Latex Posters

scripts/generate_schematic_ai.py

#!/usr/bin/env python3
"""
AI-powered scientific schematic generation using Nano Banana 2.

This script uses a smart iterative refinement approach:
1. Generate initial image with Nano Banana 2
2. AI quality review using Gemini 3.7 Flash for scientific critique
3. Only regenerate if quality is below threshold for document type
4. Repeat until quality meets standards (max iterations)

Requirements:
    - OPENROUTER_API_KEY environment variable
    - requests library

Usage:
    python generate_schematic_ai.py "Create a flowchart showing CONSORT participant flow" -o flowchart.png
    python generate_schematic_ai.py "Neural network architecture diagram" -o architecture.png --iterations 2
    python generate_schematic_ai.py "Simple block diagram" -o diagram.png --doc-type poster
"""

import argparse
import base64
import json
import os
import re
import sys
import time
from pathlib import Path
from typing import Optional, Dict, Any, List, NamedTuple

try:
    import requests
except ImportError:
    print("Error: requests library not found. Install with: uv pip install requests")
    sys.exit(1)


class ReviewResult(NamedTuple):
    """The outcome of one quality review.

    A named tuple rather than a bare tuple because the review has several
    failure modes -- the model returns no choices, the request raises, the
    answer arrives in an unexpected shape -- and each of them used to be a
    chance to return the wrong number of values to the caller.

    `score` is None whenever no usable number was parsed, and `reviewed` is
    False whenever the review never produced an answer at all. Neither case
    invents a score: a diagram whose quality was not measured is reported as
    unmeasured, not as passing.
    """

    critique: str
    score: Optional[float]
    needs_improvement: bool
    reviewed: bool
    error: Optional[str] = None


# Markdown decoration the reviewer routinely wraps its answer in. Gemini writes
# "**SCORE:** 8.5" far more often than the bare "SCORE: 8.5" the rubric asks
# for, and a pattern that does not allow for it silently scores every diagram
# at whatever the default happens to be.
_MARKUP = re.compile(r"[*_`#]")

_SCORE_PATTERN = re.compile(r"\bSCORE\s*:?\s*(\d+(?:\.\d+)?)", re.IGNORECASE)
_LOOSE_SCORE_PATTERN = re.compile(
    r"(?:score|rating|quality)\s*[:\s]\s*(\d+(?:\.\d+)?)\s*(?:/\s*10)?", re.IGNORECASE
)
# The underscore is optional because normalization strips it out of
# NEEDS_IMPROVEMENT along with the surrounding markdown.
_VERDICT_PATTERN = re.compile(
    r"^\s*VERDICT\s*:?\s*(ACCEPTABLE|NEEDS[_ ]?IMPROVEMENT)", re.IGNORECASE | re.MULTILINE
)


def _normalize_review_text(text: str) -> str:
    """Strip markdown decoration so the rubric's fields can be found."""
    return _MARKUP.sub("", text or "")


def _parse_score(text: str) -> Optional[float]:
    """Return the reviewer's 0-10 score, or None when there isn't a usable one.

    None rather than a default: a fabricated score is indistinguishable from a
    real one once it reaches the review log, and a default near the middle of
    the range quietly passes lenient document types while failing strict ones.
    """
    normalized = _normalize_review_text(text)
    match = _SCORE_PATTERN.search(normalized) or _LOOSE_SCORE_PATTERN.search(normalized)
    if not match:
        return None
    score = float(match.group(1))
    # A number outside the rubric's range means something other than the score
    # was matched -- a figure count, a percentage, an iteration number.
    return score if 0.0 <= score <= 10.0 else None


def _parse_verdict(text: str) -> Optional[str]:
    """Return ACCEPTABLE / NEEDS_IMPROVEMENT from the verdict line, or None.

