amd-gaia 0.14.1__py3-none-any.whl → 0.14.3__py3-none-any.whl

gaia/agents/emr/cli.py

@@ -0,0 +1,1322 @@
+# Copyright(C) 2025-2026 Advanced Micro Devices, Inc. All rights reserved.
+# SPDX-License-Identifier: MIT
+
+"""CLI for Medical Intake Agent."""
+
+import argparse
+import logging
+import sys
+from pathlib import Path
+
+from rich.console import Console
+from rich.panel import Panel
+from rich.table import Table
+
+from gaia.agents.emr.agent import MedicalIntakeAgent
+
+logger = logging.getLogger(__name__)
+console = Console()
+
+
+def _print_header(watch_dir: str, db_path: str):
+    """Print a styled header for the CLI."""
+    console.print()
+    console.print(
+        Panel.fit(
+            "[bold cyan]Medical Intake Agent[/bold cyan]\n"
+            "[dim]Automatic Patient Form Processing[/dim]",
+            border_style="cyan",
+        )
+    )
+
+    # Status table
+    table = Table(show_header=False, box=None, padding=(0, 2))
+    table.add_column(style="dim")
+    table.add_column()
+    table.add_row("📁 Watch folder:", watch_dir)
+    table.add_row("💾 Database:", db_path)
+    console.print(table)
+    console.print()
+
+    # Commands help
+    console.print("[dim]Commands:[/dim]")
+    console.print("  [cyan]stats[/cyan]      Show processing statistics")
+    console.print("  [cyan]quit[/cyan]       Stop and exit")
+    console.print("  [dim]Or type questions about patients[/dim]")
+    console.print()
+
+
+def _print_prompt():
+    """Print the input prompt with visual separators."""
+    console.print("─" * 80, style="dim")
+    console.print("> ", end="", style="bold green")
+    sys.stdout.flush()  # Ensure prompt is displayed before input() blocks
+
+
+def cmd_watch(args):
+    """Start watching directory for intake forms."""
+    _print_header(args.watch_dir, args.db)
+
+    console.print("[dim]Starting agent...[/dim]")
+
+    agent = MedicalIntakeAgent(
+        watch_dir=args.watch_dir,
+        db_path=args.db,
+        vlm_model=args.vlm_model,
+    )
+
+    console.print("[green]✓ Ready![/green] Drop intake forms to process them.\n")
+    sys.stdout.flush()  # Ensure Ready message appears before prompt
+
+    try:
+        while True:
+            try:
+                _print_prompt()
+                user_input = input().strip()
+            except EOFError:
+                break
+
+            if not user_input:
+                continue
+
+            if user_input.lower() in ("quit", "exit", "q"):
+                break
+
+            console.print("─" * 80, style="dim")
+
+            if user_input.lower() == "stats":
+                cmd_stats_inline(agent)
+            else:
+                # Process the query
+                agent.process_query(user_input)
+            print()
+
+    except KeyboardInterrupt:
+        print()
+    finally:
+        console.print("[dim]Stopping agent...[/dim]")
+        agent.stop()
+        console.print("[green]✓ Stopped[/green]")
+
+
+def cmd_stats_inline(agent):
+    """Show stats inline during watch mode."""
+    try:
+        stats = agent.get_stats()
+        _print_stats_table(stats)
+    except Exception as e:
+        console.print(f"[red]Error getting stats: {e}[/red]")
+
+
+def cmd_process(args):
+    """Process a single intake form file."""
+    if not Path(args.file).exists():
+        console.print(f"[red]Error: File not found: {args.file}[/red]")
+        return 1
+
+    console.print(f"[dim]Processing: {args.file}[/dim]")
+
+    agent = MedicalIntakeAgent(
+        watch_dir=args.watch_dir,
+        db_path=args.db,
+        vlm_model=args.vlm_model,
+        auto_start_watching=False,
+    )
+
+    try:
+        # pylint: disable=protected-access
+        patient_data = agent._process_intake_form(args.file)
+
+        if patient_data:
+            # Agent already prints success and patient details
+            return 0
+        else:
+            console.print(f"[red]Failed to process: {args.file}[/red]")
+            return 1
+
+    finally:
+        agent.stop()
+
+
+def cmd_query(args):
+    """Query patient database."""
+    agent = MedicalIntakeAgent(
+        watch_dir=args.watch_dir,
+        db_path=args.db,
+        auto_start_watching=False,
+    )
+
+    try:
+        agent.process_query(args.question)
+        return 0
+    finally:
+        agent.stop()
+
+
+def _print_stats_table(stats: dict):
+    """Print statistics using Rich formatting."""
+    console.print()
+
+    # Time savings highlight
+    time_table = Table(show_header=False, box=None, padding=(0, 1))
+    time_table.add_column(style="bold green")
+    time_table.add_column(style="green")
+    time_table.add_row(
+        f"⏱️  {stats['time_saved_minutes']} min saved",
+        f"({stats['time_saved_percent']} faster)",
+    )
+    console.print(Panel(time_table, title="Time Savings", border_style="green"))
+
+    # Main stats grid
+    grid = Table.grid(expand=True, padding=(0, 2))
+    grid.add_column()
+    grid.add_column()
+
+    # Patients table
+    patients = Table(show_header=False, box=None)
+    patients.add_column(style="dim")
+    patients.add_column(style="bold")
+    patients.add_row("Total", str(stats["total_patients"]))
+    patients.add_row("New", str(stats["new_patients"]))
+    patients.add_row("Returning", str(stats["returning_patients"]))
+    patients.add_row("Today", str(stats["processed_today"]))
+
+    # Processing table
+    processing = Table(show_header=False, box=None)
+    processing.add_column(style="dim")
+    processing.add_column(style="bold")
+    processing.add_row("Processed", str(stats["files_processed"]))
+    processing.add_row("Success", str(stats["extraction_success"]))
+    processing.add_row("Failed", str(stats["extraction_failed"]))
+    processing.add_row("Rate", stats["success_rate"])
+
+    grid.add_row(
+        Panel(patients, title="👥 Patients", border_style="cyan"),
+        Panel(processing, title="📋 Processing", border_style="cyan"),
+    )
+    console.print(grid)
+
+    # Alerts (if any)
+    if stats.get("unacknowledged_alerts", 0) > 0:
+        console.print(
+            f"[bold red]🚨 {stats['unacknowledged_alerts']} unacknowledged alert(s)[/bold red]"
+        )
+    console.print()
+
+
+def cmd_stats(args):
+    """Show processing statistics."""
