abstractcore 2.5.0__py3-none-any.whl → 2.5.3__py3-none-any.whl

abstractcore/architectures/detection.py

@@ -17,6 +17,10 @@ logger = logging.getLogger(__name__)
 _architecture_formats: Optional[Dict[str, Any]] = None
 _model_capabilities: Optional[Dict[str, Any]] = None
 
+# Cache for resolved model names and architectures to reduce redundant logging
+_resolved_aliases_cache: Dict[str, str] = {}
+_detected_architectures_cache: Dict[str, str] = {}
+
 
 def _load_json_assets():
     """Load architecture formats and model capabilities from JSON files."""
@@ -58,9 +62,14 @@ def detect_architecture(model_name: str) -> str:
     Returns:
         Architecture name (e.g., 'qwen', 'llama', 'openai')
     """
+    # Check cache first to avoid redundant logging
+    if model_name in _detected_architectures_cache:
+        return _detected_architectures_cache[model_name]
+    
     _load_json_assets()
 
     if not _architecture_formats or "architectures" not in _architecture_formats:
+        _detected_architectures_cache[model_name] = "generic"
         return "generic"
 
     model_lower = model_name.lower()
@@ -71,10 +80,12 @@ def detect_architecture(model_name: str) -> str:
         for pattern in patterns:
             if pattern.lower() in model_lower:
                 logger.debug(f"Detected architecture '{arch_name}' for model '{model_name}' (pattern: '{pattern}')")
+                _detected_architectures_cache[model_name] = arch_name
                 return arch_name
 
     # Fallback to generic
     logger.debug(f"No specific architecture detected for '{model_name}', using generic")
+    _detected_architectures_cache[model_name] = "generic"
     return "generic"
 
 
@@ -101,6 +112,7 @@ def resolve_model_alias(model_name: str, models: Dict[str, Any]) -> str:
     Resolve a model name to its canonical name by checking aliases.
 
     Automatically converts "--" to "/" for HuggingFace cache format compatibility.
+    Normalizes Claude version numbers (e.g., "claude-3-5-sonnet" -> "claude-3.5-sonnet").
 
     Args:
         model_name: Model name that might be an alias
@@ -109,17 +121,35 @@ def resolve_model_alias(model_name: str, models: Dict[str, Any]) -> str:
     Returns:
         Canonical model name
     """
+    # Check cache first to avoid redundant logging
+    if model_name in _resolved_aliases_cache:
+        return _resolved_aliases_cache[model_name]
+
     # First check if it's already a canonical name
     if model_name in models:
+        _resolved_aliases_cache[model_name] = model_name
         return model_name
 
+    # Normalize model name
+    normalized_model_name = model_name
+
     # Convert "--" to "/" for HuggingFace cache format compatibility
-    normalized_model_name = model_name.replace("--", "/")
+    normalized_model_name = normalized_model_name.replace("--", "/")
+
+    # Normalize Claude version numbers: convert "-X-Y-" to "-X.Y-" or "-X-Y" to "-X.Y"
+    # Examples:
+    #   "claude-3-5-sonnet" -> "claude-3.5-sonnet"
+    #   "claude-4-1-opus" -> "claude-4.1-opus"
+    #   "claude-3-5-sonnet-20241022" -> "claude-3.5-sonnet-20241022"
+    import re
+    normalized_model_name = re.sub(r'(claude-\d+)-(\d+)(?=-|$)', r'\1.\2', normalized_model_name)
+
     if normalized_model_name != model_name:
         logger.debug(f"Normalized model name '{model_name}' to '{normalized_model_name}'")
 
     # Check if normalized name is a canonical name
     if normalized_model_name in models:
+        _resolved_aliases_cache[model_name] = normalized_model_name
         return normalized_model_name
 
     # Check if it's an alias of any model (try both original and normalized)
@@ -127,9 +157,11 @@ def resolve_model_alias(model_name: str, models: Dict[str, Any]) -> str:
         aliases = model_info.get("aliases", [])
         if model_name in aliases or normalized_model_name in aliases:
             logger.debug(f"Resolved alias '{model_name}' to canonical name '{canonical_name}'")
+            _resolved_aliases_cache[model_name] = canonical_name
             return canonical_name
 
     # Return normalized name if no alias found
+    _resolved_aliases_cache[model_name] = normalized_model_name
     return normalized_model_name
 
 
@@ -165,9 +197,10 @@ def get_model_capabilities(model_name: str) -> Dict[str, Any]:
         return capabilities
 
