app.py

#!/usr/bin/env python3
"""
Hugging Face Gradio App for RDF Validation with MCP Server and Anthropic AI

This app serves both as a web interface and can expose MCP server functionality.
Deploy this on Hugging Face Spaces with your Anthropic API key.
"""

import gradio as gr
import os
import json
import sys
import asyncio
import logging
import re
import hashlib
import threading
import time
from collections import OrderedDict
from typing import Any, Dict, List, Optional

# Add current directory to path
sys.path.append(os.path.dirname(os.path.abspath(__file__)))

# Import our validation logic
try:
    from validator import validate_rdf
    VALIDATOR_AVAILABLE = True
    # Test that the function is callable
    if not callable(validate_rdf):
        print("⚠️ Warning: validate_rdf is not callable")
        VALIDATOR_AVAILABLE = False
    else:
        print("✅ Validator module loaded successfully")
except ImportError as e:
    VALIDATOR_AVAILABLE = False
    print(f"⚠️ Warning: validator.py not found or has import errors: {e}")
    print("Some features may be limited.")
except Exception as e:
    VALIDATOR_AVAILABLE = False
    print(f"⚠️ Warning: Error loading validator: {e}")

# Optional: Check if OpenAI and requests are available
try:
    from openai import OpenAI
    OPENAI_AVAILABLE = True
except ImportError:
    OPENAI_AVAILABLE = False
    print("💡 Install 'openai' package for AI-powered corrections: pip install openai")

try:
    import requests
    HF_INFERENCE_AVAILABLE = True
except ImportError:
    HF_INFERENCE_AVAILABLE = False
    print("💡 Install 'requests' package for AI-powered corrections: pip install requests")

# Set up logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# Configuration - Your specific Hugging Face Inference Endpoint (hardcoded)
HF_API_KEY = os.getenv('HF_API_KEY', '')  # Hugging Face API key from Secret
HF_ENDPOINT_URL = "https://evxgv66ksxjlfrts.us-east-1.aws.endpoints.huggingface.cloud/v1/"
HF_MODEL = "lmstudio-community/Llama-3.3-70B-Instruct-GGUF"  # Correct model name for your endpoint

# AI Correction Configuration
MAX_CORRECTION_ATTEMPTS = 2  # Reduced for speed (rapid fix handles most cases)
ENABLE_VALIDATION_LOOP = True  # Enable validation loop by default

# MCP4BibFrame Documentation API Configuration
MCP4BIBFRAME_DOCS_URL = "https://jimfhahn-mcp4bibframe-docs.hf.space/api/mcp"
MCP4BIBFRAME_DOCS_ENABLED = True  # Set to False to disable doc integration

# Cache BibFrame documentation responses to avoid repeated network calls
BIBFRAME_DOCS_CACHE: Dict[str, tuple[Any, float]] = {}
BIBFRAME_DOCS_CACHE_TTL = 3600  # seconds

# Cache successful correction outputs to accelerate repeated error patterns
FIX_CACHE: OrderedDict[str, str] = OrderedDict()
FIX_CACHE_MAX_SIZE = 100


def _make_fix_cache_key(validation_results: str, rdf_content: str, template: str) -> str:
    """Generate a deterministic cache key for correction attempts."""
    hasher = hashlib.sha256()
    hasher.update(template.strip().encode("utf-8"))
    hasher.update(b"\x1f")
    hasher.update(validation_results.strip().encode("utf-8", errors="ignore"))
    hasher.update(b"\x1f")
    hasher.update(rdf_content.strip().encode("utf-8", errors="ignore"))
    return hasher.hexdigest()


def _get_cached_correction(cache_key: str, steps_log: Optional[List[str]] = None) -> Optional[str]:
    """Retrieve a cached correction, updating its recency ordering."""
    cached = FIX_CACHE.get(cache_key)
    if cached is not None:
        FIX_CACHE.move_to_end(cache_key)
        if steps_log is not None:
            steps_log.append("Using cached correction for repeated validation errors")
    return cached


def _store_correction_in_cache(cache_key: str, corrected_rdf: str, steps_log: Optional[List[str]] = None) -> None:
    """Store a correction in the cache and evict the oldest entry if needed."""
    if not corrected_rdf:
        return
    FIX_CACHE[cache_key] = corrected_rdf
    FIX_CACHE.move_to_end(cache_key)
    if len(FIX_CACHE) > FIX_CACHE_MAX_SIZE:
        removed_key, _ = FIX_CACHE.popitem(last=False)
        if steps_log is not None:
            steps_log.append("Cache full; evicted oldest correction entry")
    elif steps_log is not None:
        steps_log.append("Cached correction for future reuse")

# Cache successful correction outputs to accelerate repeated error patterns
FIX_CACHE: OrderedDict[str, str] = OrderedDict()
FIX_CACHE_MAX_SIZE = 100


def rapid_fix_missing_properties(rdf_content: str, validation_results: str, template: str, steps_log: Optional[List[str]] = None) -> Optional[str]:
    """Ultra-fast fix for simple missing property errors - no AI needed."""
    import re
    
    # Quick pattern match for missing properties
    missing = re.findall(r"Less than \d+ values on.*->bf:(\w+)", validation_results)
    if not missing:
        if steps_log:
            steps_log.append("❌ Rapid fix: No missing properties detected in validation results")
        return None
    
    if steps_log:
        steps_log.append(f"🔍 Rapid fix detected {len(missing)} missing properties: {', '.join(set(missing))}")
    
    # Pre-compiled property templates (no API calls)
    INSTANT_FIXES = {
        "title": '<bf:title><bf:Title><bf:mainTitle>Untitled</bf:mainTitle></bf:Title></bf:title>',
        "language": '<bf:language><bf:Language rdf:about="http://id.loc.gov/vocabulary/languages/eng"><rdfs:label>English</rdfs:label><bf:code>eng</bf:code></bf:Language></bf:language>',
        "content": '<bf:content><bf:Content rdf:about="http://id.loc.gov/vocabulary/contentTypes/txt"><rdfs:label>text</rdfs:label><bf:code>txt</bf:code></bf:Content></bf:content>',
        "adminMetadata": '''<bf:adminMetadata>
    <bf:AdminMetadata>
        <bf:status>
            <bf:Status rdf:about="http://id.loc.gov/vocabulary/mstatus/n">
                <rdfs:label>new</rdfs:label>
                <bf:code>n</bf:code>
            </bf:Status>
        </bf:status>
        <bf:date rdf:datatype="http://www.w3.org/2001/XMLSchema#date">2024-01-01</bf:date>
        <bf:agent>
            <bf:Agent rdf:about="http://id.loc.gov/vocabulary/organizations/dlc">
                <rdf:type rdf:resource="http://id.loc.gov/ontologies/bibframe/Organization"/>
                <rdfs:label>Library of Congress</rdfs:label>
            </bf:Agent>
        </bf:agent>
        <bf:assigner>
            <bf:Agent rdf:about="http://id.loc.gov/vocabulary/organizations/dlc">
                <rdf:type rdf:resource="http://id.loc.gov/ontologies/bibframe/Organization"/>
                <rdfs:label>Library of Congress</rdfs:label>
            </bf:Agent>
        </bf:assigner>
    </bf:AdminMetadata>
</bf:adminMetadata>''',
        "assigner": '''<bf:assigner>
    <bf:Agent rdf:about="http://id.loc.gov/vocabulary/organizations/dlc">
        <rdf:type rdf:resource="http://id.loc.gov/ontologies/bibframe/Organization"/>
        <rdfs:label>Library of Congress</rdfs:label>
    </bf:Agent>
</bf:assigner>'''
    }
    
    # Find insertion point
    work_match = re.search(r'(<bf:Work[^>]*>)(.*?)(</bf:Work>)', rdf_content, re.DOTALL)
    instance_match = re.search(r'(<bf:Instance[^>]*>)(.*?)(</bf:Instance>)', rdf_content, re.DOTALL)
    
    if not work_match and not instance_match:
        if steps_log:
            steps_log.append("❌ Rapid fix: No bf:Work or bf:Instance found in RDF")
        return None
    
    match = work_match or instance_match
    target_type = "Work" if work_match else "Instance"
    opening_tag = match.group(1)
    content = match.group(2)
    closing_tag = match.group(3)
    
    if steps_log:
        steps_log.append(f"📍 Rapid fix target: bf:{target_type}")
        has_admin = "<bf:adminMetadata>" in content or "<bf:AdminMetadata>" in content
        steps_log.append(f"🔍 Current state: AdminMetadata {'EXISTS' if has_admin else 'MISSING'}")
    
    # Build fixes
    fixes = []
    assigner_fixed = False
    
    for prop in missing[:10]:  # Limit to 10 properties
        prop_lower = prop.lower()
        
        # Special handling for assigner within AdminMetadata
        if prop_lower == "assigner":
            if steps_log:
                steps_log.append("🔧 Processing missing 'assigner' property...")
            # Look for existing AdminMetadata blocks that need assigner
            admin_pattern = re.compile(r'(<bf:AdminMetadata[^>]*>)(.*?)(</bf:AdminMetadata>)', re.DOTALL)
            
            def add_assigner(match):
                nonlocal assigner_fixed
                admin_open = match.group(1)
                admin_content = match.group(2)
                admin_close = match.group(3)
                
                # Skip if already has assigner
                if '<bf:assigner' in admin_content:
                    return match.group(0)
                
                # Extract agent URI if present to reuse for assigner
                agent_uri = None
                agent_match = re.search(r'<bf:agent\s+rdf:resource="([^"]+)"', admin_content)
                if not agent_match:
                    agent_match = re.search(r'<bf:agent[^>]*>\s*<[^>]+\s+rdf:about="([^"]+)"', admin_content)
                if agent_match:
                    agent_uri = agent_match.group(1)
                
                # Build assigner element
                if agent_uri:
                    assigner_element = f'        <bf:assigner rdf:resource="{agent_uri}"/>'
                else:
                    # Use default Library of Congress
                    assigner_element = '''        <bf:assigner>
            <bf:Agent rdf:about="http://id.loc.gov/vocabulary/organizations/dlc">
                <rdf:type rdf:resource="http://id.loc.gov/ontologies/bibframe/Organization"/>
                <rdfs:label>Library of Congress</rdfs:label>
            </bf:Agent>
        </bf:assigner>'''
                
                assigner_fixed = True
                if steps_log:
                    steps_log.append(f"   ✅ Injected assigner into existing AdminMetadata (agent URI: {agent_uri or 'default'})")
                # Insert before closing tag
                return admin_open + admin_content + '\n' + assigner_element + '\n    ' + admin_close
            
            original_content = content
            content = admin_pattern.sub(add_assigner, content)
            
            if assigner_fixed and steps_log:
                steps_log.append("   ✅ Assigner successfully added to existing AdminMetadata")
            elif steps_log and content == original_content:
                steps_log.append("   ℹ️  No AdminMetadata found to inject assigner (will add with full block if adminMetadata is missing)")
            
        elif prop in INSTANT_FIXES and f"<bf:{prop}" not in content:
            fixes.append(INSTANT_FIXES[prop])
            if steps_log:
                steps_log.append(f"   ✅ Will add missing '{prop}' property")
        elif prop in INSTANT_FIXES:
            if steps_log:
                steps_log.append(f"   ℹ️  Property '{prop}' already exists, skipping")
        elif steps_log:
            steps_log.append(f"   ⚠️  No template for '{prop}', skipping")
    
    if not fixes and not assigner_fixed:
        if steps_log:
            steps_log.append("❌ Rapid fix: No properties could be fixed")
        return None
    
    # Insert all at once
    if fixes:
        if steps_log:
            steps_log.append(f"🔨 Adding {len(fixes)} missing properties to {target_type}")
        fixed_content = opening_tag + content + '\n    ' + '\n    '.join(fixes) + '\n' + closing_tag
    else:
        if steps_log:
            steps_log.append(f"🔨 Modified content (assigner injection only)")
        fixed_content = opening_tag + content + closing_tag
    
    # Replace in original RDF
    result = rdf_content.replace(match.group(0), fixed_content)
    
    if steps_log:
        steps_log.append(f"✅ Rapid fix complete: Added {len(fixes)} properties, assigner_injected={assigner_fixed}")
    
    return result


def get_ai_correction_minimal(errors: str, rdf: str, max_tokens: int = 800) -> str:
    """Ultra-minimal prompt for faster AI response."""
    
    if not OPENAI_AVAILABLE or not os.getenv('HF_API_KEY'):
        return rdf
    
    try:
        client = get_openai_client()
        if not client:
            return rdf
        