+    agent = MedicalIntakeAgent(
+        watch_dir=args.watch_dir,
+        db_path=args.db,
+        auto_start_watching=False,
+        silent_mode=True,
+    )
+
+    try:
+        stats = agent.get_stats()
+        _print_stats_table(stats)
+        return 0
+    finally:
+        agent.stop()
+
+
+def cmd_reset(args):
+    """Reset by deleting the database file."""
+    import os
+
+    from rich.prompt import Confirm
+
+    db_path = Path(args.db)
+
+    # Check if database exists
+    if not db_path.exists():
+        console.print("[dim]Database file does not exist. Nothing to reset.[/dim]")
+        return 0
+
+    # Get stats before deletion to show what will be deleted
+    total_patients = 0
+    agent = None
+    try:
+        agent = MedicalIntakeAgent(
+            watch_dir=args.watch_dir,
+            db_path=args.db,
+            auto_start_watching=False,
+            silent_mode=True,
+        )
+        stats = agent.get_stats()
+        total_patients = stats.get("total_patients", 0)
+    except Exception:
+        pass  # If we can't read stats, proceed anyway
+    finally:
+        if agent:
+            agent.stop()
+
+    # Confirmation prompt unless --force is used
+    if not args.force:
+        console.print()
+        console.print(
+            "[bold yellow]⚠️  WARNING:[/bold yellow] This will permanently delete:"
+        )
+        if total_patients > 0:
+            console.print(f"  • {total_patients} patient record(s)")
+            console.print("  • All associated alerts and intake sessions")
+        console.print(f"  • Database file: {db_path}")
+        console.print()
+
+        if not Confirm.ask("[bold red]Are you sure you want to continue?[/bold red]"):
+            console.print("[dim]Cancelled.[/dim]")
+            return 0
+
+    # Delete the database file
+    try:
+        os.remove(db_path)
+        console.print()
+        console.print("[bold green]✓ Database deleted successfully[/bold green]")
+        console.print(f"  Removed: {db_path}")
+        console.print()
+        console.print(
+            "[dim]A fresh database will be created when you next run the agent.[/dim]"
+        )
+        return 0
+    except Exception as e:
+        console.print(f"[red]Error deleting database: {e}[/red]")
+        return 1
+
+
+def cmd_init(args):
+    """Initialize EMR agent by downloading and loading required models."""
+    import time
+
+    from gaia.llm.lemonade_client import LemonadeClient
+
+    console.print()
+    console.print(
+        Panel.fit(
+            "[bold cyan]EMR Agent Setup[/bold cyan]\n"
+            "[dim]Downloading and loading required models[/dim]",
+            border_style="cyan",
+        )
+    )
+    console.print()
+
+    # Required models for EMR agent
+    vlm_model = args.vlm_model  # Default: Qwen3-VL-4B-Instruct-GGUF
+    llm_model = "Qwen3-Coder-30B-A3B-Instruct-GGUF"  # For chat/query processing
+    embed_model = "nomic-embed-text-v2-moe-GGUF"  # For similarity search
+
+    REQUIRED_CONTEXT_SIZE = 32768
+
+    # Step 1: Check Lemonade server and context size
+    console.print("[bold]Step 1:[/bold] Checking Lemonade server...")
+    try:
+        client = LemonadeClient(model=vlm_model)
+        health = client.health_check()
+        if health.get("status") == "ok":
+            console.print("  [green]✓[/green] Lemonade server is running")
+
+            # Check context size
+            context_size = health.get("context_size", 0)
+            if context_size >= REQUIRED_CONTEXT_SIZE:
+                console.print(
+                    f"  [green]✓[/green] Context size: [cyan]{context_size:,}[/cyan] tokens (recommended: {REQUIRED_CONTEXT_SIZE:,})"
+                )
+            elif context_size > 0:
+                console.print(
+                    f"  [yellow]⚠[/yellow] Context size: [yellow]{context_size:,}[/yellow] tokens"
+                )
+                console.print(
+                    f"    [yellow]Warning:[/yellow] Context size should be at least [cyan]{REQUIRED_CONTEXT_SIZE:,}[/cyan] for reliable form processing"
+                )
+                console.print(
+                    "    [dim]To fix: Right-click Lemonade tray → Settings → Context Size → 32768[/dim]"
+                )
+            else:
+                console.print(
+                    "  [dim]Context size: Not reported (will check after model load)[/dim]"
+                )
+        else:
+            console.print("  [red]✗[/red] Lemonade server not responding")
+            console.print()
+            console.print("[yellow]Please start Lemonade server first:[/yellow]")
+            console.print("  1. Open Lemonade from the system tray")
+            console.print("  2. Or run: [cyan]lemonade-server[/cyan]")
+            return 1
+    except Exception as e:
+        console.print(f"  [red]✗[/red] Cannot connect to Lemonade: {e}")
+        console.print()
+        console.print("[yellow]Please start Lemonade server first:[/yellow]")
+        console.print("  1. Open Lemonade from the system tray")
+        console.print("  2. Or run: [cyan]lemonade-server[/cyan]")
+        return 1
+
+    # Step 2: Check required models
+    console.print()
+    console.print("[bold]Step 2:[/bold] Checking required models...")