    Anchored to the start of a line. The rubric prompt spells out both tokens
    when it explains the response format, so a reviewer that echoes those
    instructions back would fail a perfectly good diagram under a plain
    substring search.
    """
    match = _VERDICT_PATTERN.search(_normalize_review_text(text))
    if not match:
        return None
    verdict = match.group(1).upper().replace(" ", "").replace("_", "")
    return "NEEDS_IMPROVEMENT" if verdict == "NEEDSIMPROVEMENT" else "ACCEPTABLE"


def _resolve_api_key(explicit: Optional[str] = None) -> Optional[str]:
    """Resolve the OpenRouter key from --api-key, the environment, then any .env file.

    The .env scan walks up from the working directory and finally checks the
    script's own directory, so running from anywhere inside a project picks up
    the key at its root. Only the standard library is used: python-dotenv is a
    common omission, and a missing optional dependency should not read as a
    missing credential.
    """
    if explicit:
        return explicit

    from_env = os.environ.get("OPENROUTER_API_KEY", "").strip()
    if from_env:
        return from_env

    cwd = Path.cwd()
    for directory in [cwd, *cwd.parents, Path(__file__).resolve().parent]:
        env_file = directory / ".env"
        if not env_file.is_file():
            continue
        try:
            content = env_file.read_text(encoding="utf-8", errors="replace")
        except OSError:
            continue
        for raw in content.splitlines():
            line = raw.strip()
            if line.startswith("#") or "=" not in line:
                continue
            name, _, value = line.partition("=")
            if name.strip() == "OPENROUTER_API_KEY":
                value = value.strip().strip('"').strip("'")
                if value:
                    return value

    return None


class ScientificSchematicGenerator:
    """Generate scientific schematics using AI with smart iterative refinement.
    
    Uses Gemini 3.7 Flash for quality review to determine if regeneration is needed.
    Multiple passes only occur if the generated schematic doesn't meet the
    quality threshold for the target document type.
    """
    
    # Quality thresholds by document type (score out of 10)
    # Higher thresholds for more formal publications
    QUALITY_THRESHOLDS = {
        "journal": 8.5,      # Nature, Science, etc. - highest standards
        "conference": 8.0,   # Conference papers - high standards
        "poster": 7.0,       # Academic posters - good quality
        "presentation": 6.5, # Slides/talks - clear but less formal
        "report": 7.5,       # Technical reports - professional
        "grant": 8.0,        # Grant proposals - must be compelling
        "thesis": 8.0,       # Dissertations - formal academic
        "preprint": 7.5,     # arXiv, etc. - good quality
        "default": 7.5,      # Default threshold
    }
    
    # Scientific diagram best practices prompt template
    SCIENTIFIC_DIAGRAM_GUIDELINES = """
Create a high-quality scientific diagram with these requirements:

VISUAL QUALITY:
- Clean white or light background (no textures or gradients)
- High contrast for readability and printing
- Professional, publication-ready appearance
- Sharp, clear lines and text
- Adequate spacing between elements to prevent crowding

TYPOGRAPHY:
- Clear, readable sans-serif fonts (Arial, Helvetica style)
- Minimum 10pt font size for all labels
- Consistent font sizes throughout
- All text horizontal or clearly readable
- No overlapping text

SCIENTIFIC STANDARDS:
- Accurate representation of concepts
- Clear labels for all components
- Include scale bars, legends, or axes where appropriate
- Use standard scientific notation and symbols
- Include units where applicable

ACCESSIBILITY:
- Colorblind-friendly color palette (use Okabe-Ito colors if using color)
- High contrast between elements
- Redundant encoding (shapes + colors, not just colors)
- Works well in grayscale

LAYOUT:
- Logical flow (left-to-right or top-to-bottom)
- Clear visual hierarchy
- Balanced composition
- Appropriate use of whitespace
- No clutter or unnecessary decorative elements

IMPORTANT - NO FIGURE NUMBERS:
- Do NOT include "Figure 1:", "Fig. 1", or any figure numbering in the image
- Do NOT add captions or titles like "Figure: ..." at the top or bottom
- Figure numbers and captions are added separately in the document/LaTeX
- The diagram should contain only the visual content itself
"""
    
    def __init__(self, api_key: Optional[str] = None, verbose: bool = False):
        """
        Initialize the generator.
        