        # Extract just the critical errors
        error_lines = []
        for line in errors.split('\n'):
            if any(term in line for term in ['Less than', 'missing', 'required', '->bf:', 'adminMetadata', 'assigner']):
                error_lines.append(line.strip()[:100])
                if len(error_lines) >= 5:
                    break
        
        if not error_lines:
            return rdf
        
        # Ultra-concise prompt
        prompt = f"""Fix these BibFrame errors:

{chr(10).join(error_lines[:3])}

Add only what's missing to this RDF:
{rdf[:800]}...{rdf[-200:] if len(rdf) > 1000 else ''}

Return complete valid RDF/XML only."""
        
        response = client.chat.completions.create(
            model=HF_MODEL,
            messages=[
                {"role": "system", "content": "Fix RDF. Output only valid RDF/XML. No explanations."},
                {"role": "user", "content": prompt}
            ],
            max_tokens=max_tokens,
            temperature=0,
            timeout=20  # Much shorter timeout
        )
        
        result = response.choices[0].message.content
        result = extract_rdf_from_response(result)
        result = fix_common_rdf_errors(result)
        
        return result
        
    except Exception:
        return rdf


def test_validator_functionality():
    """Test if the validator is actually working"""
    if not VALIDATOR_AVAILABLE:
        print("❌ Validator not available for testing")
        return False
    
    try:
        # Test with minimally valid RDF/XML that matches SHACL targets but is missing required properties
        # This ensures SHACL finds focus nodes (bf:Text Work) and reports violations
        test_rdf = '''<?xml version="1.0"?>
<rdf:RDF xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#"
         xmlns:bf="http://id.loc.gov/ontologies/bibframe/">
  <bf:Work rdf:about="http://example.org/work/1">
    <rdf:type rdf:resource="http://id.loc.gov/ontologies/bibframe/Text"/>
    <!-- Intentionally missing title, language, content, adminMetadata to trigger SHACL violations -->
  </bf:Work>
</rdf:RDF>'''
        conforms, results = validate_rdf(test_rdf.encode('utf-8'), 'monograph')
        
        # This should fail validation due to missing required properties
        if conforms:
            print("⚠️ WARNING: Validator returned 'conforms=True' for invalid RDF. Validator may not be working correctly!")
            return False
        else:
            preview = (results or '').strip()
            preview = preview[:200] + ('…' if len(preview) > 200 else '')
            print(f"✅ Validator test passed. Got expected SHACL violations. Preview: {preview if preview else 'No results text returned'}")
            return True
            
    except Exception as e:
        print(f"❌ Validator test failed with error: {e}")
        return False

# Run the test on startup
if VALIDATOR_AVAILABLE:
    test_validator_functionality()

def query_bibframe_docs(tool_name: str, params: dict, timeout: int = 10) -> Optional[dict]:
    """
    Query the MCP4BibFrame documentation API using the MCP protocol.
    
    Args:
        tool_name (str): Name of the tool to invoke
        params (dict): Parameters for the tool
        timeout (int): Request timeout in seconds
        
    Returns:
        Optional[dict]: Response data or None if failed
    """
    if not MCP4BIBFRAME_DOCS_ENABLED:
        return None
        
    try:
        # Construct MCP request
        mcp_request = {
            "jsonrpc": "2.0",
            "method": "tools/call",
            "params": {
                "name": tool_name,
                "arguments": params
            },
            "id": 1
        }
        
        logger.info(f"Querying BibFrame docs: {tool_name} with {params}")
        
        # Make SSE request to MCP endpoint
        response = requests.post(
            MCP4BIBFRAME_DOCS_URL,
            json=mcp_request,
            timeout=timeout,
            headers={"Accept": "text/event-stream"}
        )
        
        if response.status_code == 200:
            # Parse SSE response
            for line in response.text.split('\n'):
                if line.startswith('data: '):
                    try:
                        data = json.loads(line[6:])
                        if 'result' in data:
                            return data['result']
                    except json.JSONDecodeError:
                        continue
        else:
            logger.warning(f"BibFrame docs API returned status {response.status_code}")
            
    except requests.exceptions.Timeout:
        logger.warning("Timeout querying BibFrame documentation")
    except Exception as e:
        logger.error(f"Error querying BibFrame documentation: {str(e)}")
    
    return None


def query_bibframe_docs_cached(tool_name: str, params: dict, timeout: int = 10) -> Optional[dict]:
    """Cached wrapper around ``query_bibframe_docs`` to avoid repeated HTTP calls."""
    if not MCP4BIBFRAME_DOCS_ENABLED:
        return None

    try:
        cache_key = f"{tool_name}:{json.dumps(params, sort_keys=True)}"
    except TypeError:
        cache_key = f"{tool_name}:{str(params)}"

    cached = BIBFRAME_DOCS_CACHE.get(cache_key)
    if cached:
        payload, timestamp = cached
        if time.time() - timestamp < BIBFRAME_DOCS_CACHE_TTL:
            logger.debug(f"Using cached BibFrame docs response for {cache_key}")
            return payload

    response = query_bibframe_docs(tool_name, params, timeout)
    if response is not None:
        BIBFRAME_DOCS_CACHE[cache_key] = (response, time.time())

    return response

def extract_bibframe_terms_from_errors(validation_results: str) -> dict:
    """
    Extract BibFrame properties and classes mentioned in validation errors.
    
    Args:
        validation_results (str): Validation error text
        
    Returns:
        dict: Dictionary with 'properties' and 'classes' lists
    """
    import re
    
    terms = {
        'properties': set(),
        'classes': set()
    }
    
    # Common patterns in validation results
    # Properties often appear as bf:propertyName or ->bf:propertyName
    property_patterns = [
        r'bf:(\w+)',
        r'->bf:(\w+)',
        r'property (\w+)',
        r'missing (\w+)',
        r'requires? (\w+)'
    ]
    
    # Classes often appear as bf:ClassName or "a ClassName"
    class_patterns = [
        r'bf:([A-Z]\w+)',
        r'type ([A-Z]\w+)',
        r'class ([A-Z]\w+)',
        r'<bf:([A-Z]\w+)',
        r'a ([A-Z]\w+)'
    ]
    
    text = validation_results.lower()
    
    # Extract properties
    for pattern in property_patterns:
        matches = re.findall(pattern, validation_results, re.IGNORECASE)
        for match in matches:
            if match and len(match) > 2:  # Skip very short matches
                terms['properties'].add(match.lower())
    
    # Extract classes
    for pattern in class_patterns:
        matches = re.findall(pattern, validation_results)
        for match in matches:
            if match and len(match) > 2:
                terms['classes'].add(match)
    
    # Convert sets to lists
    terms['properties'] = list(terms['properties'])[:5]  # Limit to top 5
    terms['classes'] = list(terms['classes'])[:3]  # Limit to top 3
    
    return terms

def fetch_bibframe_guidance(validation_results: str, rdf_content: str) -> str:
    """
    Fetch relevant BibFrame guidance from the documentation API based on errors.
    
    Args:
        validation_results (str): Validation error messages
        rdf_content (str): Original RDF content
        
    Returns:
        str: Formatted guidance text for inclusion in prompts
    """
    if not MCP4BIBFRAME_DOCS_ENABLED:
        return ""
    
    guidance_parts = []
    
    try:
        # Extract terms from validation errors
        terms = extract_bibframe_terms_from_errors(validation_results)
        logger.info(f"Extracted terms - properties: {terms['properties']}, classes: {terms['classes']}")
        
        # Query information for key properties
        for prop in terms['properties'][:3]:  # Limit queries
            prop_uri = _resolve_bibframe_uri(prop)
            result = query_bibframe_docs_cached("get_property_info", {"property_uri": prop_uri})
            if result and isinstance(result, dict):
                guidance_parts.append(f"\n**{result.get('label', prop)}** ({prop}):")
                if 'definition' in result:
                    guidance_parts.append(f"- Definition: {result['definition']}")
                if 'domain' in result:
                    guidance_parts.append(f"- Used in: {', '.join(result['domain'])}")
                if 'range' in result:
                    guidance_parts.append(f"- Values: {', '.join(result['range'])}")
                if 'examples' in result and result['examples']:
                    guidance_parts.append(f"- Example: {result['examples'][0]}")
        
        # Query information for key classes
        for cls in terms['classes'][:2]:  # Limit queries
            cls_uri = _resolve_bibframe_uri(cls)
            result = query_bibframe_docs_cached("get_class_info", {"class_uri": cls_uri})
            if result and isinstance(result, dict):
                guidance_parts.append(f"\n**{result.get('label', cls)}** class:")
                if 'definition' in result:
                    guidance_parts.append(f"- Definition: {result['definition']}")
                if 'applicable_properties' in result:
                    props = [p.get('label', p.get('property', '')) for p in result['applicable_properties'][:5]]
                    guidance_parts.append(f"- Key properties: {', '.join(props)}")
        
        # If we found AdminMetadata issues, get specific usage guidance
        if any(term in validation_results.lower() for term in ['adminmetadata', 'assigner', '->bf:assigner']):
            result = query_bibframe_docs_cached("get_property_usage", {
                "property_name": "assigner",
                "class_name": "AdminMetadata"
            })
            if result and isinstance(result, dict):
                guidance_parts.append("\n**AdminMetadata/assigner usage:**")
                if 'usage' in result:
                    guidance_parts.append(f"- {result['usage']}")
                if 'examples' in result and result['examples']:
                    guidance_parts.append(f"- Pattern: {result['examples'][0]}")
        
    except Exception as e:
        logger.error(f"Error fetching BibFrame guidance: {str(e)}")
    
    if guidance_parts:
        return "\n".join(guidance_parts)
    return ""

# OpenAI client configuration for the endpoint
def get_openai_client():
    """Get configured OpenAI client for HF Inference Endpoint"""
    if not HF_API_KEY:
        print("❌ No HF_API_KEY available for OpenAI client")
        return None
    
    print(f"🔗 Creating OpenAI client with:")
    print(f"   base_url: {HF_ENDPOINT_URL}")
    print(f"   api_key: {'***' + HF_API_KEY[-4:] if len(HF_API_KEY) > 4 else 'HIDDEN'}")
    
    return OpenAI(
        base_url=HF_ENDPOINT_URL,
        api_key=HF_API_KEY,
        timeout=120.0  # Increase timeout for cold starts
    )

# Sample RDF data for examples (based on real Library of Congress BibFrame)
SAMPLE_VALID_RDF = '''<?xml version="1.0" encoding="UTF-8"?>
<rdf:RDF xmlns:bf="http://id.loc.gov/ontologies/bibframe/"
         xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#"
         xmlns:rdfs="http://www.w3.org/2000/01/rdf-schema#"
         xmlns:bflc="http://id.loc.gov/ontologies/bflc/"
         xmlns:madsrdf="http://www.loc.gov/mads/rdf/v1#">
    
    <bf:Work rdf:about="http://example.org/work/1">
        <rdf:type rdf:resource="http://id.loc.gov/ontologies/bibframe/Text"/>
        <rdf:type rdf:resource="http://id.loc.gov/ontologies/bibframe/Monograph"/>
        
        <bf:title>
            <bf:Title>
                <bf:mainTitle>The knitter's handy book of patterns</bf:mainTitle>
                <bf:subtitle>basic designs in multiple sizes & gauges</bf:subtitle>
            </bf:Title>
        </bf:title>
        
        <bf:contribution>
            <bf:PrimaryContribution>
                <bf:agent>
                    <bf:Agent rdf:about="http://id.loc.gov/rwo/agents/n2001017606">
                        <rdf:type rdf:resource="http://id.loc.gov/ontologies/bibframe/Person"/>
                        <rdfs:label>Budd, Ann, 1956-</rdfs:label>
                    </bf:Agent>
                </bf:agent>
                <bf:role>
                    <bf:Role rdf:about="http://id.loc.gov/vocabulary/relators/aut">
                        <rdfs:label>author</rdfs:label>
                        <bf:code>aut</bf:code>
                    </bf:Role>
                </bf:role>
            </bf:PrimaryContribution>
        </bf:contribution>
        
        <bf:language>
            <bf:Language rdf:about="http://id.loc.gov/vocabulary/languages/eng">
                <rdfs:label xml:lang="en">English</rdfs:label>
                <bf:code rdf:datatype="http://www.w3.org/2001/XMLSchema#string">eng</bf:code>
            </bf:Language>
        </bf:language>
        
        <bf:content>
            <bf:Content rdf:about="http://id.loc.gov/vocabulary/contentTypes/txt">
                <rdfs:label>text</rdfs:label>
                <bf:code>txt</bf:code>
            </bf:Content>
        </bf:content>
        
        <bf:classification>
            <bf:ClassificationLcc>
                <bf:classificationPortion>TT820</bf:classificationPortion>
                <bf:itemPortion>.B877 2002</bf:itemPortion>
                <bf:assigner>
                    <bf:Organization rdf:about="http://id.loc.gov/vocabulary/organizations/dlc">
                        <rdfs:label>United States, Library of Congress</rdfs:label>
                    </bf:Organization>
                </bf:assigner>
            </bf:ClassificationLcc>
        </bf:classification>
        