+
+    try:
+        models_response = client.list_models()
+        available_models = models_response.get("data", [])
+        downloaded_model_ids = [m.get("id", "") for m in available_models]
+
+        # Check each required model
+        required_models = [
+            ("VLM", vlm_model, "Form extraction"),
+            ("LLM", llm_model, "Chat/query processing"),
+            ("Embedding", embed_model, "Similarity search"),
+        ]
+
+        models_to_download = []
+        for model_type, model_name, _purpose in required_models:
+            is_downloaded = model_name in downloaded_model_ids
+            if is_downloaded:
+                console.print(
+                    f"  [green]✓[/green] {model_type}: [cyan]{model_name}[/cyan]"
+                )
+            else:
+                console.print(
+                    f"  [dim]○[/dim] {model_type}: [cyan]{model_name}[/cyan] [dim](not downloaded)[/dim]"
+                )
+                models_to_download.append((model_type, model_name))
+
+        if models_to_download:
+            console.print()
+            console.print(
+                f"  [yellow]⚠️  {len(models_to_download)} model(s) need to be downloaded[/yellow]"
+            )
+
+    except Exception as e:
+        console.print(f"  [red]✗[/red] Failed to check models: {e}")
+        return 1
+
+    # Step 3: Load all required models
+    console.print()
+    console.print("[bold]Step 3:[/bold] Loading required models...")
+    console.print("  [dim]Loading models into memory for fast inference...[/dim]")
+    console.print()
+
+    models_loaded = {}
+
+    # Load VLM model first (most important for form processing)
+    for model_type, model_name in [
+        ("VLM", vlm_model),
+        ("LLM", llm_model),
+        ("Embedding", embed_model),
+    ]:
+        console.print(f"  Loading {model_type}: [cyan]{model_name}[/cyan]...")
+
+        try:
+            start_time = time.time()
+            client.load_model(model_name, timeout=1800, auto_download=True)
+            elapsed = time.time() - start_time
+            models_loaded[model_type] = True
+            console.print(f"  [green]✓[/green] {model_type} loaded ({elapsed:.1f}s)")
+        except Exception as e:
+            error_msg = str(e)
+            models_loaded[model_type] = False
+
+            # Check for common errors
+            if "being used by another process" in error_msg:
+                console.print(
+                    f"  [yellow]![/yellow] {model_type}: File locked, try again later"
+                )
+            elif (
+                "not found" in error_msg.lower()
+                or "does not exist" in error_msg.lower()
+            ):
+                console.print(
+                    f"  [yellow]![/yellow] {model_type}: Model not available in registry"
+                )
+            else:
+                console.print(f"  [yellow]![/yellow] {model_type}: {error_msg[:50]}...")
+
+    # Check if critical models loaded
+    if not models_loaded.get("VLM"):
+        console.print()
+        console.print(
+            "[red]✗ VLM model failed to load - form processing will not work[/red]"
+        )
+        return 1
+
+    # Clear VLM context to ensure fresh memory allocation
+    console.print()
+    console.print("  [dim]Clearing VLM context for clean memory...[/dim]")
+    try:
+        client.unload_model()
+        client.load_model(vlm_model, timeout=300, auto_download=True)
+        console.print("  [green]✓[/green] VLM context cleared")
+    except Exception as e:
+        console.print(f"  [dim]Context clear skipped: {e}[/dim]")
+
+    # Step 4: Verify models and check context size
+    console.print()
+    console.print("[bold]Step 4:[/bold] Verifying models are ready...")
+
+    vlm_ready = False
+    llm_ready = False
+    embed_ready = False
+    final_context_size = 0
+
+    try:
+        # Check health for context size
+        health = client.health_check()
+        final_context_size = health.get("context_size", 0)
+
+        # Check each model
+        vlm_ready = client.check_model_loaded(vlm_model)
+        llm_ready = client.check_model_loaded(llm_model)
+        embed_ready = client.check_model_loaded(embed_model)
+
+        if vlm_ready:
+            console.print("  [green]✓[/green] VLM: Ready for form extraction")
+        else:
+            console.print("  [yellow]![/yellow] VLM: Will load on first use")
+
+        if llm_ready:
+            console.print("  [green]✓[/green] LLM: Ready for chat queries")
+        else:
+            console.print("  [dim]○[/dim] LLM: Will load on first use")
+
+        if embed_ready:
+            console.print("  [green]✓[/green] Embedding: Ready for search")
+        else:
+            console.print("  [dim]○[/dim] Embedding: Will load on first use")
+
+        # Report context size
+        if final_context_size >= REQUIRED_CONTEXT_SIZE:
+            console.print(
+                f"  [green]✓[/green] Context size: [cyan]{final_context_size:,}[/cyan] tokens"
+            )
+        elif final_context_size > 0:
+            console.print(
+                f"  [yellow]⚠[/yellow] Context size: [yellow]{final_context_size:,}[/yellow] tokens (need {REQUIRED_CONTEXT_SIZE:,})"
+            )
+
+    except Exception as e:
+        console.print(f"  [yellow]![/yellow] Could not verify: {e}")
+
+    # Step 5: Show all downloaded and loaded models
+    console.print()
+    console.print("[bold]Step 5:[/bold] Model inventory...")