        Args:
            api_key: OpenRouter API key (or use OPENROUTER_API_KEY env var)
            verbose: Print detailed progress information
        """
        # Priority: 1) explicit api_key param, 2) environment variable, 3) .env file
        self.api_key = _resolve_api_key(api_key)

        if not self.api_key:
            raise ValueError(
                "OPENROUTER_API_KEY not found. Please either:\n"
                "  1. Set the OPENROUTER_API_KEY environment variable\n"
                "  2. Add OPENROUTER_API_KEY to your .env file\n"
                "  3. Pass api_key parameter to the constructor\n"
                "Get your API key from: https://openrouter.ai/keys"
            )
        
        self.verbose = verbose
        self._last_error = None  # Track last error for better reporting
        self.base_url = "https://openrouter.ai/api/v1"
        # Nano Banana 2 - Google's advanced image generation model. The slug must
        # be an image-output model; a text-only chat model is rejected with
        # "No endpoints found that support the requested output modalities".
        # https://openrouter.ai/google/gemini-3.1-flash-image
        self.image_model = "google/gemini-3.1-flash-image"
        # Gemini 3.7 Flash for quality review - excellent vision and reasoning
        self.review_model = "google/gemini-3.7-flash"
        
    def _log(self, message: str):
        """Log message if verbose mode is enabled."""
        if self.verbose:
            print(f"[{time.strftime('%H:%M:%S')}] {message}")
    
    def _post_request(self, endpoint: str, payload: Dict[str, Any]) -> Dict[str, Any]:
        """Send one non-streaming request; never automatically replay a paid call."""
        headers = {
            "Authorization": f"Bearer {self.api_key}",
            "Content-Type": "application/json",
            "X-Title": "Scientific Schematic Generator",
        }
        self._log(f"POST {endpoint} using {payload['model']}")
        try:
            response = requests.post(
                f"{self.base_url}/{endpoint}", headers=headers,
                json=payload, timeout=120,
            )
            try:
                response_json = response.json()
            except ValueError:
                raise RuntimeError(
                    f"API returned non-JSON data (HTTP {response.status_code})"
                ) from None
            if not isinstance(response_json, dict):
                raise RuntimeError("API response must be a JSON object")
            if response.status_code != 200 or "error" in response_json:
                error = response_json.get("error", {})
                detail = error.get("message", "request rejected") if isinstance(error, dict) else str(error)
                # Providers occasionally echo input; do not dump whole bodies or credentials.
                detail = str(detail).replace(self.api_key, "[redacted]")[:500]
                raise RuntimeError(f"API request failed (HTTP {response.status_code}): {detail}")
            return response_json
        except requests.exceptions.Timeout:
            raise RuntimeError("API request timed out after 120 seconds") from None
        except requests.exceptions.RequestException as exc:
            detail = str(exc).replace(self.api_key, "[redacted]")
            raise RuntimeError(f"API request failed: {detail}") from None

    def _make_request(self, model: str, messages: List[Dict[str, Any]],
                      modalities: Optional[List[str]] = None) -> Dict[str, Any]:
        """Use chat completions for the text-and-image vision review."""
        payload = {"model": model, "messages": messages}
        if modalities:
            payload["modalities"] = modalities
        return self._post_request("chat/completions", payload)

    def _extract_image_from_response(self, response: Dict[str, Any]) -> bytes:
        """Decode the dedicated Image API response and enforce this helper's PNG contract.