        <bf:adminMetadata>
            <bf:AdminMetadata>
                <bf:status>
                    <bf:Status rdf:about="http://id.loc.gov/vocabulary/mstatus/n">
                        <rdfs:label>new</rdfs:label>
                        <bf:code>n</bf:code>
                    </bf:Status>
                </bf:status>
                <bf:date rdf:datatype="http://www.w3.org/2001/XMLSchema#date">2001-12-12</bf:date>
                <bf:agent>
                    <bf:Agent rdf:about="http://id.loc.gov/vocabulary/organizations/dlc">
                        <rdf:type rdf:resource="http://id.loc.gov/ontologies/bibframe/Organization"/>
                        <rdfs:label>United States, Library of Congress</rdfs:label>
                    </bf:Agent>
                </bf:agent>
            </bf:AdminMetadata>
        </bf:adminMetadata>
    </bf:Work>
    
</rdf:RDF>'''

SAMPLE_INVALID_RDF = '''<?xml version="1.0" encoding="UTF-8"?>
<rdf:RDF xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#"
         xmlns:bf="http://id.loc.gov/ontologies/bibframe/">
    <!-- Well-formed RDF/XML, but missing required properties to trigger SHACL violations -->
    <bf:Work rdf:about="http://example.org/work/invalid-1">
        <!-- Ensure target class is hit so SHACL runs -->
        <rdf:type rdf:resource="http://id.loc.gov/ontologies/bibframe/Text"/>
        <!-- Missing proper title structure, language, content, adminMetadata -->
        <bf:title>Incomplete Title</bf:title>
    </bf:Work>
</rdf:RDF>'''

# BibFrame Few-Shot Examples (based on real Library of Congress records)
BIBFRAME_CORRECTION_EXAMPLES = {
    "title_structure": {
        "pattern": r"bf:title",
        "wrong": """<bf:title>Simple Title String</bf:title>""",
        "correct": """<bf:title>
    <bf:Title>
        <bf:mainTitle>The knitter's handy book of patterns</bf:mainTitle>
        <bf:subtitle>basic designs in multiple sizes & gauges</bf:subtitle>
    </bf:Title>
</bf:title>"""
    },
    "adminmetadata": {
        "pattern": r"bf:adminMetadata|->bf:assigner",
        "wrong": """<bf:adminMetadata>
    <bf:AdminMetadata>
        <bf:agent rdf:resource="http://example.org/org"/>
        <bf:status>new</bf:status>
    </bf:AdminMetadata>
</bf:adminMetadata>""",
        "correct": """<bf:adminMetadata>
    <bf:AdminMetadata>
        <bf:status>
            <bf:Status rdf:about="http://id.loc.gov/vocabulary/mstatus/n">
                <rdfs:label>new</rdfs:label>
                <bf:code>n</bf:code>
            </bf:Status>
        </bf:status>
        <bf:date rdf:datatype="http://www.w3.org/2001/XMLSchema#date">2001-12-12</bf:date>
        <bf:agent>
            <bf:Agent rdf:about="http://id.loc.gov/vocabulary/organizations/dlc">
                <rdf:type rdf:resource="http://id.loc.gov/ontologies/bibframe/Organization"/>
                <rdfs:label>United States, Library of Congress</rdfs:label>
            </bf:Agent>
        </bf:agent>
    </bf:AdminMetadata>
</bf:adminMetadata>"""
    },
    "contribution": {
        "pattern": r"bf:contribution",
        "wrong": """<bf:contribution>Author Name</bf:contribution>""",
        "correct": """<bf:contribution>
    <bf:PrimaryContribution>
        <bf:agent>
            <bf:Agent rdf:about="http://id.loc.gov/rwo/agents/n2001017606">
                <rdf:type rdf:resource="http://id.loc.gov/ontologies/bibframe/Person"/>
                <rdfs:label>Budd, Ann, 1956-</rdfs:label>
            </bf:Agent>
        </bf:agent>
        <bf:role>
            <bf:Role rdf:about="http://id.loc.gov/vocabulary/relators/ctb">
                <rdfs:label>contributor</rdfs:label>
                <bf:code>ctb</bf:code>
            </bf:Role>
        </bf:role>
    </bf:PrimaryContribution>
</bf:contribution>"""
    },
    "language": {
        "pattern": r"bf:language",
        "wrong": """<bf:language>English</bf:language>""",
        "correct": """<bf:language>
    <bf:Language rdf:about="http://id.loc.gov/vocabulary/languages/eng">
        <rdfs:label xml:lang="en">English</rdfs:label>
        <bf:code rdf:datatype="http://www.w3.org/2001/XMLSchema#string">eng</bf:code>
    </bf:Language>
</bf:language>"""
    },
    "content": {
        "pattern": r"bf:content",
        "wrong": """<bf:content>Text</bf:content>""",
        "correct": """<bf:content>
    <bf:Content rdf:about="http://id.loc.gov/vocabulary/contentTypes/txt">
        <rdfs:label>text</rdfs:label>
        <bf:code>txt</bf:code>
    </bf:Content>
</bf:content>"""
    },
    "classification": {
        "pattern": r"bf:classification",
        "wrong": """<bf:classification>TT820 .B877 2002</bf:classification>""",
        "correct": """<bf:classification>
    <bf:ClassificationLcc>
        <bf:classificationPortion>TT820</bf:classificationPortion>
        <bf:itemPortion>.B877 2002</bf:itemPortion>
        <bf:assigner>
            <bf:Organization rdf:about="http://id.loc.gov/vocabulary/organizations/dlc">
                <rdfs:label>United States, Library of Congress</rdfs:label>
                <bf:code rdf:datatype="http://id.loc.gov/datatypes/orgs/code">DLC</bf:code>
            </bf:Organization>
        </bf:assigner>
        <bf:status>
            <bf:Status rdf:about="http://id.loc.gov/vocabulary/mstatus/uba">
                <rdfs:label>used by assigner</rdfs:label>
                <bf:code>uba</bf:code>
            </bf:Status>
        </bf:status>
    </bf:ClassificationLcc>
</bf:classification>"""
    },
    "subject": {
        "pattern": r"bf:subject",
        "wrong": """<bf:subject>Knitting--Patterns</bf:subject>""",
        "correct": """<bf:subject>
    <bf:Topic rdf:about="http://id.loc.gov/authorities/subjects/sh85072708">
        <rdfs:label xml:lang="en">Knitting--Patterns</rdfs:label>
        <madsrdf:componentList rdf:parseType="Collection">
            <madsrdf:Authority>
                <madsrdf:authoritativeLabel xml:lang="en">Knitting</madsrdf:authoritativeLabel>
                <madsrdf:elementList>
                    <madsrdf:TopicElement>
                        <madsrdf:elementValue xml:lang="en">Knitting</madsrdf:elementValue>
                    </madsrdf:TopicElement>
                </madsrdf:elementList>
            </madsrdf:Authority>
        </madsrdf:componentList>
    </bf:Topic>
</bf:subject>"""
    }
}

# MCP Server Tools (can be used independently)
def validate_rdf_tool(rdf_content: str, template: str = "monograph") -> dict:
    """
    Validate RDF/XML content against SHACL templates.
    
    This tool validates RDF/XML data against predefined SHACL shapes to ensure
    compliance with metadata standards like BIBFRAME. Returns detailed validation
    results with conformance status and specific violation information.
    
    Args:
        rdf_content (str): The RDF/XML content to validate
        template (str): Validation template to use ('monograph' or 'custom')
    
    Returns:
        dict: Validation results with conformance status and detailed feedback
    """
    if not rdf_content:
        return {"error": "No RDF/XML content provided", "conforms": False}
    
    if not VALIDATOR_AVAILABLE:
        logger.error("Validator module not available")
        return {
            "error": "Validator not available - ensure validator.py is present",
            "conforms": False
        }
    
    try:
        # Fast syntax check before SHACL to give clearer errors on XML/prefix issues
        try:
            try:
                import rdflib  # type: ignore
            except ImportError:
                rdflib = None  # type: ignore
            if rdflib:
                g = rdflib.Graph()  # type: ignore
                # Parse as RDF/XML; raise on syntax errors like unbound prefixes
                g.parse(data=rdf_content, format="application/rdf+xml")  # type: ignore
            else:
                logger.info("rdflib not installed; skipping pre-parse RDF/XML syntax check")
        except Exception as parse_err:
            logger.error(f"RDF/XML parse error before validation: {parse_err}")
            return {
                "error": f"RDF/XML parse error: {parse_err}",
                "conforms": False
            }
        
        # Log what we're validating
        logger.info(f"Validating RDF with template '{template}', content length: {len(rdf_content)}")
        
        # Call the validator
        conforms, results_text = validate_rdf(rdf_content.encode('utf-8'), template)
        
        # Debug logging
        logger.info(f"Validation result - conforms: {conforms}, results length: {len(results_text) if results_text else 0}")
        
        # If no results text but claims to conform, something might be wrong
        if conforms and (not results_text or len(results_text.strip()) == 0):
            results_text = "Validation passed with no specific feedback."
        elif not conforms and (not results_text or len(results_text.strip()) == 0):
            results_text = "Validation failed but no specific errors were returned. Check the RDF syntax and structure."
            
        return {
            "conforms": conforms,
            "results": results_text if results_text else "",
            "template": template,
            "status": "✅ Valid RDF" if conforms else "❌ Invalid RDF"
        }
        
    except ImportError as e:
        logger.error(f"Import error in validator: {str(e)}")
        return {
            "error": f"Validator import error: {str(e)}. Check that all dependencies are installed.",
            "conforms": False
        }
    except AttributeError as e:
        logger.error(f"Validator function not found: {str(e)}")
        return {
            "error": f"Validator function error: {str(e)}. Check validator.py implementation.",
            "conforms": False
        }
    except Exception as e:
        logger.error(f"Validation error: {str(e)}")
        import traceback
        logger.error(f"Full traceback: {traceback.format_exc()}")
        return {
            "error": f"Validation failed: {str(e)}",
            "conforms": False
        }

def filter_validation_results_by_class(validation_results: str, rdf_content: str) -> dict:
    """
    Filter validation results by RDF class (Work, Instance, etc.)
    
    Args:
        validation_results (str): Full validation results
        rdf_content (str): Original RDF content
        
    Returns:
        dict: Validation results organized by class
    """
    import re
    
    # Parse validation results to extract class information
    class_results = {
        'Work': [],
        'Instance': [],
        'Title': [],
        'Contribution': [],
        'AdminMetadata': [],
        'Other': []
    }
    
    lines = validation_results.split('\n')
    current_section = []
    current_class = 'Other'
    
    for line in lines:
        # Detect which class this error relates to
        if 'bf:Work' in line or '/work/' in line:
            current_class = 'Work'
        elif 'bf:Instance' in line or '/instance/' in line:
            current_class = 'Instance'
        elif 'bf:Title' in line:
            current_class = 'Title'
        elif 'bf:Contribution' in line:
            current_class = 'Contribution'
        elif 'bf:AdminMetadata' in line or 'AdminMetadata' in line or '->bf:assigner' in line:
            # Many admin violations show assigner path; map to AdminMetadata
            current_class = 'AdminMetadata'
        
        # Collect lines for current violation
        if 'Constraint Violation' in line:
            if current_section:
                class_results[current_class].extend(current_section)
            current_section = [line]
        elif line.strip():
            current_section.append(line)
    
    # Add last section
    if current_section:
        class_results[current_class].extend(current_section)
    
    # Remove empty classes
    return {k: '\n'.join(v) for k, v in class_results.items() if v}

def get_ai_suggestions(validation_results: str, rdf_content: str, include_warnings: bool = False) -> str:
    """Generate AI-powered, plain-language suggestions based on validation results.

    Avoids RDF/SHACL jargon and focuses on actionable fixes.
    """
    if not OPENAI_AVAILABLE:
        return generate_manual_suggestions(validation_results)

    current_api_key = os.getenv('HF_API_KEY', '')
    if not current_api_key:
        return f"""
🔑 **AI suggestions disabled**: Please set your Hugging Face API key as a Secret in your Space settings.

{generate_manual_suggestions(validation_results)}
"""

    try:
        client = get_openai_client()
        if not client:
            return f"""
🔑 **AI suggestions disabled**: HF_API_KEY not configured.