+
+    try:
+        models_response = client.list_models()
+        all_models = models_response.get("data", [])
+
+        # Categorize models
+        vlm_models = []
+        llm_models = []
+        embed_models = []
+
+        for m in all_models:
+            model_id = m.get("id", "")
+            model_lower = model_id.lower()
+
+            if "vl" in model_lower or "vision" in model_lower or "vlm" in model_lower:
+                vlm_models.append(model_id)
+            elif (
+                "embed" in model_lower
+                or "bge" in model_lower
+                or "e5" in model_lower
+                or "nomic" in model_lower
+            ):
+                embed_models.append(model_id)
+            else:
+                llm_models.append(model_id)
+
+        # Show categorized models
+        if vlm_models:
+            console.print(
+                f"  [cyan]VLM Models:[/cyan] {', '.join(vlm_models[:3])}"
+                + (f" (+{len(vlm_models)-3} more)" if len(vlm_models) > 3 else "")
+            )
+        if llm_models:
+            console.print(
+                f"  [cyan]LLM Models:[/cyan] {', '.join(llm_models[:3])}"
+                + (f" (+{len(llm_models)-3} more)" if len(llm_models) > 3 else "")
+            )
+        if embed_models:
+            console.print(
+                f"  [cyan]Embedding Models:[/cyan] {', '.join(embed_models[:3])}"
+                + (f" (+{len(embed_models)-3} more)" if len(embed_models) > 3 else "")
+            )
+
+        console.print(f"  [dim]Total models available: {len(all_models)}[/dim]")
+
+    except Exception as e:
+        console.print(f"  [dim]Could not list models: {e}[/dim]")
+
+    # Success summary
+    console.print()
+
+    # Build model status lines
+    model_status_lines = []
+
+    # VLM status
+    if vlm_ready:
+        model_status_lines.append(
+            f"[green]✓[/green] VLM: [cyan]{vlm_model}[/cyan] - Ready"
+        )
+    else:
+        model_status_lines.append(
+            f"[yellow]![/yellow] VLM: [cyan]{vlm_model}[/cyan] - Will load on first use"
+        )
+
+    # LLM status
+    if llm_ready:
+        model_status_lines.append(
+            f"[green]✓[/green] LLM: [cyan]{llm_model}[/cyan] - Ready"
+        )
+    else:
+        model_status_lines.append(
+            f"[dim]○[/dim] LLM: [cyan]{llm_model}[/cyan] - Will load on first use"
+        )
+
+    # Embedding status
+    if embed_ready:
+        model_status_lines.append(
+            f"[green]✓[/green] Embedding: [cyan]{embed_model}[/cyan] - Ready"
+        )
+    else:
+        model_status_lines.append(
+            f"[dim]○[/dim] Embedding: [cyan]{embed_model}[/cyan] - Will load on first use"
+        )
+
+    # Context size status
+    if final_context_size >= REQUIRED_CONTEXT_SIZE:
+        model_status_lines.append(
+            f"[green]✓[/green] Context size: {final_context_size:,} tokens"
+        )
+    elif final_context_size > 0:
+        model_status_lines.append(
+            f"[yellow]⚠[/yellow] Context size: {final_context_size:,} tokens (need {REQUIRED_CONTEXT_SIZE:,})"
+        )
+
+    # Count ready models
+    ready_count = sum([vlm_ready, llm_ready, embed_ready])
+
+    console.print(
+        Panel.fit(
+            f"[bold green]✓ EMR Agent initialized ({ready_count}/3 models ready)[/bold green]\n\n"
+            + "\n".join(model_status_lines)
+            + "\n\n"
+            "[dim]You can now run:[/dim]\n"
+            "  [cyan]gaia-emr dashboard[/cyan]  - Start the web dashboard\n"
+            "  [cyan]gaia-emr watch[/cyan]      - Watch folder for intake forms\n"
+            "  [cyan]gaia-emr process <file>[/cyan] - Process a single file",
+            border_style="green",
+        )
+    )
+    console.print()
+
+    # Context size warning if needed
+    if 0 < final_context_size < REQUIRED_CONTEXT_SIZE:
+        console.print(
+            Panel.fit(
+                "[yellow]⚠️  Context Size Warning[/yellow]\n\n"
+                f"Current context size ({final_context_size:,}) may be too small for processing intake forms.\n"
+                "Large images can require 4,000-8,000+ tokens.\n\n"
+                "[bold]To fix:[/bold]\n"
+                "  1. Right-click Lemonade tray icon → Settings\n"
+                "  2. Set Context Size to [cyan]32768[/cyan]\n"
+                "  3. Click Apply and restart the model",
+                border_style="yellow",
+            )
+        )
+        console.print()
+
+    return 0
+
+
+def cmd_test(args):
+    """Test VLM extraction pipeline on a single file."""
+    import io
+    import json
+    import time
+
+    from PIL import Image
+
+    from gaia.llm.vlm_client import VLMClient
+
+    file_path = Path(args.file)
+    if not file_path.exists():
+        console.print(f"[red]Error: File not found: {file_path}[/red]")
+        return 1
+
+    console.print()
+    console.print(
+        Panel.fit(
+            "[bold cyan]EMR Agent - VLM Pipeline Test[/bold cyan]\n"
+            f"[dim]Testing extraction on: {file_path.name}[/dim]",
+            border_style="cyan",
+        )
+    )
+    console.print()
+
+    # Step 1: Read and analyze file
+    console.print("[bold]Step 1:[/bold] Reading file...")