        The provider may omit media_type; the PNG signature still has to match.
        This is a format check, not a full image-decoder or scientific-content check.
        """
        data = response.get("data")
        if not isinstance(data, list) or not data or not isinstance(data[0], dict):
            raise RuntimeError("Image API returned no data[0] image")
        item = data[0]
        if item.get("media_type") not in (None, "image/png"):
            raise RuntimeError(f"Expected PNG output; received {item['media_type']}")
        encoded = item.get("b64_json")
        if not isinstance(encoded, str) or not encoded:
            raise RuntimeError("Image API returned no b64_json image")
        try:
            image_data = base64.b64decode(encoded, validate=True)
        except ValueError:
            raise RuntimeError("Image API returned invalid base64 image data") from None
        if not image_data.startswith(b"\x89PNG\r\n\x1a\n"):
            raise RuntimeError("Image API output does not have a PNG signature")
        return image_data

    def _image_to_base64(self, image_path: str) -> str:
        """
        Convert image file to base64 data URL.
        
        Args:
            image_path: Path to image file
            
        Returns:
            Base64 data URL string
        """
        with open(image_path, "rb") as f:
            image_data = f.read()
        
        # Determine image type from extension
        ext = Path(image_path).suffix.lower()
        mime_type = {
            ".png": "image/png",
            ".jpg": "image/jpeg",
            ".jpeg": "image/jpeg",
            ".gif": "image/gif",
            ".webp": "image/webp"
        }.get(ext, "image/png")
        
        base64_data = base64.b64encode(image_data).decode("utf-8")
        return f"data:{mime_type};base64,{base64_data}"
    
    def generate_image(self, prompt: str) -> Optional[bytes]:
        """
        Generate an image using Nano Banana 2.
        
        Args:
            prompt: Description of the diagram to generate
            
        Returns:
            Image bytes or None if generation failed
        """
        self._last_error = None  # Reset error
        
        try:
            # Image generation has its own body and response, distinct from chat.
            # n=1 is supported by both current Gemini image-provider endpoints.
            # output_format is absent from their capability maps: use the default
            # and validate the returned MIME/signature rather than send that knob.
            response = self._post_request("images", {
                "model": self.image_model, "prompt": prompt, "n": 1,
            })
            image_data = self._extract_image_from_response(response)
            self._log(f"[OK] Generated PNG ({len(image_data)} bytes)")
            return image_data
        except RuntimeError as exc:
            self._last_error = str(exc)
            self._log(f"[FAIL] Generation failed: {self._last_error}")
            return None

    def review_image(self, image_path: str, original_prompt: str, 
                    iteration: int, doc_type: str = "default",
                    max_iterations: int = 2) -> ReviewResult:
        """
        Review the generated image for quality, using the vision review model.

        Args:
            image_path: Path to the generated image
            original_prompt: Original user prompt
            iteration: Current iteration number
            doc_type: Document type (journal, poster, presentation, etc.)
            max_iterations: Maximum iterations allowed

        Returns:
            A ReviewResult. When the review cannot be completed -- the request
            fails, or the model answers with nothing usable -- `reviewed` is
            False and `score` is None rather than a stand-in number, and
            `needs_improvement` is False: the fault is in the reviewer, not in
            the diagram, so regenerating would not help.
        """
        image_data_url = self._image_to_base64(image_path)
        
        # Get quality threshold for this document type
        threshold = self.QUALITY_THRESHOLDS.get(doc_type.lower(), 
                                                 self.QUALITY_THRESHOLDS["default"])
        
        review_prompt = f"""You are an expert reviewer evaluating a scientific diagram for publication quality.