{generate_manual_suggestions(validation_results)}
"""

        severity_instruction = (
            "Focus only on violations (errors) and ignore any warnings."
            if not include_warnings else
            "Address both violations and warnings."
        )

        # Get BibFrame documentation for context
        bibframe_guidance = fetch_bibframe_guidance(validation_results, rdf_content)
        doc_section = ""
        if bibframe_guidance:
            doc_section = f"""
Reference information from BibFrame ontology:
{bibframe_guidance}
"""

        # Group errors by class to focus the prompt
        class_results = filter_validation_results_by_class(validation_results, rdf_content)
        if class_results:
            primary_class = max(class_results.keys(), key=lambda k: len(class_results[k]))
            focused_results = class_results[primary_class]
        else:
            primary_class = "Record"
            focused_results = validation_results

        simplified_summary = parse_shacl_results_for_ai(focused_results)
        relevant_rdf = extract_relevant_rdf_section(rdf_content, primary_class)

        prompt = f"""
You are a helpful metadata librarian. Write in plain language (no RDF/SHACL jargon). Analyze the validation errors for the {primary_class} and provide concise, actionable fixes.

{severity_instruction}
{doc_section}

Validation Errors for {primary_class}:
{focused_results[:1500]}

Validation Summary (plain language):
{simplified_summary}

Relevant RDF Section:
{relevant_rdf[:800]}

Instructions:
1. ONE sentence: What's wrong with this {primary_class}?
2. List errors (max 3 words each)
3. Show exact XML fixes

Format:
**Issue:** [One sentence about the {primary_class} problem]

**Errors:**
• Error 1
• Error 2

**Fix:**
```xml
[Complete corrected {primary_class} section]
```

Be ultra-concise. Show the fix, not explanations."""

        chat_completion = client.chat.completions.create(
            model=HF_MODEL,
            messages=[
                {
                    "role": "system",
                    "content": "You are a friendly librarian helping fix catalog records. Never use technical RDF or SHACL terminology. Use the BibFrame documentation provided to ensure accuracy."
                },
                {
                    "role": "user",
                    "content": prompt
                }
            ],
            max_tokens=800,
            temperature=0.5,
            top_p=0.9
        )

        generated_text = chat_completion.choices[0].message.content
        generated_text = clean_technical_jargon(generated_text)

        other_classes = [k for k in class_results.keys() if k != primary_class]
        class_note = (
            f"\n\n📌 **Note:** Focused on {primary_class} errors. " +
            (f"Also found issues in: {', '.join(other_classes)}" if other_classes else "")
        )

        return f"🤖 **AI-Powered Suggestions ({('Violations + Warnings' if include_warnings else 'Violations Only')}):**\n\n{generated_text}{class_note}"

    except Exception as e:
        logger.error(f"OpenAI/HF Inference Endpoint error: {str(e)}")
        return f"""
❌ **AI suggestions error**: {str(e)}

{generate_manual_suggestions(validation_results)}
"""

def extract_relevant_rdf_section(rdf_content: str, class_name: str) -> str:
    """
    Extract only the relevant RDF section for a specific class
    
    Args:
        rdf_content (str): Full RDF content
        class_name (str): Class name to extract (Work, Instance, etc.)
        
    Returns:
        str: Relevant RDF section
    """
    import re
    
    # Map class names to RDF patterns
    patterns = {
        'Work': r'<bf:Work.*?</bf:Work>',
        'Instance': r'<bf:Instance.*?</bf:Instance>',
        'Title': r'<bf:Title.*?</bf:Title>',
        'Contribution': r'<bf:Contribution.*?</bf:Contribution>',
        'AdminMetadata': r'<bf:AdminMetadata.*?</bf:AdminMetadata>'
    }
    
    pattern = patterns.get(class_name)
    if not pattern:
        return rdf_content[:1000]  # Fallback to first 1000 chars
    
    # Extract matching section
    match = re.search(pattern, rdf_content, re.DOTALL)
    if match:
        section = match.group(0)
        # Also include namespace declarations
        namespaces = re.findall(r'xmlns:\w+="[^"]*"', rdf_content[:500])
        if namespaces:
            return f"<!-- Namespaces: {' '.join(namespaces[:3])} -->\n{section}"
        return section
    
    return rdf_content[:1000]  # Fallback

## [Removed duplicate get_ai_correction definition – unified below]

def merge_corrected_sections(original_rdf: str, corrected_sections: dict) -> str:
    """
    Merge corrected class sections back into the original RDF
    
    Args:
        original_rdf (str): Original RDF content
        corrected_sections (dict): Corrected sections by class
        
    Returns:
        str: Merged RDF with corrections
    """
    import re
    
    result = original_rdf
    
    # Replace each corrected section
    for class_name, corrected_section in corrected_sections.items():
        patterns = {
            'Work': r'<bf:Work.*?</bf:Work>',
            'Instance': r'<bf:Instance.*?</bf:Instance>',
            'Title': r'<bf:Title.*?</bf:Title>',
            'Contribution': r'<bf:Contribution.*?</bf:Contribution>',
            'AdminMetadata': r'<bf:AdminMetadata.*?</bf:AdminMetadata>'
        }
        
        pattern = patterns.get(class_name)
        if pattern:
            result = re.sub(pattern, corrected_section, result, count=1, flags=re.DOTALL)
    
    return result

# Sample RDF data for examples
# MCP Server Tools (can be used independently)
# Note: This section exists earlier in the file, we're removing the duplicates
    """
    Generate AI-powered fix suggestions for invalid RDF/XML.
    
    This tool analyzes validation results and provides actionable suggestions
    for fixing RDF/XML validation errors using AI or rule-based analysis.
    
    Args:
        validation_results (str): The validation error messages
        rdf_content (str): The original RDF/XML content that failed validation
        include_warnings (bool): Whether to include warnings in suggestions
    
    Returns:
        str: Detailed suggestions for fixing the RDF validation issues
    """
    
    if not OPENAI_AVAILABLE:
        return generate_manual_suggestions(validation_results)
    
    # Get API key dynamically at runtime
    current_api_key = os.getenv('HF_API_KEY', '')
    if not current_api_key:
        return f"""
🔑 **AI suggestions disabled**: Please set your Hugging Face API key as a Secret in your Space settings.

{generate_manual_suggestions(validation_results)}
"""
    
    try:
        # Use OpenAI client with your Hugging Face Inference Endpoint
        client = get_openai_client()
        if not client:
            return f"""
🔑 **AI suggestions disabled**: HF_API_KEY not configured.

{generate_manual_suggestions(validation_results)}
"""
        
        severity_instruction = "Focus only on violations (errors) and ignore any warnings." if not include_warnings else "Address both violations and warnings."
        
        prompt = f"""You are an expert in RDF/XML and SHACL validation. Analyze the validation errors and provide CONCISE, ACTIONABLE fix suggestions.

{severity_instruction}

Validation Results:
{validation_results}

Original RDF (first 1000 chars):
{rdf_content[:1000]}...

Instructions:
1. Start with a ONE-SENTENCE summary of the main issue
2. List the specific errors in bullet points (max 5 words per error)
3. Provide the exact fix for each error with code snippets
4. Keep explanations minimal - focus on solutions

Format:
**Main Issue:** [One sentence]

**Errors Found:**
• Error 1 name
• Error 2 name

**Fixes:**
1. **Error 1**: 
   ```xml
   [exact code to add/fix]
   ```
2. **Error 2**:
   ```xml
   [exact code to add/fix]
   ```

Be direct and solution-focused. No lengthy explanations."""
        
        # Make API call using OpenAI client
        print(f"🔄 Making API call to: {HF_ENDPOINT_URL}")
        print(f"🔄 Using model: {HF_MODEL}")
        print(f"🔄 Include warnings: {include_warnings}")
        
        chat_completion = client.chat.completions.create(
            model=HF_MODEL,
            messages=[
                {
                    "role": "system",
                    "content": "You are a friendly librarian helping fix catalog records. Never use technical RDF or SHACL terminology."
                },
                {
                    "role": "user",
                    "content": prompt
                }
            ],
            max_tokens=1500,
            temperature=0.6,
            top_p=0.9
        )
        
        print("✅ API call successful")
        generated_text = chat_completion.choices[0].message.content
        return f"🤖 **AI-Powered Suggestions ({('Violations + Warnings' if include_warnings else 'Violations Only')}):**\n\n{generated_text}"
        
    except Exception as e:
        logger.error(f"OpenAI/HF Inference Endpoint error: {str(e)}")
        return f"""
❌ **AI suggestions error**: {str(e)}

{generate_manual_suggestions(validation_results)}
"""

def extract_rdf_from_response(response: str) -> str:
    """
    Extract RDF/XML content from AI response, handling code blocks.
    
    Args:
        response (str): AI response that may contain RDF wrapped in code blocks
    
    Returns:
        str: Extracted RDF/XML content
    """
    response = response.strip()
    
    # Handle ```xml code blocks
    if "```xml" in response:
        try:
            return response.split("```xml")[1].split("```")[0].strip()
        except IndexError:
            pass
    
    # Handle generic ``` code blocks
    if "```" in response and response.count("```") >= 2:
        try:
            return response.split("```")[1].split("```")[0].strip()
        except IndexError:
            pass
    
    # If no code blocks found, return the response as-is
    return response

def fix_common_rdf_errors(rdf_xml: str) -> str:
    """
    Fix common RDF/XML errors that AI models generate.
    
    Args:
        rdf_xml (str): RDF/XML that may contain common errors
        
    Returns:
        str: Fixed RDF/XML
    """
    import re
    
    # Remove any rdf:parseType attributes (common AI mistake)
    rdf_xml = re.sub(r'\s+rdf:parseType="[^"]*"', '', rdf_xml)
    
    # Fix bf:title if it's just a string (should be nested structure)
    rdf_xml = re.sub(
        r'<bf:title>([^<]+)</bf:title>',
        r'<bf:title><bf:Title><bf:mainTitle>\1</bf:mainTitle></bf:Title></bf:title>',
        rdf_xml
    )
    
    # Fix bf:language if it's a string instead of URI
    language_map = {
        'English': 'http://id.loc.gov/vocabulary/languages/eng',
        'eng': 'http://id.loc.gov/vocabulary/languages/eng',
        'Spanish': 'http://id.loc.gov/vocabulary/languages/spa',
        'French': 'http://id.loc.gov/vocabulary/languages/fre',
    }
    for lang_text, lang_uri in language_map.items():
        rdf_xml = re.sub(
            f'<bf:language>{lang_text}</bf:language>',
            f'<bf:language rdf:resource="{lang_uri}"/>',
            rdf_xml,
            flags=re.IGNORECASE
        )
    
    # Fix bf:content if it's a string
    content_map = {
        'Text': 'http://id.loc.gov/vocabulary/contentTypes/txt',
        'text': 'http://id.loc.gov/vocabulary/contentTypes/txt',
    }
    for content_text, content_uri in content_map.items():
        rdf_xml = re.sub(
            f'<bf:content>{content_text}</bf:content>',
            f'<bf:content rdf:resource="{content_uri}"/>',
            rdf_xml,
            flags=re.IGNORECASE
        )
    
    return rdf_xml


def extract_error_focus_points(validation_results: str) -> Dict[str, List[str]]:
    """Identify the specific focus nodes and properties mentioned in validation errors."""
    import re

    focus = {
        "properties": [],
        "focus_nodes": [],
        "missing_properties": [],
        "classes": [],
    }

    if not validation_results:
        return focus

    property_set = set()
    missing_set = set()
    node_set = set()

    for match in re.finditer(r"Focus Node:\s*(?:<)?([^\s>]+)(?:>)?", validation_results):
        node_set.add(match.group(1))

    for match in re.finditer(r"Result Path:\s*(?:http://[^/]+/)?([A-Za-z]+)", validation_results):
        property_set.add(match.group(1))

    for match in re.finditer(r"Less than \d+ values on .*->bf:([A-Za-z]+)", validation_results):
        missing_set.add(match.group(1))

    focus["properties"] = sorted(property_set)
    focus["focus_nodes"] = sorted(node_set)
    focus["missing_properties"] = sorted(missing_set)
    return focus


def _resolve_bibframe_uri(name: str) -> str:
    if not name:
        return name
    if name.startswith("http://") or name.startswith("https://"):
        return name
    if ":" in name:
        prefix, local = name.split(":", 1)
        if prefix == "bf":
            return f"http://id.loc.gov/ontologies/bibframe/{local}"
    return f"http://id.loc.gov/ontologies/bibframe/{name}"


def get_targeted_bibframe_guidance(properties: List[str], classes: List[str]) -> Dict[str, dict]:
    """Fetch BibFrame documentation for only the specified properties/classes."""
    guidance: Dict[str, dict] = {}

    if not MCP4BIBFRAME_DOCS_ENABLED:
        return guidance

    for prop in properties[:5]:
        prop_uri = _resolve_bibframe_uri(prop)
        result = query_bibframe_docs_cached("get_property_info", {"property_uri": prop_uri}, timeout=5)
        if result:
            guidance[prop] = result