+    try:
+        raw_bytes = file_path.read_bytes()
+        file_size_kb = len(raw_bytes) / 1024
+        console.print(f"  File size: {file_size_kb:.1f} KB")
+
+        # Get image dimensions
+        img = Image.open(io.BytesIO(raw_bytes))
+        orig_width, orig_height = img.size
+        console.print(f"  Dimensions: {orig_width}x{orig_height} pixels")
+
+        # Auto-rotate based on EXIF orientation
+        try:
+            from PIL import ExifTags
+
+            exif = img._getexif()  # pylint: disable=protected-access
+            if exif:
+                for tag, value in exif.items():
+                    if ExifTags.TAGS.get(tag) == "Orientation":
+                        if value == 3:
+                            img = img.rotate(180, expand=True)
+                            console.print("  [dim]Auto-rotated 180°[/dim]")
+                        elif value == 6:
+                            img = img.rotate(270, expand=True)
+                            console.print("  [dim]Auto-rotated 90° CW[/dim]")
+                        elif value == 8:
+                            img = img.rotate(90, expand=True)
+                            console.print("  [dim]Auto-rotated 90° CCW[/dim]")
+                        orig_width, orig_height = img.size
+                        break
+        except Exception:
+            pass  # No EXIF or rotation info
+    except Exception as e:
+        console.print(f"  [red]✗[/red] Failed to read file: {e}")
+        return 1
+
+    # Step 2: Optimize image (same as agent)
+    console.print()
+    console.print("[bold]Step 2:[/bold] Optimizing image...")
+    max_dimension = args.max_dimension
+    jpeg_quality = args.jpeg_quality
+
+    try:
+        if orig_width > max_dimension or orig_height > max_dimension:
+            scale = min(max_dimension / orig_width, max_dimension / orig_height)
+            new_width = int(orig_width * scale)
+            new_height = int(orig_height * scale)
+            img = img.resize((new_width, new_height), Image.Resampling.LANCZOS)
+            console.print(
+                f"  Resized: {orig_width}x{orig_height} → {new_width}x{new_height}"
+            )
+        else:
+            new_width, new_height = orig_width, orig_height
+            console.print(f"  No resize needed (under {max_dimension}px)")
+
+        # Convert to RGB and JPEG
+        if img.mode in ("RGBA", "P"):
+            img = img.convert("RGB")
+
+        output = io.BytesIO()
+        img.save(output, format="JPEG", quality=jpeg_quality, optimize=True)
+        image_bytes = output.getvalue()
+
+        opt_size_kb = len(image_bytes) / 1024
+        reduction = (1 - opt_size_kb / file_size_kb) * 100
+        console.print(
+            f"  Optimized: {file_size_kb:.0f}KB → {opt_size_kb:.0f}KB ({reduction:.0f}% smaller)"
+        )
+
+        # Estimate image tokens (rough: ~1 token per 14x14 pixel patch)
+        est_img_tokens = (new_width // 14) * (new_height // 14)
+        console.print(f"  Est. image tokens: ~{est_img_tokens:,}")
+    except Exception as e:
+        console.print(f"  [red]✗[/red] Failed to optimize: {e}")
+        return 1
+
+    # Step 3: Initialize VLM
+    console.print()
+    console.print("[bold]Step 3:[/bold] Initializing VLM...")
+    try:
+        vlm = VLMClient(vlm_model=args.vlm_model)
+
+        # Clear context if requested (unload and reload model)
+        if getattr(args, "clear_context", False):
+            console.print("  [dim]Clearing VLM context (unload + reload)...[/dim]")
+            try:
+                vlm.client.unload_model()
+                vlm.client.load_model(args.vlm_model, timeout=300, auto_download=True)
+                console.print("  [green]✓[/green] Context cleared")
+            except Exception as e:
+                console.print(f"  [yellow]⚠[/yellow] Could not clear context: {e}")
+
+        console.print(f"  [green]✓[/green] VLM ready: [cyan]{vlm.vlm_model}[/cyan]")
+    except Exception as e:
+        console.print(f"  [red]✗[/red] Failed to initialize VLM: {e}")
+        return 1
+
+    # Step 4: Extract data with auto-retry on memory errors
+    console.print()
+    console.print("[bold]Step 4:[/bold] Extracting patient data...")
+
+    extraction_prompt = """Extract ALL patient information from this medical intake form.
+
+Return a JSON object with these fields (use null for missing/unclear):
+{
+    "form_date": "YYYY-MM-DD (date form was filled, today's date)",
+    "first_name": "...",
+    "last_name": "...",
+    "date_of_birth": "YYYY-MM-DD",
+    "age": "patient's age if listed",
+    "gender": "Male/Female/Other",
+    "preferred_pronouns": "he/him, she/her, they/them if listed",
+    "ssn": "XXX-XX-XXXX (social security number)",
+    "marital_status": "Single/Married/Divorced/Widowed/Partnered",
+    "spouse_name": "spouse's name if listed",
+    "phone": "home phone number",
+    "mobile_phone": "cell/mobile phone number",
+    "work_phone": "work phone number",
+    "email": "...",
+    "address": "street address",
+    "city": "...",
+    "state": "...",
+    "zip_code": "...",
+    "preferred_language": "English/Spanish/etc if listed",
+    "race": "if listed",
+    "ethnicity": "if listed",
+    "contact_preference": "preferred contact method if listed",
+    "emergency_contact_name": "name of emergency contact person",
+    "emergency_contact_relationship": "relationship to patient (e.g. Mom, Spouse, Friend)",
+    "emergency_contact_phone": "emergency contact's phone number",
+    "referring_physician": "name of referring physician/doctor",
+    "referring_physician_phone": "phone number next to referring physician",
+    "primary_care_physician": "PCP name if different from referring",
+    "preferred_pharmacy": "pharmacy name if listed",
+    "employment_status": "Employed/Self Employed/Unemployed/Retired/Student/Disabled/Military",
+    "occupation": "job title if listed",
+    "employer": "employer/company name",
+    "employer_address": "employer address if listed",
+    "insurance_provider": "insurance company name",
+    "insurance_id": "policy number",
+    "insurance_group_number": "group number",
+    "insured_name": "name of insured person (may differ from patient)",
+    "insured_dob": "YYYY-MM-DD",
+    "insurance_phone": "insurance contact number",
+    "billing_address": "billing address if different from home",
+    "guarantor_name": "person responsible for payment if listed",
+    "reason_for_visit": "chief complaint or reason for visit",
+    "date_of_injury": "YYYY-MM-DD (date of injury or onset of symptoms)",
+    "pain_location": "where pain is located if listed",
+    "pain_onset": "when pain began (e.g. three months ago)",
+    "pain_cause": "what caused the pain/condition",
+    "pain_progression": "Improved/Worsened/Stayed the same",
+    "work_related_injury": "Yes/No",
+    "car_accident": "Yes/No",
+    "medical_conditions": "existing medical conditions",
+    "allergies": "known allergies",
+    "medications": "current medications",
+    "signature_date": "YYYY-MM-DD (date signed)"
+}
+
+IMPORTANT: Return ONLY the JSON object, no other text."""