ORIGINAL REQUEST: {original_prompt}

DOCUMENT TYPE: {doc_type} (quality threshold: {threshold}/10)
ITERATION: {iteration}/{max_iterations}

Carefully evaluate this diagram on these criteria:

1. **Scientific Accuracy** (0-2 points)
   - Correct representation of concepts
   - Proper notation and symbols
   - Accurate relationships shown

2. **Clarity and Readability** (0-2 points)
   - Easy to understand at a glance
   - Clear visual hierarchy
   - No ambiguous elements

3. **Label Quality** (0-2 points)
   - All important elements labeled
   - Labels are readable (appropriate font size)
   - Consistent labeling style

4. **Layout and Composition** (0-2 points)
   - Logical flow (top-to-bottom or left-to-right)
   - Balanced use of space
   - No overlapping elements

5. **Professional Appearance** (0-2 points)
   - Publication-ready quality
   - Clean, crisp lines and shapes
   - Appropriate colors/contrast

RESPOND IN THIS EXACT FORMAT:
SCORE: [total score 0-10]

STRENGTHS:
- [strength 1]
- [strength 2]

ISSUES:
- [issue 1 if any]
- [issue 2 if any]

VERDICT: [ACCEPTABLE or NEEDS_IMPROVEMENT]

If score >= {threshold}, the diagram is ACCEPTABLE for {doc_type} publication.
If score < {threshold}, mark as NEEDS_IMPROVEMENT with specific suggestions."""

        messages = [
            {
                "role": "user",
                "content": [
                    {
                        "type": "text",
                        "text": review_prompt
                    },
                    {
                        "type": "image_url",
                        "image_url": {
                            "url": image_data_url
                        }
                    }
                ]
            }
        ]
        
        try:
            response = self._make_request(
                model=self.review_model,
                messages=messages
            )

            # Extract text response
            choices = response.get("choices", [])
            if not choices:
                # A normal enough outcome -- a rate limit, a content filter, a
                # provider hiccup. The diagram itself is fine; only its review
                # is missing, and saying so beats inventing a score.
                reason = "the review model returned no choices"
                self._log(f"[WARN] Review unavailable: {reason}")
                return ReviewResult(
                    f"Review unavailable: {reason}.",
                    None, False, reviewed=False, error=reason,
                )

            message = choices[0].get("message", {})
            content = message.get("content", "")

            # Some models put their analysis in a separate reasoning field
            reasoning = message.get("reasoning", "")
            if reasoning and not content:
                content = reasoning

            if isinstance(content, list):
                # Extract text from content blocks
                text_parts = []
                for block in content:
                    if isinstance(block, dict) and block.get("type") == "text":
                        text_parts.append(block.get("text", ""))
                content = "\n".join(text_parts)

            score = _parse_score(content)
            verdict = _parse_verdict(content)

            if score is None and verdict is None:
                reason = "no score or verdict found in the review"
                self._log(f"[WARN] Review unusable: {reason}")
                return ReviewResult(
                    content if content else f"Review unusable: {reason}.",
                    None, False, reviewed=True, error=reason,
                )

            needs_improvement = verdict == "NEEDS_IMPROVEMENT" or (
                score is not None and score < threshold
            )

            shown = f"{score}/10" if score is not None else "not stated"
            self._log(f"[OK] Review complete (Score: {shown}, Threshold: {threshold}/10)")
            self._log(f"  Verdict: {'Needs improvement' if needs_improvement else 'Acceptable'}")

            return ReviewResult(
                content if content else "Review returned no text",
                score,
                needs_improvement,
                reviewed=True,
                error=None if score is not None else "no score found in the review",
            )
        except Exception as e:
            # A failed review must not fail the run -- the image is already
            # generated and saved -- but it must not read as a pass either.
            self._log(f"[WARN] Review failed: {str(e)}")
            return ReviewResult(
                f"Review failed: {str(e)}",
                None, False, reviewed=False, error=str(e),
            )
    
    def improve_prompt(self, original_prompt: str, critique: str, 
                      iteration: int) -> str:
        """
        Improve the generation prompt based on critique.
        