    for cls in classes[:5]:
        cls_uri = _resolve_bibframe_uri(cls)
        result = query_bibframe_docs_cached("get_class_info", {"class_uri": cls_uri}, timeout=5)
        if result:
            guidance[cls] = result

    return guidance


def generate_property_specific_fix(property_name: str, guidance: Optional[dict] = None) -> str:
    """Generate a BibFrame-compliant snippet for a specific missing property."""
    guidance = guidance or {}
    prop = property_name.lower() if property_name else ""

    if prop == "title":
        return """<bf:title>
    <bf:Title>
        <bf:mainTitle>PLACEHOLDER_TITLE</bf:mainTitle>
    </bf:Title>
</bf:title>"""

    if prop == "language":
        return """<bf:language>
    <bf:Language rdf:about="http://id.loc.gov/vocabulary/languages/eng">
        <rdfs:label xml:lang="en">English</rdfs:label>
        <bf:code rdf:datatype="http://www.w3.org/2001/XMLSchema#string">eng</bf:code>
    </bf:Language>
</bf:language>"""

    if prop == "content":
        return """<bf:content>
    <bf:Content rdf:about="http://id.loc.gov/vocabulary/contentTypes/txt">
        <rdfs:label>text</rdfs:label>
        <bf:code>txt</bf:code>
    </bf:Content>
</bf:content>"""

    if prop == "contribution":
        return """<bf:contribution>
    <bf:PrimaryContribution>
        <bf:agent>
            <bf:Agent>
                <rdf:type rdf:resource="http://id.loc.gov/ontologies/bibframe/Person"/>
                <rdfs:label>Author Name</rdfs:label>
            </bf:Agent>
        </bf:agent>
        <bf:role>
            <bf:Role rdf:about="http://id.loc.gov/vocabulary/relators/aut">
                <rdfs:label>author</rdfs:label>
                <bf:code>aut</bf:code>
            </bf:Role>
        </bf:role>
    </bf:PrimaryContribution>
</bf:contribution>"""

    if prop == "classification":
        return """<bf:classification>
    <bf:ClassificationLcc>
        <bf:classificationPortion>TT820</bf:classificationPortion>
        <bf:itemPortion>.B877 2002</bf:itemPortion>
        <bf:assigner>
            <bf:Organization rdf:about="http://id.loc.gov/vocabulary/organizations/dlc">
                <rdfs:label>United States, Library of Congress</rdfs:label>
            </bf:Organization>
        </bf:assigner>
    </bf:ClassificationLcc>
</bf:classification>"""

    if prop == "adminmetadata":
        return """<bf:adminMetadata>
    <bf:AdminMetadata>
        <bf:status>
            <bf:Status rdf:about="http://id.loc.gov/vocabulary/mstatus/n">
                <rdfs:label>new</rdfs:label>
                <bf:code>n</bf:code>
            </bf:Status>
        </bf:status>
        <bf:date rdf:datatype="http://www.w3.org/2001/XMLSchema#date">2024-01-01</bf:date>
        <bf:agent>
            <bf:Agent rdf:about="http://id.loc.gov/vocabulary/organizations/dlc">
                <rdf:type rdf:resource="http://id.loc.gov/ontologies/bibframe/Organization"/>
                <rdfs:label>United States, Library of Congress</rdfs:label>
            </bf:Agent>
        </bf:agent>
    </bf:AdminMetadata>
</bf:adminMetadata>"""

    # Fallback: simple literal placeholder
    return f"<bf:{property_name}>PLACEHOLDER_VALUE</bf:{property_name}>"

def get_ai_correction(validation_results: str, rdf_content: str, template: str = 'monograph', max_attempts: int = None, include_warnings: bool = False, enable_validation_loop: bool | None = None, cache_key: Optional[str] = None, steps_log: Optional[List[str]] = None) -> str:
    """
    Generate AI-powered corrected RDF/XML based on validation errors.
    
    This tool takes invalid RDF/XML and validation results, then generates
    a corrected version that addresses all identified validation issues.
    The generated correction is validated before being returned to the user.
    
    Args:
        validation_results (str): The validation error messages
        rdf_content (str): The original invalid RDF/XML content
        template (str): The validation template to use
        max_attempts (int): Maximum number of attempts to generate valid RDF (uses MAX_CORRECTION_ATTEMPTS if None)
        include_warnings (bool): Whether to fix warnings in addition to violations
    
    Returns:
        str: Corrected RDF/XML that should pass validation
    """
    
    # Determine whether to iterate based on parameter or global default
    iterate_enabled = ENABLE_VALIDATION_LOOP if enable_validation_loop is None else enable_validation_loop
    if steps_log is not None:
        steps_log.append(f"Planning correction: iterate_enabled={iterate_enabled}, include_warnings={include_warnings}")
    # Use configuration default if not specified
    if max_attempts is None:
        max_attempts = MAX_CORRECTION_ATTEMPTS
    if steps_log is not None:
        steps_log.append(f"Max attempts set to {max_attempts}")
    # If iteration disabled, force single attempt
    if not iterate_enabled:
        max_attempts = 1
        if steps_log is not None:
            steps_log.append("Iteration disabled; forcing single attempt")

    if cache_key is None and validation_results and rdf_content:
        cache_key = _make_fix_cache_key(validation_results, rdf_content, template)
    if cache_key:
        cached_result = _get_cached_correction(cache_key, steps_log)
        if cached_result is not None:
            return cached_result
    
    if not OPENAI_AVAILABLE:
        if steps_log is not None:
            steps_log.append("OPENAI client not available; falling back to manual hints")
        return generate_manual_correction_hints(validation_results, rdf_content)
    
    # Get API key dynamically at runtime
    current_api_key = os.getenv('HF_API_KEY', '')
    if not current_api_key:
        if steps_log is not None:
            steps_log.append("HF_API_KEY not set; cannot call model; returning manual hints")
        return f"""<!-- AI correction disabled: Set HF_API_KEY as a Secret in your Space settings -->

{generate_manual_correction_hints(validation_results, rdf_content)}"""
    
    try:
        client = get_openai_client()
        if not client:
            if steps_log is not None:
                steps_log.append("Failed to initialize OpenAI client; returning manual hints")
            return f"""<!-- AI correction disabled: HF_API_KEY not configured -->

{generate_manual_correction_hints(validation_results, rdf_content)}"""
        
        # Fetch BibFrame documentation guidance
        if steps_log is not None:
            steps_log.append("Fetching BibFrame documentation guidance...")
        
        bibframe_guidance = fetch_bibframe_guidance(validation_results, rdf_content)
        
        if bibframe_guidance:
            if steps_log is not None:
                steps_log.append(f"Retrieved BibFrame guidance ({len(bibframe_guidance)} chars)")
            guidance_section = f"""
BIBFRAME DOCUMENTATION (from official ontology):
{bibframe_guidance}

Apply the above BibFrame definitions and patterns when correcting the RDF/XML.
"""
        else:
            guidance_section = ""
            if steps_log is not None:
                steps_log.append("No specific BibFrame guidance retrieved")
        
        # Add timeout protection
        import time
        start_time = time.time()
        timeout = 45  # Reduced to 45 second total timeout for speed
        if steps_log is not None:
            steps_log.append(f"Timeout budget: {timeout}s total")

        severity_instruction = "Fix only the violations (errors) and ignore any warnings." if not include_warnings else "Fix both violations and warnings."
        
        # Try multiple attempts to generate valid RDF
        for attempt in range(max_attempts):
            # Check timeout
            elapsed = time.time() - start_time
            if elapsed > timeout:
                if steps_log is not None:
                    steps_log.append(f"Timeout reached after {int(elapsed)}s; stopping attempts")
                print(f"⏰ Timeout reached after {timeout} seconds")
                break
                
            attempt_no = attempt + 1
            if steps_log is not None:
                steps_log.append(f"Attempt {attempt_no}/{max_attempts}: requesting model correction")
            print(f"🔄 Correction attempt {attempt_no}/{max_attempts}")
            
            # Targeted AdminMetadata guidance inferred from results text
            needs_assigner = ("->bf:assigner" in validation_results) or (" bf:assigner" in validation_results)
            admin_guidance = ""
            if needs_assigner:
                admin_guidance = """
IMPORTANT: For each <bf:AdminMetadata>, ensure it has a direct child <bf:assigner>.
Rules:
- If <bf:agent rdf:resource=\"...\"/> exists, add <bf:assigner rdf:resource=\"...\"/> with the SAME URI.
- Else if <bf:descriptionModifier rdf:resource=\"...\"/> exists, add <bf:assigner rdf:resource=\"...\"/> with the SAME URI.
- Else if a <bf:identifiedBy> block contains <bf:assigner rdf:resource=\"...\"/>, copy that URI to a TOP-LEVEL <bf:assigner>.
Keep all existing content; only add missing <bf:assigner> where required.
"""

            # Build few-shot examples based on the errors found
            examples_to_include = []
            validation_lower = validation_results.lower()
            
            # Check each example pattern against validation results
            for name, example in BIBFRAME_CORRECTION_EXAMPLES.items():
                pattern = example.get("pattern", name)
                if re.search(pattern, validation_results, re.IGNORECASE):
                    examples_to_include.append((name, example))
                    if steps_log is not None:
                        steps_log.append(f"Including {name} example based on pattern match")
            
            few_shot_section = ""
            if examples_to_include:
                few_shot_section = "\n\nCORRECT BIBFRAME PATTERNS (from Library of Congress records):\n"
                few_shot_section += "NEVER use simple strings - always use nested structures as shown below:\n\n"
                for name, example in examples_to_include:
                    few_shot_section += f"{name.upper()}:\n"
                    few_shot_section += f"❌ WRONG:\n```xml\n{example['wrong']}\n```\n"
                    few_shot_section += f"✅ CORRECT:\n```xml\n{example['correct']}\n```\n\n"

            # Add critical rules based on real patterns
            critical_rules = """
CRITICAL RDF/XML RULES (from real BibFrame):
1. NEVER use rdf:parseType except for "Collection" on madsrdf:componentList
2. Properties like bf:title, bf:language, bf:content MUST have nested typed resources
3. Use rdf:about for resource URIs, not rdf:resource on the property element
4. bf:adminMetadata can appear multiple times in one record
5. Status, Role, Language etc. are OBJECTS with rdf:about URIs, not literals
6. Date values use rdf:datatype for typing (e.g., xsd:date, xsd:dateTime)
7. Every bf:AdminMetadata needs BOTH bf:agent AND bf:assigner if validation requires it
"""

            prompt = f"""You are an expert in RDF/XML and BibFrame cataloging. Fix the following RDF/XML based on the validation errors and official BibFrame documentation.

{severity_instruction}
{admin_guidance}
{guidance_section}
{critical_rules}
{few_shot_section}

Validation Errors:
{validation_results}

Original RDF/XML:
{rdf_content}

{f"Previous attempt {attempt} still had validation errors. Please fix ALL issues this time." if attempt > 0 else ""}

INSTRUCTIONS:
1. Return ONLY valid RDF/XML - no explanations
2. Follow the EXACT patterns shown in the examples above
3. Use proper nested structures - NO simple string values for complex properties
4. Keep ALL namespace declarations
5. Fix ALL validation errors"""
            
            try:
                # Update system prompt to be even more explicit
                system_prompt = """You are an RDF/XML expert following Library of Congress BibFrame patterns.
Output ONLY valid RDF/XML following these rules:
- Start with <?xml version="1.0" encoding="UTF-8"?>
- NO markdown, NO explanations
- Use EXACT structure patterns from the examples
- Complex properties need nested typed resources
- rdf:parseType ONLY for Collection on madsrdf:componentList
- Status/Role/Language are OBJECTS with URIs, not strings"""

                chat_completion = client.chat.completions.create(
                    model=HF_MODEL,
                    messages=[
                        {
                            "role": "system",
                            "content": system_prompt
                        },
                        {
                            "role": "user", 
                            "content": prompt
                        }
                    ],
                    max_tokens=1500,
                    temperature=0.0,
                    timeout=20  # Reduced to 20 second timeout per API call for speed
                )
                
                corrected_rdf = chat_completion.choices[0].message.content.strip()
                if steps_log is not None:
                    steps_log.append(f"Attempt {attempt_no}: model responded; extracting and fixing common errors")
                
                # Extract RDF content if it's wrapped in code blocks
                corrected_rdf = extract_rdf_from_response(corrected_rdf)
                
                # Fix common AI mistakes
                corrected_rdf = fix_common_rdf_errors(corrected_rdf)
                
                # Only validate if we have the validator and haven't hit timeout
                if VALIDATOR_AVAILABLE and (time.time() - start_time < timeout - 10):
                    try:
                        # Quick validation check
                        conforms, new_results = validate_rdf(corrected_rdf.encode('utf-8'), template)
                        