+
+    # Retry loop with progressively smaller images on memory errors
+    max_retries = 3
+    current_img = img
+    current_bytes = image_bytes
+    current_width, current_height = new_width, new_height
+    current_size_kb = opt_size_kb
+
+    for attempt in range(max_retries):
+        est_img_tokens = (current_width // 14) * (current_height // 14)
+        console.print(
+            f"  Image: {current_width}x{current_height}, {current_size_kb:.0f}KB (~{est_img_tokens:,} tokens)"
+        )
+
+        if attempt == 0:
+            console.print("  [dim]This may take 30-60 seconds...[/dim]")
+        else:
+            console.print(
+                f"  [dim]Retry {attempt}/{max_retries-1} with smaller image...[/dim]"
+            )
+
+        try:
+            start_time = time.time()
+            raw_text = vlm.extract_from_image(
+                image_bytes=current_bytes,
+                prompt=extraction_prompt,
+            )
+            extraction_time = time.time() - start_time
+
+            # Check for memory-related errors
+            if (
+                "failed to process image" in raw_text
+                or "memory slot" in raw_text.lower()
+            ):
+                if attempt < max_retries - 1:
+                    console.print(
+                        "  [yellow]⚠[/yellow] Memory error, reducing image size..."
+                    )
+                    # Reduce image to 75% of current size
+                    scale = 0.75
+                    current_width = int(current_width * scale)
+                    current_height = int(current_height * scale)
+                    current_img = current_img.resize(
+                        (current_width, current_height), Image.Resampling.LANCZOS
+                    )
+                    output = io.BytesIO()
+                    current_img.save(
+                        output, format="JPEG", quality=jpeg_quality, optimize=True
+                    )
+                    current_bytes = output.getvalue()
+                    current_size_kb = len(current_bytes) / 1024
+                    continue
+                else:
+                    console.print(f"  [red]✗[/red] {raw_text}")
+                    console.print()
+                    console.print("[yellow]Suggestions:[/yellow]")
+                    console.print("  1. Try with smaller image: --max-dimension 640")
+                    console.print("  2. Restart Lemonade Server to clear memory")
+                    console.print("  3. Reload the VLM model in Lemonade")
+                    return 1
+
+            if raw_text.startswith("[VLM extraction failed:"):
+                console.print(f"  [red]✗[/red] {raw_text}")
+                return 1
+
+            # Success!
+            console.print(
+                f"  [green]✓[/green] Extraction complete ({len(raw_text)} chars, {extraction_time:.1f}s)"
+            )
+
+            # Estimate tokens/sec (output tokens only)
+            est_output_tokens = len(raw_text) / 4
+            tps = est_output_tokens / extraction_time if extraction_time > 0 else 0
+            console.print(
+                f"  Output: ~{est_output_tokens:.0f} tokens at ~{tps:.1f} TPS"
+            )
+
+            # Total throughput including prompt processing
+            total_tokens = (
+                est_img_tokens + 200 + est_output_tokens
+            )  # img + prompt + output
+            total_tps = total_tokens / extraction_time if extraction_time > 0 else 0
+            console.print(
+                f"  Total throughput: ~{total_tps:.0f} TPS (incl. {est_img_tokens:,} image tokens)"
+            )
+
+            # Update dimensions for final report
+            new_width, new_height = current_width, current_height
+            break
+
+        except Exception as e:
+            console.print(f"  [red]✗[/red] Extraction failed: {e}")
+            return 1
+    else:
+        # All retries exhausted
+        console.print("  [red]✗[/red] All retries failed")
+        return 1
+
+    # Step 5: Parse JSON
+    console.print()
+    console.print("[bold]Step 5:[/bold] Parsing JSON...")
+    try:
+        # Try direct parse
+        patient_data = json.loads(raw_text)
+        console.print("  [green]✓[/green] JSON parsed successfully")
+    except json.JSONDecodeError:
+        # Try to find JSON in text
+        try:
+            start = raw_text.find("{")
+            end = raw_text.rfind("}") + 1
+            if start >= 0 and end > start:
+                patient_data = json.loads(raw_text[start:end])
+                console.print("  [green]✓[/green] JSON extracted from response")
+            else:
+                console.print("  [red]✗[/red] No JSON found in response")
+                console.print()
+                console.print("[bold]Raw VLM Output:[/bold]")
+                console.print(raw_text[:500])
+                return 1
+        except json.JSONDecodeError as e:
+            console.print(f"  [red]✗[/red] JSON parse failed: {e}")
+            return 1
+
+    # Display results
+    console.print()
+    console.print(
+        Panel.fit(
+            "[bold green]✓ Extraction Successful[/bold green]",
+            border_style="green",
+        )
+    )
+
+    # Show extracted fields
+    console.print()
+    console.print("[bold]Extracted Fields:[/bold]")
+    fields_found = 0
+    for key, value in patient_data.items():
+        if value and value != "null":
+            fields_found += 1
+            console.print(f"  [cyan]{key}:[/cyan] {value}")
+
+    console.print()
+    console.print(f"[dim]Fields extracted: {fields_found}/53[/dim]")
+    console.print()
+
+    # Timing breakdown section
+    console.print("[bold]⏱️  TIMING BREAKDOWN[/bold]")
+    console.print("-" * 40)
+    console.print(f"   Model:              {args.vlm_model}")
+    console.print(f"   Image dimensions:   {new_width}x{new_height}")
+    console.print(f"   VLM extraction:     {extraction_time:.2f}s")
+    console.print(f"   Est. image tokens:  ~{est_img_tokens:,}")
+    console.print(f"   Est. output tokens: ~{int(est_output_tokens)}")
+    console.print(f"   Est. tokens/sec:    ~{tps:.1f} TPS")
+    console.print()
+
+    # Success summary
+    if patient_data.get("first_name") and patient_data.get("last_name"):
+        console.print(
+            f"   Patient: {patient_data.get('first_name', '')} {patient_data.get('last_name', '')}"
+        )
+        console.print("   [green]✓ Pipeline test PASSED[/green]")
+    else:
+        console.print("   [yellow]⚠ Missing required name fields[/yellow]")
+        console.print("   [red]✗ Pipeline test FAILED[/red]")
+    console.print()
+
+    return 0
+
+
+def _launch_electron(url: str, delay: float = 1.5) -> bool:
+    """
+    Launch Electron app to display the dashboard.