        Args:
            original_prompt: Original user prompt
            critique: Review critique from previous iteration
            iteration: Current iteration number
            
        Returns:
            Improved prompt for next generation
        """
        improved_prompt = f"""{self.SCIENTIFIC_DIAGRAM_GUIDELINES}

USER REQUEST: {original_prompt}

ITERATION {iteration}: Based on previous feedback, address these specific improvements:
{critique}

Generate an improved version that addresses all the critique points while maintaining scientific accuracy and professional quality."""
        
        return improved_prompt
    
    def generate_iterative(self, user_prompt: str, output_path: str,
                          iterations: int = 2, 
                          doc_type: str = "default") -> Dict[str, Any]:
        """
        Generate scientific schematic with smart iterative refinement.
        
        Only regenerates if the quality score is below the threshold for the
        specified document type. This saves API calls and time when the first
        generation is already good enough.
        
        Args:
            user_prompt: User's description of desired diagram
            output_path: Path to save final image
            iterations: Maximum refinement iterations (default: 2, max: 2)
            doc_type: Document type for quality threshold (journal, poster, etc.)
            
        Returns:
            Dictionary with generation results and metadata
        """
        if isinstance(iterations, bool) or not isinstance(iterations, int) or not 1 <= iterations <= 2:
            raise ValueError("iterations must be 1 or 2")
        if doc_type.lower() not in self.QUALITY_THRESHOLDS:
            raise ValueError(f"Unknown document type: {doc_type}")
        output_path = Path(output_path)
        if output_path.suffix.lower() != ".png":
            raise ValueError("Output path must end in .png; format conversion is not supported")
        output_dir = output_path.parent
        output_dir.mkdir(parents=True, exist_ok=True)
        
        base_name = output_path.stem
        extension = output_path.suffix or ".png"
        
        # Get quality threshold for this document type
        threshold = self.QUALITY_THRESHOLDS.get(doc_type.lower(), 
                                                 self.QUALITY_THRESHOLDS["default"])
        
        results = {
            "user_prompt": user_prompt,
            "doc_type": doc_type,
            "quality_threshold": threshold,
            "image_model": self.image_model,
            "review_model": self.review_model,
            "quality_met": False,
            "termination_reason": None,
            "iterations": [],
            "final_image": None,
            # None, not 0.0 -- a run whose review never completed has no score,
            # and a number here would be read as one the reviewer gave.
            "final_score": None,
            "final_reviewed": False,
            "success": False,
            "early_stop": False,
            "early_stop_reason": None
        }
        
        current_prompt = f"""{self.SCIENTIFIC_DIAGRAM_GUIDELINES}

USER REQUEST: {user_prompt}

Generate a publication-quality scientific diagram that meets all the guidelines above."""
        
        print(f"\n{'='*60}")
        print(f"Generating Scientific Schematic")
        print(f"{'='*60}")
        print(f"Description: {user_prompt}")
        print(f"Document Type: {doc_type}")
        print(f"Quality Threshold: {threshold}/10")
        print(f"Max Iterations: {iterations}")
        print(f"Output: {output_path}")
        print(f"{'='*60}\n")
        
        for i in range(1, iterations + 1):
            print(f"\n[Iteration {i}/{iterations}]")
            print("-" * 40)
            
            # Generate image
            print(f"Generating image...")
            image_data = self.generate_image(current_prompt)
            
            if not image_data:
                error_msg = getattr(self, '_last_error', 'Image generation failed - no image data returned')
                print(f"[FAIL] Generation failed: {error_msg}")
                results["iterations"].append({
                    "iteration": i,
                    "success": False,
                    "error": error_msg
                })
                # A failed refinement must not discard the earlier saved image.
                # Stop here rather than replay a request after an ambiguous timeout.
                results["termination_reason"] = "generation_failed"
                break

            # Save iteration image
            iter_path = output_dir / f"{base_name}_v{i}{extension}"
            with open(iter_path, "wb") as f:
                f.write(image_data)
            print(f"[OK] Saved: {iter_path}")
            