                        if conforms:
                            if steps_log is not None:
                                steps_log.append(f"Attempt {attempt_no}: correction PASSED validation")
                            print(f"✅ Correction validated successfully on attempt {attempt_no}")
                            result_text = f"""<!-- AI-generated correction validated successfully -->
{corrected_rdf}"""
                            if cache_key:
                                _store_correction_in_cache(cache_key, result_text, steps_log)
                            return result_text
                        else:
                            if steps_log is not None:
                                steps_log.append(f"Attempt {attempt_no}: still invalid; will retry with updated errors")
                            print(f"❌ Correction attempt {attempt_no} still has validation errors")
                            # Update validation_results for next attempt
                            validation_results = new_results
                            
                    except Exception as e:
                        if steps_log is not None:
                            steps_log.append(f"Attempt {attempt_no}: error during validation: {str(e)} — returning correction anyway")
                        print(f"⚠️ Error validating correction attempt {attempt_no}: {str(e)}")
                        # If validation fails, return the correction anyway
                        return f"""<!-- AI-generated correction (validation check failed) -->
{corrected_rdf}"""
                else:
                    # If validator not available or timeout approaching, return the correction
                    if steps_log is not None:
                        steps_log.append("Skipping validation check (validator unavailable or timeout)")
                    print("⚠️ Returning correction without validation")
                    return f"""<!-- AI-generated correction (validation skipped) -->
{corrected_rdf}"""
                    
            except Exception as api_error:
                if steps_log is not None:
                    steps_log.append(f"Attempt {attempt_no}: API error: {str(api_error)}")
                print(f"❌ API error on attempt {attempt_no}: {str(api_error)}")
                if attempt == max_attempts - 1:  # Last attempt
                    raise api_error
                continue
        
        # All attempts failed or timed out
        if steps_log is not None:
            steps_log.append("All attempts failed or timed out; returning manual hints")
        return f"""<!-- AI correction failed after {max_attempts} attempts or timeout -->
<!-- Please correct manually using the validation results as a guide -->

{generate_manual_correction_hints(validation_results, rdf_content)}"""
        
    except Exception as e:
        logger.error(f"LLM API error: {str(e)}")
        if steps_log is not None:
            steps_log.append(f"Fatal error invoking model: {str(e)}")
        return f"""<!-- Error generating AI correction: {str(e)} -->

{generate_manual_correction_hints(validation_results, rdf_content)}"""


def get_ai_correction_targeted(validation_results: str, rdf_content: str, template: str = 'monograph', max_attempts: int = None, include_warnings: bool = False, enable_validation_loop: bool | None = None, steps_log: Optional[List[str]] = None) -> str:
    """Fast path that attempts structured quick fixes before invoking the full AI loop."""

    if steps_log:
        steps_log.append("\n" + "=" * 70)
        steps_log.append("📊 INITIAL VALIDATION ERRORS:")
        steps_log.append("=" * 70)
        # Show summary of validation errors
        error_lines = [line.strip() for line in validation_results.split('\n') if 'Less than' in line or 'Message:' in line or 'Module:' in line]
        for line in error_lines[:15]:  # Show first 15 error lines
            steps_log.append(f"   {line}")
        if len(error_lines) > 15:
            steps_log.append(f"   ... and {len(error_lines) - 15} more errors")
        steps_log.append("")

    cache_key: Optional[str] = None
    if validation_results and rdf_content:
        cache_key = _make_fix_cache_key(validation_results, rdf_content, template)
        cached = _get_cached_correction(cache_key, steps_log)
        if cached is not None:
            if steps_log:
                steps_log.append("💾 Cache hit! Returning previously successful correction")
            return cached

    # Try rapid fix FIRST - this should handle most cases in < 5 seconds
    if steps_log:
        steps_log.append("=" * 60)
        steps_log.append("🚀 STARTING RAPID FIX")
        steps_log.append("=" * 60)
    
    quick_fix = rapid_fix_missing_properties(rdf_content, validation_results, template, steps_log)
    
    if quick_fix:
        if steps_log:
            steps_log.append("=" * 60)
            steps_log.append("🔍 RE-VALIDATING AFTER RAPID FIX")
            steps_log.append("=" * 60)
    
    if quick_fix and VALIDATOR_AVAILABLE:
        try:
            conforms, new_results = validate_rdf(quick_fix.encode('utf-8'), template)
            if conforms:
                if steps_log:
                    steps_log.append("=" * 60)
                    steps_log.append("✅✅✅ RAPID FIX SUCCESSFUL - VALIDATION PASSED!")
                    steps_log.append("=" * 60)
                if cache_key:
                    _store_correction_in_cache(cache_key, quick_fix, steps_log)
                return quick_fix
            else:
                # Update for next attempt
                if steps_log:
                    steps_log.append("=" * 60)
                    steps_log.append("⚠️ RAPID FIX INCOMPLETE - Still has errors:")
                    steps_log.append("=" * 60)
                    # Show first few errors
                    error_lines = new_results.split('\n')[:10] if new_results else []
                    for line in error_lines:
                        if 'Less than' in line or 'Message:' in line:
                            steps_log.append(f"   {line.strip()}")
                
                validation_results = new_results or validation_results
                rdf_content = quick_fix
                if steps_log:
                    steps_log.append("📋 Continuing to minimal AI correction...")
        except Exception as e:
            if steps_log:
                steps_log.append("=" * 60)
                steps_log.append(f"❌ RAPID FIX VALIDATION ERROR: {e}")
                steps_log.append("=" * 60)
                steps_log.append("📋 Continuing to minimal AI correction...")
    elif quick_fix and steps_log:
        steps_log.append("⚠️ Validator not available, cannot re-validate rapid fix")
    elif steps_log:
        steps_log.append("ℹ️ Rapid fix returned None, moving to AI correction")

    # If rapid fix didn't fully work, try minimal AI correction
    if OPENAI_AVAILABLE and os.getenv('HF_API_KEY'):
        if steps_log:
            steps_log.append("Attempting minimal AI correction...")
        
        corrected = get_ai_correction_minimal(validation_results, rdf_content, max_tokens=1000)
        
        if corrected and corrected != rdf_content and VALIDATOR_AVAILABLE:
            try:
                conforms, new_results = validate_rdf(corrected.encode('utf-8'), template)
                if conforms:
                    if steps_log:
                        steps_log.append("✅ Minimal AI correction successful!")
                    if cache_key:
                        _store_correction_in_cache(cache_key, corrected, steps_log)
                    return corrected
                else:
                    validation_results = new_results or validation_results
                    rdf_content = corrected
                    if steps_log:
                        steps_log.append("Minimal AI correction partial; falling back to full AI...")
            except Exception as e:
                if steps_log:
                    steps_log.append(f"Minimal AI validation error: {e}; falling back...")

    focus_points = extract_error_focus_points(validation_results)
    missing_props = focus_points.get("missing_properties", [])

    if steps_log is not None:
        steps_log.append(f"Targeted fix: detected {len(missing_props)} missing properties")
        if missing_props:
            preview = ", ".join(missing_props[:5])
            if len(missing_props) > 5:
                preview += ", ..."
            steps_log.append(f"Missing list: {preview}")

    working_rdf = rdf_content
    quick_fix_attempted = False

    if missing_props and len(missing_props) <= 5:
        guidance = get_targeted_bibframe_guidance(missing_props, focus_points.get("classes", []))
        if steps_log is not None:
            steps_log.append(f"Retrieved guidance entries: {len(guidance)}")

        import re

        def _inject_snippets(match: re.Match) -> str:
            nonlocal quick_fix_attempted
            opening, inner, closing = match.groups()
            new_bits = []
            for prop in missing_props:
                if f"<bf:{prop}" not in inner:
                    snippet = generate_property_specific_fix(prop, guidance.get(prop))
                    new_bits.append(snippet)
            if not new_bits:
                return match.group(0)
            quick_fix_attempted = True
            if steps_log is not None:
                steps_log.append(f"Injected {len(new_bits)} snippets into {match.group(1).split()[0][1:]}")
            combined = opening + inner
            if not inner.endswith("\n"):
                combined += "\n"
            combined += "    " + "\n    ".join(new_bits) + "\n" + closing
            return combined

        work_pattern = re.compile(r"(<bf:Work[^>]*>)([\s\S]*?)(</bf:Work>)")
        instance_pattern = re.compile(r"(<bf:Instance[^>]*>)([\s\S]*?)(</bf:Instance>)")

        if work_pattern.search(working_rdf):
            working_rdf = work_pattern.sub(_inject_snippets, working_rdf, count=1)
        elif instance_pattern.search(working_rdf):
            working_rdf = instance_pattern.sub(_inject_snippets, working_rdf, count=1)

        if quick_fix_attempted and VALIDATOR_AVAILABLE:
            try:
                conforms, new_results = validate_rdf(working_rdf.encode('utf-8'), template)
                if conforms:
                    if steps_log is not None:
                        steps_log.append("Quick fix succeeded; validation now passes")
                    if cache_key:
                        _store_correction_in_cache(cache_key, working_rdf, steps_log)
                    return working_rdf
                else:
                    if steps_log is not None:
                        steps_log.append("Quick fix incomplete; falling back to AI loop")
                    validation_results = new_results or validation_results
            except Exception as quick_err:
                if steps_log is not None:
                    steps_log.append(f"Quick fix validation error: {quick_err}; using AI fallback")

    if validation_results and working_rdf:
        cache_key = _make_fix_cache_key(validation_results, working_rdf, template)

    return get_ai_correction(
        validation_results,
        working_rdf,
        template,
        max_attempts=max_attempts,
        include_warnings=include_warnings,
        enable_validation_loop=enable_validation_loop,
        cache_key=cache_key,
        steps_log=steps_log,
    )


def generate_manual_suggestions(validation_results: str) -> str:
    """Generate generic, pattern-based suggestions when AI is not available.

    Note: Avoid hardcoding SHACL rules or specific property requirements; rely only on
    patterns present in the validation output text.
    """
    vr_lower = validation_results.lower() if validation_results else ""
    suggestions: List[str] = []

    # Missing/required
    if ("mincount" in vr_lower) or ("missing" in vr_lower) or ("required" in vr_lower):
        suggestions.append("• Some required fields are missing. Add the missing information where indicated.")

    # Too many values
    if ("maxcount" in vr_lower) or ("too many" in vr_lower) or ("more than allowed" in vr_lower):
        suggestions.append("• Some fields have too many values. Keep only the main/one value as required.")

    # Datatype/format issues
    if ("datatype" in vr_lower) or ("type mismatch" in vr_lower) or ("expected" in vr_lower and "datatype" in vr_lower):
        suggestions.append("• Some values are in the wrong format. Use the expected format (e.g., dates like YYYY-MM-DD).")

    # URI/identifier issues
    if ("iri" in vr_lower) or ("uri" in vr_lower) or ("identifier" in vr_lower and "invalid" in vr_lower):
        suggestions.append("• Some identifiers look malformed. Use complete, valid web addresses or proper identifiers.")

    # Namespace/prefix issues
    if ("namespace" in vr_lower) or ("prefix" in vr_lower):
        suggestions.append("• Define all XML namespace prefixes at the top and use them consistently.")

    # XML syntax/structure
    if ("xml" in vr_lower) or ("syntax" in vr_lower) or ("well-formed" in vr_lower):
        suggestions.append("• Fix XML structure issues (unclosed tags, invalid characters, or nesting problems).")

    # Fallback
    if not suggestions:
        suggestions.append("• Review the validation details and update the record where issues are highlighted.")
        suggestions.append("• Follow the selected template; add missing fields and correct formats as needed.")

    suggestions_text = "\n".join(suggestions)

    return f"""
📋 **What needs fixing:**

{suggestions_text}

💡 **Quick tips:**
• Include required fields when noted
• Keep single-value fields to one value
• Use the expected formats (e.g., for dates)
• Declare and use XML namespace prefixes consistently
• Ensure the XML is well‑formed

Need help? Load an example and compare the structure.
"""

def clean_technical_jargon(text: str) -> str:
    """Replace technical RDF/SHACL terms with plain language for end users."""
    if not text:
        return text
    replacements = {
        # RDF/SHACL jargon
        "URIRef": "identifier",
        "URI": "identifier",
        "IRI": "identifier",
        "Literal": "text value",
        "triple": "field entry",
        "graph": "dataset",
        "node": "record",
        "subject": "record",
        "predicate": "field type",
        "object": "value",
        "SHACL": "validation",
        "constraint": "rule",
        "conformance": "compliance",
        "violation": "issue",
        "sh:": "",
        "rdf:": "",
        "rdfs:": "",
        "xsd:": "",
        # Tone softening
        "Error:": "Issue:",
        "Invalid": "Incorrect",
        "Failed": "Did not pass",
        "Missing": "Not found",
    }
    cleaned = text
    for k, v in replacements.items():
        cleaned = cleaned.replace(k, v)
    return cleaned

def parse_shacl_results_for_ai(results_text: str) -> str:
    """Simplify SHACL results into clearer sentences for AI processing.