+
+    Returns True if Electron was launched successfully, False otherwise.
+    """
+    import os
+    import platform
+    import shutil
+    import subprocess
+    import time
+
+    time.sleep(delay)  # Wait for server to start
+
+    # Find the Electron wrapper directory
+    electron_dir = Path(__file__).parent / "dashboard" / "electron"
+    main_js = electron_dir / "main.js"
+
+    if not main_js.exists():
+        logger.warning(f"Electron wrapper not found at {electron_dir}")
+        return False
+
+    # On Windows, use npm.cmd and npx.cmd
+    is_windows = platform.system() == "Windows"
+    npm_cmd = "npm.cmd" if is_windows else "npm"
+    npx_cmd = "npx.cmd" if is_windows else "npx"
+
+    # Check if npx/electron is available
+    npx_path = shutil.which(npx_cmd)
+    if not npx_path:
+        logger.warning(f"{npx_cmd} not found in PATH, cannot launch Electron")
+        return False
+
+    try:
+        # Check if node_modules exists, if not run npm install first
+        node_modules = electron_dir / "node_modules"
+        if not node_modules.exists():
+            console.print("[dim]Installing Electron dependencies...[/dim]")
+            subprocess.run(
+                [npm_cmd, "install"],
+                cwd=str(electron_dir),
+                capture_output=True,
+                check=True,
+                shell=is_windows,  # Use shell on Windows for .cmd files
+            )
+
+        # Launch Electron with the dashboard URL
+        env = os.environ.copy()
+        env["EMR_DASHBOARD_URL"] = url
+
+        subprocess.Popen(
+            [npx_cmd, "electron", "."],
+            cwd=str(electron_dir),
+            env=env,
+            stdout=subprocess.DEVNULL,
+            stderr=subprocess.DEVNULL,
+            shell=is_windows,  # Use shell on Windows for .cmd files
+        )
+        return True
+    except Exception as e:
+        logger.warning(f"Failed to launch Electron: {e}")
+        return False
+
+
+def cmd_dashboard(args):
+    """Start web dashboard."""
+    import threading
+    import time
+    import webbrowser
+
+    def open_browser(url: str, delay: float = 1.5):
+        """Open browser after a short delay to allow server to start."""
+        time.sleep(delay)
+        webbrowser.open(url)
+
+    def open_electron_or_browser(url: str, use_browser: bool, delay: float = 1.5):
+        """Open Electron app, falling back to browser if needed."""
+        if use_browser:
+            open_browser(url, delay)
+        else:
+            if not _launch_electron(url, delay):
+                console.print(
+                    "[yellow]Electron not available, opening in browser instead...[/yellow]"
+                )
+                open_browser(url, delay)
+
+    try:
+        from gaia.agents.emr.dashboard.server import run_dashboard
+
+        console.print()
+        console.print(
+            Panel.fit(
+                "[bold cyan]EMR Dashboard[/bold cyan]\n"
+                "[dim]Real-time Patient Processing Monitor[/dim]",
+                border_style="cyan",
+            )
+        )
+
+        url = f"http://localhost:{args.port}"
+
+        table = Table(show_header=False, box=None, padding=(0, 2))
+        table.add_column(style="dim")
+        table.add_column()
+        table.add_row("📁 Watch folder:", args.watch_dir)
+        table.add_row("💾 Database:", args.db)
+        table.add_row("🌐 URL:", f"[bold cyan]{url}[/bold cyan]")
+        console.print(table)
+        console.print("\n[dim]Press Ctrl+C to stop[/dim]\n")
+
+        # Auto-open unless --no-open flag is set
+        if not getattr(args, "no_open", False):
+            use_browser = getattr(args, "browser", False)
+            if use_browser:
+                console.print("[dim]Opening dashboard in browser...[/dim]\n")
+            else:
+                console.print("[dim]Opening dashboard in Electron app...[/dim]\n")
+
+            open_thread = threading.Thread(
+                target=open_electron_or_browser,
+                args=(url, use_browser),
+                daemon=True,
+            )
+            open_thread.start()
+
+        run_dashboard(
+            watch_dir=args.watch_dir,
+            db_path=args.db,
+            host=args.host,
+            port=args.port,
+        )
+    except ImportError:
+        console.print("[red]Error: Dashboard dependencies not installed[/red]")
+        console.print("Install with: [cyan]pip install 'amd-gaia[api]'[/cyan]")
+        return 1
+    except KeyboardInterrupt:
+        console.print("\n[dim]Shutting down dashboard...[/dim]")
+        return 0
+
+
+def _add_common_args(parser):
+    """Add common arguments to a parser."""