            # Review the image with the vision review model
            print(f"Reviewing image with {self.review_model}...")
            review = self.review_image(
                str(iter_path), user_prompt, i, doc_type, iterations
            )
            if review.score is not None:
                print(f"[OK] Score: {review.score}/10 (threshold: {threshold}/10)")
            else:
                print(f"[WARN] Review unavailable - image kept, quality not verified")
                print(f"  Reason: {review.error}")

            # Save iteration results
            iteration_result = {
                "iteration": i,
                "image_path": str(iter_path),
                "prompt": current_prompt,
                "critique": review.critique,
                "score": review.score,
                "reviewed": review.reviewed,
                "review_error": review.error,
                "needs_improvement": review.needs_improvement,
                "success": True
            }
            results["iterations"].append(iteration_result)
            # Preserve the latest successful image even if a later API call fails.
            results["final_image"] = str(iter_path)
            results["final_score"] = review.score
            results["final_reviewed"] = review.reviewed and review.score is not None
            results["success"] = True
            results["quality_met"] = (
                results["final_reviewed"] and review.score >= threshold
                and not review.needs_improvement
            )

            # Check if quality is acceptable - STOP EARLY if so
            if not review.needs_improvement:
                if review.score is not None:
                    print(f"\n[OK] Quality meets {doc_type} threshold ({review.score} >= {threshold})")
                    print(f"  No further iterations needed!")
                    reason = (f"Quality score {review.score} meets threshold "
                              f"{threshold} for {doc_type}")
                else:
                    # Regenerating cannot fix a reviewer that did not answer.
                    print(f"\n[WARN] Stopping without a verified score - review the image yourself")
                    reason = f"Review did not produce a score: {review.error}"
                results["final_image"] = str(iter_path)
                results["final_score"] = review.score
                results["final_reviewed"] = review.reviewed and review.score is not None
                results["success"] = True
                results["early_stop"] = True
                results["early_stop_reason"] = reason
                results["termination_reason"] = (
                    "quality_threshold_met" if results["quality_met"] else "review_unavailable"
                )
                break

            # If this is the last iteration, we're done regardless
            if i == iterations:
                print(f"\n[WARN] Maximum iterations reached")
                results["termination_reason"] = "max_iterations"
                results["final_image"] = str(iter_path)
                results["final_score"] = review.score
                results["final_reviewed"] = review.reviewed and review.score is not None
                results["success"] = True
                break

            # Quality below threshold - improve prompt for next iteration
            print(f"\n[WARN] Review requests improvement (score: {review.score}; threshold: {threshold})")
            print(f"Improving prompt based on feedback...")
            current_prompt = self.improve_prompt(user_prompt, review.critique, i + 1)
        
        # Copy final version to output path
        if results["success"] and results["final_image"]:
            final_iter_path = Path(results["final_image"])
            if final_iter_path != output_path:
                import shutil
                shutil.copy(final_iter_path, output_path)
                print(f"\n[OK] Final image: {output_path}")
        
        # Save review log
        log_path = output_dir / f"{base_name}_review_log.json"
        with open(log_path, "w", encoding="utf-8") as f:
            json.dump(results, f, indent=2)
        print(f"[OK] Review log: {log_path}")
        
        print(f"\n{'='*60}")
        print(f"Generation Complete!")
        if results["final_score"] is not None:
            print(f"Final Score: {results['final_score']}/10")
        else:
            print(f"Final Score: unavailable (the review did not produce one)")
        if results["early_stop"]:
            print(f"Iterations Used: {len([r for r in results['iterations'] if r.get('success')])}/{iterations} (early stop)")
        print(f"{'='*60}\n")
        
        return results


def main():
    """Command-line interface."""
    parser = argparse.ArgumentParser(
        description="Generate scientific schematics using AI with smart iterative refinement",
        formatter_class=argparse.RawDescriptionHelpFormatter,
        epilog="""
Examples:
  # Generate a flowchart for a journal paper
  python generate_schematic_ai.py "CONSORT participant flow diagram" -o flowchart.png --doc-type journal
  