    Pattern-based only; does not depend on any SHACL rule definitions.
    """
    if not results_text:
        return ""
    import re
    simplified: List[str] = []

    # Generic patterns
    patterns = [
        (re.compile(r"minCount", re.IGNORECASE), "A required field is missing."),
        (re.compile(r"maxCount", re.IGNORECASE), "A field has more values than allowed; only one may be permitted."),
        (re.compile(r"datatype", re.IGNORECASE), "A field has a value in the wrong format."),
        (re.compile(r"iri|uri", re.IGNORECASE), "An identifier looks malformed or incomplete."),
        (re.compile(r"namespace|prefix", re.IGNORECASE), "A namespace prefix is undefined or inconsistent."),
        (re.compile(r"xml|syntax|well-formed", re.IGNORECASE), "The XML structure has an error (e.g., unclosed tag)."),
    ]

    lines = [ln.strip() for ln in results_text.splitlines() if ln.strip()]
    for ln in lines:
        matched = False
        for regex, message in patterns:
            if regex.search(ln):
                simplified.append(message)
                matched = True
                break
        if not matched and ("Constraint Violation" in ln or "Violation" in ln):
            simplified.append("A record rule was not met.")

    # Deduplicate while preserving order
    seen = set()
    unique = []
    for s in simplified:
        if s not in seen:
            unique.append(s)
            seen.add(s)

    return "\n".join(unique) if unique else results_text

def generate_manual_correction_hints(validation_results: str, rdf_content: str) -> str:
    """Generate manual correction hints when AI is not available"""
    return f"""<!-- Manual correction hints based on validation results -->
<!-- Set HF_API_KEY as a Secret in your Space settings for AI-powered corrections -->

{rdf_content}

<!-- 
VALIDATION ISSUES FOUND:
{validation_results[:500]}...

MANUAL CORRECTION STEPS:
1. Add missing namespace declarations
2. Include required properties (rdf:type, etc.)
3. Fix XML syntax errors
4. Ensure proper URI formats
5. Validate data types
-->"""

def extract_xml_from_text(text: str) -> str:
    """Extract RDF/XML from model output that may include extra formatting.

    Looks for the first <rdf:RDF ...> ... </rdf:RDF> block. If not found,
    returns the original text unchanged.
    """
    if not text:
        return text
    import re
    # Try to capture XML block even if fenced in code blocks
    # Use DOTALL to span multiple lines
    pattern = re.compile(r"<rdf:RDF[\s\S]*?</rdf:RDF>", re.IGNORECASE)
    m = pattern.search(text)
    if m:
        return m.group(0)
    # Strip common markdown fences if present
    fenced = re.sub(r"^```[a-zA-Z]*\n|```$", "", text.strip())
    return fenced if fenced else text

def clean_xml_for_validation(xml_text: str) -> str:
    """
    Clean XML text for validation by removing comments and extra formatting.
    
    Args:
        xml_text (str): XML text that may contain comments or formatting
        
    Returns:
        str: Clean XML ready for validation
    """
    import re
    
    if not xml_text:
        return xml_text
    
    # Remove all HTML comments
    cleaned = re.sub(r'<!--.*?-->', '', xml_text, flags=re.DOTALL)
    
    # Remove any leading/trailing whitespace
    cleaned = cleaned.strip()
    
    # If the text starts with "```" code fence, extract content
    if cleaned.startswith("```"):
        try:
            # Extract content between code fences
            parts = cleaned.split("```")
            if len(parts) >= 3:
                # Second part should be the XML content
                cleaned = parts[1]
                # Remove language identifier if present (e.g., "xml")
                if cleaned.startswith("xml"):
                    cleaned = cleaned[3:]
        except:
            pass
    
    return cleaned.strip()

# --- Namespace and wrapper helpers to avoid XML parser errors ---
STANDARD_NAMESPACES = {
    "rdf": "http://www.w3.org/1999/02/22-rdf-syntax-ns#",
    "bf": "http://id.loc.gov/ontologies/bibframe/",
    "rdfs": "http://www.w3.org/2000/01/rdf-schema#",
    "xsd": "http://www.w3.org/2001/XMLSchema#",
}

def _extract_declared_namespaces(xml_text: str) -> dict:
    import re
    decls = {}
    for prefix, uri in re.findall(r"xmlns:([A-Za-z0-9_-]+)=\"([^\"]+)\"", xml_text[:2000]):
        decls[prefix] = uri
    return decls

def _detect_used_prefixes(xml_text: str) -> set:
    import re
    used = set()
    # Tag prefixes like <bf:Work ...> and attribute prefixes like rdf:type="..."
    for m in re.finditer(r"<\s*([A-Za-z0-9_-]+):[A-Za-z0-9_-]+", xml_text):
        used.add(m.group(1))
    for m in re.finditer(r"\s([A-Za-z0-9_-]+):[A-Za-z0-9_-]+=", xml_text):
        used.add(m.group(1))
    return used

def ensure_rdf_wrapper_and_namespaces(xml_text: str, original_text: Optional[str] = None, steps_log: Optional[List[str]] = None) -> str:
    """Ensure the XML has an <rdf:RDF> wrapper and required xmlns declarations for used prefixes.

    - If wrapper exists, add any missing xmlns: declarations for standard, used prefixes.
    - If wrapper is missing, wrap the content and include standard namespaces for used prefixes.
    """
    if not xml_text or not isinstance(xml_text, str):
        return xml_text
    import re

    declared = _extract_declared_namespaces(xml_text)
    if original_text:
        # Merge any declarations present in the original input
        declared.update(_extract_declared_namespaces(original_text))

    used = _detect_used_prefixes(xml_text)
    # Always consider rdf used for wrapper
    used.add("rdf")

    # Only inject namespaces for known standards to avoid guessing
    missing = [p for p in used if p not in declared and p in STANDARD_NAMESPACES]
    added_attrs = " ".join([f"xmlns:{p}=\"{STANDARD_NAMESPACES[p]}\"" for p in missing])

    has_wrapper = bool(re.search(r"<rdf:RDF[^>]*>", xml_text))
    updated = xml_text

    if has_wrapper:
        if added_attrs:
            # Inject before the closing '>' of the first <rdf:RDF ...>
            def _inject(match):
                start_tag = match.group(0)
                if start_tag.endswith('>'):
                    return start_tag[:-1] + ' ' + added_attrs + '>'
                return start_tag + ' ' + added_attrs
            updated = re.sub(r"<rdf:RDF[^>]*>", _inject, updated, count=1)
            if steps_log is not None and missing:
                steps_log.append(f"Injected missing namespace declarations: {', '.join(missing)}")
    else:
        # Build a wrapper with standard namespaces for used prefixes we know
        attrs = [f"xmlns:rdf=\"{STANDARD_NAMESPACES['rdf']}\""]
        for p in used:
            if p == 'rdf':
                continue
            uri = declared.get(p) or STANDARD_NAMESPACES.get(p)
            if uri:
                attrs.append(f"xmlns:{p}=\"{uri}\"")
        wrapper_open = "<rdf:RDF " + " ".join(attrs) + ">\n"
        wrapper_close = "\n</rdf:RDF>"
        updated = wrapper_open + xml_text + wrapper_close
        if steps_log is not None:
            steps_log.append("Wrapped snippet in <rdf:RDF> with standard namespace declarations")

    return updated

def validate_rdf_interface(rdf_content: str, template: str, use_ai: bool = True, include_warnings: bool = False, iterate_until_valid: bool = True, max_attempts: int = 5, show_steps: bool = True):
    """Main validation function for Gradio interface"""
    if not rdf_content.strip():
        return "❌ Error", "No RDF/XML data provided", "", "", "", "", ""

    steps_log: List[str] = []
    
    # Check if validator is available
    if not VALIDATOR_AVAILABLE:
        error_msg = "Validator module is not available. Please check that validator.py is present and all dependencies are installed."
        steps_log.append(f"ERROR: {error_msg}")
        return "❌ Error", error_msg, "", "\n".join(steps_log) if show_steps else "", "", "", ""
    
    # Prepare and validate RDF
    steps_log.append(f"Preparing RDF for validation (original length: {len(rdf_content)} chars)")
    prepped_input = ensure_rdf_wrapper_and_namespaces(rdf_content, steps_log=steps_log if show_steps else None)
    steps_log.append(f"Preprocessed RDF (new length: {len(prepped_input)} chars)")
    
    # Call validation
    steps_log.append(f"Calling validator with template '{template}'")
    result = validate_rdf_tool(prepped_input, template)
    
    if "error" in result:
        steps_log.append(f"Validation error: {result['error']}")
        return f"❌ Error: {result['error']}", "", "", "\n".join(steps_log) if show_steps else "", "", "", ""
    
    status = result["status"]
    results_text = result["results"]
    conforms = result["conforms"]
    
    steps_log.append(f"Initial validation: {'PASSED' if conforms else 'FAILED'} using template '{template}'")
    
    # Log if we got unexpected empty results
    if not results_text or len(results_text.strip()) == 0:
        steps_log.append("WARNING: Validator returned empty results text")
    
    # Filter results if warnings should be excluded
    filtered_results = results_text
    if not include_warnings and "Warning" in results_text:
        # Split results into lines and filter out warnings
        lines = results_text.split('\n')
        filtered_lines = []
        skip_until_next_section = False
        
        for line in lines:
            if "Warning" in line and ("Constraint Violation" in line or "sh:Warning" in line):
                skip_until_next_section = True
            elif "Constraint Violation" in line and "Warning" not in line:
                skip_until_next_section = False
                filtered_lines.append(line)
            elif not skip_until_next_section:
                filtered_lines.append(line)
        
        filtered_results = '\n'.join(filtered_lines)
        if not include_warnings:
            steps_log.append("Filtered out warnings from results")
    
    corrected_status = ""
    corrected_results = ""
    
    if not include_warnings:
        steps_log.append("Configured to ignore warnings in AI processing")
    if iterate_until_valid:
        steps_log.append(f"Iteration enabled with max_attempts={max_attempts}")
    if conforms:
        suggestions = "✅ No issues found! Your RDF/XML is valid according to the selected template."
        corrected_rdf = ""
        corrected_status = "—"
        corrected_results = ""
        steps_log.append("No correction needed; record already conforms")
    else:
        if use_ai:
            # Pass filtered results to AI functions
            suggestions = get_ai_suggestions(filtered_results, rdf_content, include_warnings)
            steps_log.append("Requested AI suggestions for concise guidance")
            corrected_rdf = get_ai_correction_targeted(
                filtered_results,
                rdf_content,
                template,
                max_attempts=max_attempts,
                include_warnings=include_warnings,
                enable_validation_loop=iterate_until_valid,
                steps_log=steps_log,
            )
            # Attempt re-validation of corrected RDF
            try:
                # Clean the corrected output for validation
                corrected_xml = clean_xml_for_validation(corrected_rdf)
                corrected_xml = extract_xml_from_text(corrected_xml)
                corrected_xml = ensure_rdf_wrapper_and_namespaces(corrected_xml, original_text=prepped_input, steps_log=steps_log)
                
                # Debug logging
                steps_log.append(f"Re-validating cleaned RDF ({len(corrected_xml)} chars)")
                if show_steps:
                    # Log first 200 chars of what we're validating
                    preview = corrected_xml[:200] + "..." if len(corrected_xml) > 200 else corrected_xml
                    steps_log.append(f"Preview: {preview}")
                
                reval = validate_rdf_tool(corrected_xml, template)
                if "error" in reval:
                    corrected_status = f"❌ Re-validation Error: {reval['error']}"
                    corrected_results = ""
                    steps_log.append(f"Re-validation failed with error: {reval['error']}")
                else:
                    corrected_status = reval.get("status", "")
                    corrected_results = reval.get("results", "")
                    conforms = reval.get('conforms', False)
                    steps_log.append(f"Re-validation: {corrected_status} - Conforms: {conforms}")
            except Exception as re_ex:
                corrected_status = f"❌ Re-validation Error: {re_ex}"
                corrected_results = ""
                steps_log.append(f"Re-validation error: {re_ex}")
        else:
            suggestions = generate_manual_suggestions(filtered_results)
            corrected_rdf = generate_manual_correction_hints(filtered_results, rdf_content)
            corrected_status = "—"
            corrected_results = ""
            steps_log.append("AI disabled; produced manual suggestions and hints")
    
    steps_text = "\n".join(steps_log) if show_steps else ""
    return status, results_text, suggestions, steps_text, corrected_rdf, corrected_status, corrected_results

def get_rdf_examples(example_type: str = "valid") -> str:
    """
    Retrieve example RDF/XML snippets for testing and learning.
    