+    parser.add_argument(
+        "--watch-dir",
+        default="./intake_forms",
+        help="Directory to watch for intake forms (default: ./intake_forms)",
+    )
+    parser.add_argument(
+        "--db",
+        default="./data/patients.db",
+        help="Path to patient database (default: ./data/patients.db)",
+    )
+    parser.add_argument(
+        "--vlm-model",
+        default="Qwen3-VL-4B-Instruct-GGUF",
+        help="VLM model for extraction (default: Qwen3-VL-4B-Instruct-GGUF)",
+    )
+    parser.add_argument(
+        "--debug",
+        action="store_true",
+        help="Enable debug logging",
+    )
+
+
+def main():
+    """Main CLI entry point."""
+    parser = argparse.ArgumentParser(
+        description="Medical Intake Agent CLI",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+
+    subparsers = parser.add_subparsers(dest="command", help="Command to run")
+
+    # Watch command
+    parser_watch = subparsers.add_parser(
+        "watch",
+        help="Watch directory for new intake forms",
+    )
+    _add_common_args(parser_watch)
+    parser_watch.set_defaults(func=cmd_watch)
+
+    # Process command
+    parser_process = subparsers.add_parser(
+        "process",
+        help="Process a single intake form",
+    )
+    _add_common_args(parser_process)
+    parser_process.add_argument("file", help="Path to intake form file")
+    parser_process.set_defaults(func=cmd_process)
+
+    # Query command
+    parser_query = subparsers.add_parser(
+        "query",
+        help="Query patient database",
+    )
+    _add_common_args(parser_query)
+    parser_query.add_argument("question", help="Question to ask")
+    parser_query.set_defaults(func=cmd_query)
+
+    # Stats command
+    parser_stats = subparsers.add_parser(
+        "stats",
+        help="Show processing statistics",
+    )
+    _add_common_args(parser_stats)
+    parser_stats.set_defaults(func=cmd_stats)
+
+    # Reset command
+    parser_reset = subparsers.add_parser(
+        "reset",
+        help="Clear all patient data from the database",
+    )
+    _add_common_args(parser_reset)
+    parser_reset.add_argument(
+        "--force",
+        "-f",
+        action="store_true",
+        help="Skip confirmation prompt",
+    )
+    parser_reset.set_defaults(func=cmd_reset)
+
+    # Init command
+    parser_init = subparsers.add_parser(
+        "init",
+        help="Download and setup required VLM models",
+    )
+    parser_init.add_argument(
+        "--vlm-model",
+        default="Qwen3-VL-4B-Instruct-GGUF",
+        help="VLM model to download (default: Qwen3-VL-4B-Instruct-GGUF)",
+    )
+    parser_init.add_argument(
+        "--debug",
+        action="store_true",
+        help="Enable debug logging",
+    )
+    parser_init.set_defaults(func=cmd_init)
+
+    # Test command
+    parser_test = subparsers.add_parser(
+        "test",
+        help="Test VLM extraction pipeline on a single file",
+    )
+    parser_test.add_argument(
+        "file",
+        help="Path to intake form image (PNG, JPG, PDF)",
+    )
+    parser_test.add_argument(
+        "--vlm-model",
+        default="Qwen3-VL-4B-Instruct-GGUF",
+        help="VLM model to use (default: Qwen3-VL-4B-Instruct-GGUF)",
+    )
+    parser_test.add_argument(
+        "--max-dimension",
+        type=int,
+        default=1024,
+        help="Max image dimension in pixels (default: 1024)",
+    )
+    parser_test.add_argument(
+        "--jpeg-quality",
+        type=int,
+        default=85,
+        help="JPEG compression quality (default: 85)",
+    )
+    parser_test.add_argument(
+        "--clear-context",
+        action="store_true",
+        help="Clear VLM context before processing (fixes memory errors)",
+    )
+    parser_test.add_argument(
+        "--debug",
+        action="store_true",
+        help="Enable debug logging",
+    )
+    parser_test.set_defaults(func=cmd_test)
+
+    # Dashboard command
+    parser_dashboard = subparsers.add_parser(
+        "dashboard",
+        help="Start web dashboard",
+    )
+    _add_common_args(parser_dashboard)
+    parser_dashboard.add_argument(
+        "--host",
+        default="127.0.0.1",
+        help="Dashboard host (default: 127.0.0.1)",
+    )
+    parser_dashboard.add_argument(
+        "--port",
+        type=int,
+        default=8080,
+        help="Dashboard port (default: 8080)",
+    )
+    parser_dashboard.add_argument(
+        "--no-open",
+        action="store_true",
+        help="Don't automatically open dashboard",
+    )
+    parser_dashboard.add_argument(
+        "--browser",
+        action="store_true",
+        help="Open in web browser instead of Electron app",
+    )
+    parser_dashboard.set_defaults(func=cmd_dashboard)
+
+    args = parser.parse_args()
+
+    # Run command
+    if not args.command:
+        parser.print_help()
+        return 0
+
+    # Configure logging - WARNING by default, DEBUG with --debug flag
+    if getattr(args, "debug", False):
+        logging.basicConfig(
+            level=logging.DEBUG,
+            format="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
+            datefmt="%H:%M:%S",
+        )
+    else:
+        # Suppress all logs from gaia modules for cleaner output
+        logging.basicConfig(level=logging.WARNING)
+        for logger_name in [
+            "gaia",
+            "gaia.llm",
+            "gaia.database",
+            "gaia.agents",
+            "gaia.utils",
+        ]:
+            logging.getLogger(logger_name).setLevel(logging.WARNING)
+
+    return args.func(args)
+
+
+if __name__ == "__main__":
+    sys.exit(main())