  # Generate neural network architecture for presentation (lower threshold)
  python generate_schematic_ai.py "Transformer encoder-decoder architecture" -o transformer.png --doc-type presentation
  
  # Generate with custom max iterations for poster
  python generate_schematic_ai.py "Biological signaling pathway" -o pathway.png --iterations 2 --doc-type poster
  
  # Verbose output
  python generate_schematic_ai.py "Circuit diagram" -o circuit.png -v

Document Types (quality thresholds):
  journal      8.5/10  - Nature, Science, peer-reviewed journals
  conference   8.0/10  - Conference papers
  thesis       8.0/10  - Dissertations, theses
  grant        8.0/10  - Grant proposals
  preprint     7.5/10  - arXiv, bioRxiv, etc.
  report       7.5/10  - Technical reports
  poster       7.0/10  - Academic posters
  presentation 6.5/10  - Slides, talks
  default      7.5/10  - General purpose

Note: Multiple iterations only occur if quality is BELOW the threshold.
      If the first generation meets the threshold, no extra API calls are made.

Environment:
  OPENROUTER_API_KEY    OpenRouter API key (required)
        """
    )
    
    parser.add_argument("prompt", help="Description of the diagram to generate")
    parser.add_argument("-o", "--output", required=True, 
                       help="Output image path (e.g., diagram.png)")
    parser.add_argument("--iterations", type=int, default=2,
                       help="Maximum refinement iterations (default: 2, max: 2)")
    parser.add_argument("--doc-type", default="default",
                       choices=["journal", "conference", "poster", "presentation", 
                               "report", "grant", "thesis", "preprint", "default"],
                       help="Document type for quality threshold (default: default)")
    parser.add_argument("--api-key", help="OpenRouter API key (or set OPENROUTER_API_KEY)")
    parser.add_argument("-v", "--verbose", action="store_true",
                       help="Verbose output")
    
    args = parser.parse_args()
    
    # Check for API key — resolves --api-key, the environment, then any .env file
    api_key = _resolve_api_key(args.api_key)
    if not api_key:
        print("Error: OPENROUTER_API_KEY not found")
        print("\nSet it with:")
        print("  export OPENROUTER_API_KEY='your_api_key'")
        print("\nOr add OPENROUTER_API_KEY=your_api_key to a .env file")
        print("Or provide via --api-key flag")
        sys.exit(1)
    
    # Validate iterations - enforce max of 2
    if args.iterations < 1 or args.iterations > 2:
        print("Error: Iterations must be between 1 and 2")
        sys.exit(1)
    
    try:
        generator = ScientificSchematicGenerator(api_key=api_key, verbose=args.verbose)
        results = generator.generate_iterative(
            user_prompt=args.prompt,
            output_path=args.output,
            iterations=args.iterations,
            doc_type=args.doc_type
        )
        
        if results["success"]:
            print(f"\n[OK] Success! Image saved to: {args.output}")
            used = len([r for r in results['iterations'] if r.get('success')])
            if results.get("final_reviewed"):
                if results.get("quality_met"):
                    print(f"  (Completed in {used} iteration(s) - quality threshold met)")
                else:
                    print("  (Image saved; review did not meet the quality criteria.)")
            else:
                # The image is real; the quality claim is not. Say which.
                print(f"  (Completed in {used} iteration(s) - quality NOT verified,"
                      f" the review produced no score. Check the image yourself.)")
            sys.exit(0)
        else:
            print(f"\n[FAIL] Generation failed. Check review log for details.")
            sys.exit(1)
    except Exception as e:
        print(f"\n[FAIL] Error: {str(e)}")
        sys.exit(1)


if __name__ == "__main__":
    main()