    This tool provides sample RDF/XML content that can be used to test
    the validation system or learn proper RDF structure. Examples include
    valid BibFrame Work records, invalid records for testing corrections,
    and BibFrame Instance records.
    
    Args:
        example_type (str): Type of example to retrieve. Options:
            - 'valid': A complete, valid BibFrame Work record
            - 'invalid': An incomplete BibFrame Work with validation errors
            - 'bibframe': A BibFrame Instance record example
    
    Returns:
        str: Complete RDF/XML example content ready for validation testing
    """
    examples = {
        "valid": SAMPLE_VALID_RDF,
        "invalid": SAMPLE_INVALID_RDF,
        "bibframe": '''<?xml version="1.0" encoding="UTF-8"?>
<rdf:RDF xmlns:rdf="http://www.w3.org/1999/02/22-rdf-syntax-ns#"
         xmlns:bf="http://id.loc.gov/ontologies/bibframe/"
         xmlns:rdfs="http://www.w3.org/2000/01/rdf-schema#">
    
    <bf:Instance rdf:about="http://example.org/instance/1">
        <rdf:type rdf:resource="http://id.loc.gov/ontologies/bibframe/Print"/>
        <bf:instanceOf rdf:resource="http://example.org/work/1"/>
        <bf:title>
            <bf:Title>
                <bf:mainTitle>Example Book Title</bf:mainTitle>
            </bf:Title>
        </bf:title>
        <bf:provisionActivity>
            <bf:Publication>
                <bf:date>2024</bf:date>
                <bf:place>
                    <bf:Place>
                        <rdfs:label>New York</rdfs:label>
                    </bf:Place>
                </bf:place>
            </bf:Publication>
        </bf:provisionActivity>
    </bf:Instance>
    
</rdf:RDF>'''
    }
    
    return examples.get(example_type, examples["valid"])

# Create Gradio Interface
def create_interface():
    """Create the main Gradio interface"""
    
    # Check API key status dynamically
    current_api_key = os.getenv('HF_API_KEY', '')
    api_status = "🔑 AI features enabled" if (OPENAI_AVAILABLE and current_api_key) else "⚠️ AI features disabled (set HF_API_KEY)"
    
    with gr.Blocks(
        title="RDF Validation Server with AI",
        theme=gr.themes.Soft(),
        css="""
        .status-box { 
            font-weight: bold; 
            padding: 10px; 
            border-radius: 5px; 
        }
        .header-text {
            text-align: center;
            padding: 20px;
        }
        """
    ) as demo:
        
        # Header
        debug_info = f"""
        Debug Info:
        - OPENAI_AVAILABLE: {OPENAI_AVAILABLE}
        - HF_INFERENCE_AVAILABLE: {HF_INFERENCE_AVAILABLE}
        - HF_API_KEY set: {'Yes' if current_api_key else 'No'}
        - HF_API_KEY length: {len(current_api_key) if current_api_key else 0}
        - HF_ENDPOINT_URL: {HF_ENDPOINT_URL}
        - HF_MODEL: {HF_MODEL}
        """
        
        gr.HTML(f"""
        <div class="header-text">
            <h1>🔍 RDF Validation Server with AI</h1>
            <p>Validate RDF/XML against SHACL schemas with AI-powered suggestions and corrections</p>
            <p><strong>Status:</strong> {api_status}</p>
            <details><summary>Debug Info</summary><pre>{debug_info}</pre></details>
        </div>
        """)
        
        # Main interface
        with gr.Row():
            with gr.Column(scale=1):
                gr.Markdown("### 📝 Input")
                
                rdf_input = gr.Textbox(
                    label="RDF/XML Content",
                    placeholder="Paste your RDF/XML content here...",
                    lines=15,
                    show_copy_button=True
                )
                
                # Keep the main form simple and tuck options into an accordion
                with gr.Accordion("Advanced options", open=False):
                    with gr.Row():
                        template_dropdown = gr.Dropdown(
                            label="Validation Template",
                            choices=["monograph", "custom"],
                            value="monograph",
                            info="Select the SHACL template to validate against"
                        )
                        use_ai_checkbox = gr.Checkbox(
                            label="Use AI Features",
                            value=True,
                            info="Enable AI-powered suggestions and corrections"
                        )
                        include_warnings_checkbox = gr.Checkbox(
                            label="Include Warnings",
                            value=False,
                            info="Include warnings in AI corrections (violations only by default)"
                        )
                    with gr.Row():
                        iterate_checkbox = gr.Checkbox(
                            label="Iterate until valid",
                            value=True,
                            info="Try multiple correction attempts until validation passes or attempts run out"
                        )
                        max_attempts_slider = gr.Slider(
                            label="Max attempts",
                            minimum=1,
                            maximum=3,
                            value=2,
                            step=1,
                            info="Maximum number of correction attempts (2 recommended for speed)"
                        )
                        show_steps_checkbox = gr.Checkbox(
                            label="Show steps",
                            value=False,
                            info="Display step-by-step process (turn on when you want transparency)"
                        )
                
                validate_btn = gr.Button("🔍 Validate RDF", variant="primary", size="lg")
                
                # Examples and controls
                gr.Markdown("### 📚 Examples & Tools")
                
                with gr.Row():
                    example1_btn = gr.Button("✅ Valid RDF Example", variant="secondary")
                    example2_btn = gr.Button("❌ Invalid RDF Example", variant="secondary")
                    clear_btn = gr.Button("🗑️ Clear All", variant="stop")
        
        # Results section
        with gr.Row():
            with gr.Column():
                gr.Markdown("### 📊 Results")
                
                status_output = gr.Textbox(
                    label="Validation Status",
                    interactive=False,
                    lines=1,
                    elem_classes=["status-box"]
                )
                
                results_output = gr.Textbox(
                    label="Detailed Validation Results",
                    interactive=False,
                    lines=8,
                    show_copy_button=True
                )
                
                suggestions_output = gr.Textbox(
                    label="💡 Fix Suggestions",
                    interactive=False,
                    lines=8,
                    show_copy_button=True
                )

                steps_output = gr.Textbox(
                    label="🧭 Correction Steps",
                    interactive=False,
                    lines=10,
                    show_copy_button=True,
                    placeholder="Step-by-step log of how the system derived the corrected XML"
                )
        
        # Corrected RDF section
        with gr.Row():
            with gr.Column():
                gr.Markdown("### 🛠️ AI-Generated Corrections")
                
                corrected_output = gr.Textbox(
                    label="Corrected RDF/XML",
                    interactive=False,
                    lines=15,
                    show_copy_button=True,
                    placeholder="Corrected RDF will appear here after validation..."
                )

                with gr.Row():
                    corrected_status_output = gr.Textbox(
                        label="Re-validation Status (Corrected RDF)",
                        interactive=False,
                        lines=1,
                        elem_classes=["status-box"]
                    )
                    corrected_results_output = gr.Textbox(
                        label="Re-validation Details",
                        interactive=False,
                        lines=6,
                        show_copy_button=True
                    )
        
        # Event handlers
        validate_btn.click(
            fn=validate_rdf_interface,
            inputs=[rdf_input, template_dropdown, use_ai_checkbox, include_warnings_checkbox, iterate_checkbox, max_attempts_slider, show_steps_checkbox],
            outputs=[status_output, results_output, suggestions_output, steps_output, corrected_output, corrected_status_output, corrected_results_output]
        )
        
        # Remove auto-validation to prevent processing loops
        # rdf_input.change(
        #     fn=validate_rdf_interface,
        #     inputs=[rdf_input, template_dropdown, use_ai_checkbox],
        #     outputs=[status_output, results_output, suggestions_output, corrected_output]
        # )
        
        # Example buttons
        example1_btn.click(
            lambda: get_rdf_examples("valid"),
            outputs=[rdf_input]
        )
        
        example2_btn.click(
            lambda: get_rdf_examples("invalid"),
            outputs=[rdf_input]
        )
        
        clear_btn.click(
            lambda: ("", "", "", "", "", "", "", ""),
            outputs=[rdf_input, status_output, results_output, suggestions_output, steps_output, corrected_output, corrected_status_output, corrected_results_output]
        )
        
        # Footer with instructions
        gr.Markdown("""
        ---
        ### � **Documentation & Resources:**
        
        **[📖 MCP4BibFrame Documentation](https://huggingface.co/spaces/jimfhahn/mcp4bibframe-docs)** - Complete BibFrame ontology reference with examples
        
        This validator integrates with the **MCP4BibFrame Documentation API** to provide authoritative BibFrame ontology information during AI-powered corrections.
        
        ### 🚀 **Quick Start:**
        
        1. **Paste your RDF/XML** in the input box above
        2. **Click "Validate RDF"** to check for errors
        3. **Review AI suggestions** for plain-language fixes (enhanced with BibFrame documentation)
        4. **Copy the corrected RDF** from the output
        
        ---
        ### �🚀 **Deployment Instructions for Hugging Face Spaces:**
        
        1. **Create a new Space** on [Hugging Face](https://huggingface.co/spaces)
        2. **Set up your Hugging Face Inference Endpoint** and get the endpoint URL
        3. **Set your tokens** in Space settings (use Secrets for security):
           - Go to Settings → Repository secrets  
           - Add: `HF_API_KEY` = `your_huggingface_api_key_here`
           - Endpoint is now hardcoded to your specific Inference Endpoint
        4. **Upload these files** to your Space repository
        5. **Install requirements**: The Space will auto-install from `requirements.txt`
        
        ### 🔧 **MCP Server Mode:**
        This app functions as both a web interface AND an MCP server for Claude Desktop and other MCP clients.
        
        **Available MCP Tools:**
        - `validate_rdf_tool`: Validate RDF/XML against SHACL shapes
        - `get_ai_suggestions`: Get AI-powered fix suggestions (with BibFrame docs)
        - `get_ai_correction`: Generate corrected RDF/XML (with BibFrame docs)
        - `get_rdf_examples`: Retrieve example RDF snippets
        - `validate_rdf_interface`: Complete validation with AI suggestions and corrections (primary tool)
        
        **MCP Configuration (Streamable HTTP):** 
        Add this configuration to your MCP client (Claude Desktop, etc.):
        
        ```json
        {
          "mcpServers": {
            "rdf-validator": {
              "url": "https://jimfhahn-mcp4rdf.hf.space/gradio_api/mcp/"
            }
          }
        }
        ```
        
        **Alternative SSE Configuration:**
        ```json
        {
          "mcpServers": {
            "rdf-validator": {
              "url": "https://jimfhahn-mcp4rdf.hf.space/gradio_api/mcp/sse"
            }
          }
        }
        ```
        
        ### 💡 **Features:**
        - ✅ Real-time RDF/XML validation against SHACL schemas
        - 🤖 AI-powered error suggestions and corrections (enhanced with BibFrame ontology docs)
        - 📚 Built-in examples and templates
        - 🔗 Integrated with [MCP4BibFrame Documentation API](https://huggingface.co/spaces/jimfhahn/mcp4bibframe-docs)
        - 📋 Copy results with one click
        
        **BibFrame Documentation Integration:** 
        AI corrections now use authoritative BibFrame ontology information from the MCP4BibFrame Documentation API to ensure accuracy and compliance with official specifications.
        
        ### 🔗 **Related Resources:**
        - [MCP4BibFrame Documentation](https://huggingface.co/spaces/jimfhahn/mcp4bibframe-docs) - BibFrame ontology reference
        - [BIG DCTAP Documentation](https://bf-interop.github.io/DCTap/)
        - [BIBFRAME Ontology](http://id.loc.gov/ontologies/bibframe.html)
        - [SHACL Specification](https://www.w3.org/TR/shacl/)
        
        **Note:** AI features require a valid Hugging Face API key (HF_API_KEY) set as a Secret. Manual suggestions are provided as fallback.
        """)
    
    return demo

# Launch configuration
if __name__ == "__main__":
    demo = create_interface()
    
    # Configuration for different environments
    port = int(os.getenv('PORT', 7860))  # Hugging Face uses PORT env variable
    
    demo.launch(
        server_name="0.0.0.0",      # Important for external hosting
        server_port=port,           # Use environment PORT or default to 7860
        share=False,                # Don't create gradio.live links in production
        show_error=True,            # Show errors in the interface
        show_api=True,              # Enable API endpoints
        allowed_paths=["."],        # Allow serving files from current directory
        mcp_server=True             # Enable MCP server functionality (Gradio 5.28+)
    )