagerAction, CashflowmanagerObservation from .cashflowmanager_environment import CashflowmanagerEnvironment print("DEBUG: Using relative imports") app: FastAPI = create_app( CashflowmanagerEnvironment, CashflowmanagerAction, CashflowmanagerObservation, env_name="cashflowmanager", max_concurrent_envs=1, ) print("DEBUG: FastAPI App created successfully") try: from server.client import groq_policy, clear_action_cache except ImportError: try: from cashflowmanager.server.client import groq_policy, clear_action_cache except ImportError: from .client import groq_policy, clear_action_cache # Global state for the interactive UI _env_instance = None _last_obs = None _history = [] def get_env(seed=42, difficulty="medium"): global _env_instance, _last_obs, _history if _env_instance is None: _env_instance = CashflowmanagerEnvironment() clear_action_cache() _last_obs = _env_instance.reset(seed=seed, difficulty=difficulty) _history = [] return _env_instance, _last_obs def format_invoices(invoices): if not invoices: return "No active invoices." rows = [] for inv in invoices: urgency = "🔴 OVERDUE" if inv.due_in <= 0 else "🟡 URGENT" if inv.due_in <= 2 else "đŸŸĸ OK" rows.append(f"{urgency} | {inv.id} | ₹{inv.amount:.0f} | Due: {inv.due_in}d | Vendor: {inv.vendor_id}") return "\n".join(rows) def format_receivables(receivables): if not receivables: return "No expected inflows." rows = [] for rec in receivables: rows.append(f"₹{rec.amount:.0f} from {rec.customer_id} in {rec.expected_in}d (prob: {rec.probability*100:.0f}%)") return "\n".join(rows) def process_step(action_type, invoice_id=None, amount=0.0, memo=None): global _env_instance, _last_obs, _history env, obs = get_env() if obs.done: return update_ui() # Create action action = CashflowmanagerAction(type=action_type, invoice_id=invoice_id, amount=amount, memo=memo) # Step environment new_obs = env.step(action) # Log to history entry = { "Step": env.state.step_count, "Day": new_obs.day, "Action": f"{action.type}({action.invoice_id or 'N/A'})", "Amount": f"₹{action.amount:.0f}" if action.amount else "N/A", "Cash": f"₹{new_obs.cash:.0f}", "Reward": round(new_obs.reward, 2), "Reasoning": action.memo or "Manual Action", "Events": " | ".join(new_obs.world_events) if new_obs.world_events else "None" } _history.insert(0, entry) _last_obs = new_obs return update_ui() def ai_step(): global _last_obs if _last_obs is None or _last_obs.done: return update_ui() # Let the policy decide action = groq_policy(_last_obs, []) return process_step(action.type, action.invoice_id, action.amount, memo=action.memo) def reset_sim(seed, difficulty): global _env_instance, _last_obs, _history _env_instance = None get_env(seed=int(seed), difficulty=difficulty) return update_ui() def update_ui(): global _last_obs, _history obs = _last_obs history_df = pd.DataFrame(_history) status = f"### Day {obs.day} | Step {obs.metadata.get('step', 0)}\n" status += f"**Cash:** ₹{obs.cash:.0f} | **Credit Used:** ₹{obs.credit_used:.0f}/{obs.credit_limit:.0f}\n" if obs.done: status += "## 🏁 EPISODE FINISHED\n" memos = "#### 🤖 Advisor Memos\n" for agent, msg in obs.advisor_messages.items(): memos += f"- **{agent}:** {msg}\n" world = "#### 🌍 World Events\n" if obs.world_events: for e in obs.world_events: world += f"- {e}\n" else: world += "- No events this step." invoice_list = [inv.id for inv in obs.invoices] return ( status, memos, world, format_invoices(obs.invoices), format_receivables(obs.receivables), history_df, gr.Dropdown(choices=invoice_list, value=invoice_list[0] if invoice_list else None) ) def build_ui(): with gr.Blocks(title="Cashflow Multi-Agent RL Simulator") as demo: gr.Markdown("# đŸĸ Cashflow Management Dashboard") with gr.Row(): seed_input = gr.Number(value=42, label="Sim Seed", precision=0, scale=1) difficulty_input = gr.Dropdown(choices=["easy", "medium", "hard"], value="medium", label="Difficulty", scale=1) reset_btn = gr.Button("🔄 Reset", variant="secondary", scale=1) ai_btn = gr.Button("🤖 AI Next Step", variant="primary", scale=2) with gr.Row(): # --- Left: Data View --- with gr.Column(scale=3): status_md = gr.Markdown("### 💰 Financial Status") with gr.Tabs(): with gr.TabItem("📋 Active Invoices"): invoice_display = gr.Code(label="Debts", language="markdown") with gr.TabItem("📈 Receivables"): receivable_display = gr.Code(label="Expected Inflows", language="markdown") with gr.Group(): gr.Markdown("#### 🕹ī¸ Manual Action") with gr.Row(): target_inv = gr.Dropdown(label="Select Invoice", choices=[], scale=2) pay_amount = gr.Number(label="Amount (₹)", value=0, scale=1) with gr.Row(): pay_btn = gr.Button("Pay Full", variant="stop") neg_btn = gr.Button("Negotiate", variant="primary") credit_btn = gr.Button("Draw Credit") defer_btn = gr.Button("Defer Step") # --- Right: Intelligence --- with gr.Column(scale=2): memo_md = gr.Markdown("#### 🤖 Advisor Intelligence") world_md = gr.Markdown("#### 🌍 World Events") gr.Markdown("---") history_table = gr.Dataframe( headers=["Step", "Day", "Action", "Amount", "Cash", "Reward", "Reasoning", "Events"], interactive=False ) # Event handlers demo.load(reset_sim, inputs=[seed_input, difficulty_input], outputs=[status_md, memo_md, world_md, invoice_display, receivable_display, history_table, target_inv]) reset_btn.click(reset_sim, inputs=[seed_input, difficulty_input], outputs=[status_md, memo_md, world_md, invoice_display, receivable_display, history_table, target_inv]) difficulty_input.change(reset_sim, inputs=[seed_input, difficulty_input], outputs=[status_md, memo_md, world_md, invoice_display, receivable_display, history_table, target_inv]) ai_btn.click(ai_step, outputs=[status_md, memo_md, world_md, invoice_display, receivable_display, history_table, target_inv]) pay_btn.click(lambda id, amt: process_step("pay", id, amt), inputs=[target_inv, pay_amount], outputs=[status_md, memo_md, world_md, invoice_display, receivable_display, history_table, target_inv]) neg_btn.click(lambda id: process_step("negotiate", id), inputs=[target_inv], outputs=[status_md, memo_md, world_md, invoice_display, receivable_display, history_table, target_inv]) credit_btn.click(lambda amt: process_step("credit", amount=amt), inputs=[pay_amount], outputs=[status_md, memo_md, world_md, invoice_display, receivable_display, history_table, target_inv]) defer_btn.click(lambda: process_step("defer"), outputs=[status_md, memo_md, world_md, invoice_display, receivable_display, history_table, target_inv]) return demo gradio_app = build_ui() app = gr.mount_gradio_app(app, gradio_app, path="/ui") def main(): import uvicorn uvicorn.run(app, host="0.0.0.0", port=7861) if __name__ == "__main__": main()