     # Step 3: Try partial matches for common model naming patterns
-    model_lower = model_name.lower()
+    # Use canonical_name (which has been normalized) for better matching
+    canonical_lower = canonical_name.lower()
     for model_key, capabilities in models.items():
-        if model_key.lower() in model_lower or model_lower in model_key.lower():
+        if model_key.lower() in canonical_lower or canonical_lower in model_key.lower():
             result = capabilities.copy()
             # Remove alias-specific fields
             result.pop("canonical_name", None)
@@ -293,4 +326,217 @@ def detect_model_type(model_name: str) -> str:
     elif any(x in model_lower for x in ["vision", "vl", "multimodal"]):
         return "vision"
     else:
-        return "base"
+        return "base"
+
+
+def get_vision_capabilities(model_name: str) -> Dict[str, Any]:
+    """
+    Get vision-specific capabilities for a model with fallback to generic vision model.
+    
+    Args:
+        model_name: Model name to get vision capabilities for
+        
+    Returns:
+        Dictionary with vision capabilities, using generic fallback if model not found
+    """
+    from ..utils.structured_logging import get_logger
+    logger = get_logger(__name__)
+    
+    # Get model capabilities
+    capabilities = get_model_capabilities(model_name)
+    
+    # Check if model has vision support
+    if not capabilities.get('vision_support', False):
+        logger.warning(f"Model '{model_name}' does not have vision support")
+        return {}
+    
+    # Extract vision-specific fields
+    vision_fields = [
+        'image_resolutions', 'max_image_resolution', 'image_patch_size', 
+        'max_image_tokens', 'image_tokenization_method', 'adaptive_resolution',
+        'vision_encoder', 'pixel_grouping', 'supported_resolutions',
+        'base_tokens_per_resolution', 'fixed_resolution', 'tokens_per_tile',
+        'tile_size', 'base_image_tokens', 'pixel_divisor', 'token_cap'
+    ]
+    
+    vision_capabilities = {}
+    for field in vision_fields:
+        if field in capabilities:
+            vision_capabilities[field] = capabilities[field]
+    
+    # If we have minimal vision capabilities, use generic fallback
+    if not vision_capabilities or len(vision_capabilities) < 3:
+        logger.warning(
+            f"Model '{model_name}' has limited vision metadata, using generic vision model fallback",
+            model=model_name,
+            found_fields=list(vision_capabilities.keys())
+        )
+        
+        # Get generic vision model capabilities
+        _load_json_assets()
+        if _model_capabilities and "generic_vision_model" in _model_capabilities:
+            generic_caps = _model_capabilities["generic_vision_model"]
+            for field in vision_fields:
+                if field in generic_caps:
+                    vision_capabilities[field] = generic_caps[field]
+    
+    return vision_capabilities
+
+
+def get_glyph_compression_capabilities(model_name: str) -> Dict[str, Any]:
+    """
+    Get capabilities relevant for Glyph compression with intelligent fallbacks.
+    
+    Args:
+        model_name: Model name to get Glyph capabilities for
+        
+    Returns:
+        Dictionary with Glyph-relevant capabilities and recommendations
+    """
+    from ..utils.structured_logging import get_logger
+    logger = get_logger(__name__)
+    
+    capabilities = get_model_capabilities(model_name)
+    
+    # Check if model supports vision (required for Glyph)
+    if not capabilities.get('vision_support', False):
+        logger.error(
+            f"Model '{model_name}' does not support vision, cannot use Glyph compression",
+            model=model_name
+        )
+        return {
+            'glyph_compatible': False,
+            'reason': 'no_vision_support'
+        }
+    
+    # Get vision capabilities
+    vision_caps = get_vision_capabilities(model_name)
+    
+    # Determine Glyph compatibility and optimal settings
+    glyph_caps = {
+        'glyph_compatible': True,
+        'model_name': model_name,
+        'vision_support': True
+    }
+    
+    # Add vision-specific fields for token calculation
+    glyph_caps.update(vision_caps)
+    
+    # Determine optimal compression settings based on model capabilities
+    max_image_tokens = vision_caps.get('max_image_tokens', 2048)
+    image_patch_size = vision_caps.get('image_patch_size', 16)
+    
+    # Recommend compression parameters
+    if max_image_tokens >= 16000:
+        glyph_caps['recommended_pages_per_image'] = 2
+        glyph_caps['recommended_dpi'] = 150
+    elif max_image_tokens >= 8000:
+        glyph_caps['recommended_pages_per_image'] = 1
+        glyph_caps['recommended_dpi'] = 120
+    else:
+        glyph_caps['recommended_pages_per_image'] = 1
+        glyph_caps['recommended_dpi'] = 100
+    
+    # Check for Glyph-optimized models
+    if capabilities.get('optimized_for_glyph', False):
+        glyph_caps['glyph_optimized'] = True
+        logger.info(f"Model '{model_name}' is optimized for Glyph compression")
+    
+    return glyph_caps
+
+
+def check_vision_model_compatibility(model_name: str, provider: str = None) -> Dict[str, Any]:
+    """
+    Comprehensive check for vision model compatibility with detailed recommendations.
+    
+    Args:
+        model_name: Model name to check
+        provider: Provider name (optional, for provider-specific checks)
+        
+    Returns:
+        Dictionary with compatibility status and recommendations
+    """
+    from ..utils.structured_logging import get_logger
+    logger = get_logger(__name__)
+    
+    result = {
+        'model_name': model_name,
+        'provider': provider,
+        'compatible': False,
+        'vision_support': False,
+        'glyph_compatible': False,
+        'warnings': [],
+        'recommendations': [],
+        'capabilities': {}
+    }
+    
+    # Get model capabilities
+    capabilities = get_model_capabilities(model_name)
+    
+    # Check if this is an unknown model (architecture is 'generic' means it wasn't found in database)
+    is_unknown_model = capabilities.get('architecture') == 'generic' and not capabilities.get('vision_support', False)
+    
+    if is_unknown_model:
+        result['warnings'].append(f"Model '{model_name}' not found in capabilities database")
+        result['recommendations'].append("Add model specifications to model_capabilities.json")
+        result['recommendations'].append("Using generic vision model fallback for VLM calculations")
+        
+        # Use generic fallback - assume vision support for unknown models
+        _load_json_assets()
+        if _model_capabilities and "generic_vision_model" in _model_capabilities:
+            generic_caps = _model_capabilities["generic_vision_model"].copy()
+            result['compatible'] = True
+            result['vision_support'] = True
+            result['capabilities'] = generic_caps
+            
+            # Also get vision capabilities using the generic model
+            vision_caps = generic_caps.copy()
+            result['vision_capabilities'] = vision_caps
+            
+            # Check Glyph compatibility with generic model
+            glyph_caps = {
+                'glyph_compatible': True,
+                'model_name': model_name,
+                'vision_support': True,
+                'recommended_pages_per_image': 1,
+                'recommended_dpi': 100
+            }
+            glyph_caps.update(vision_caps)
+            result['glyph_compatible'] = True
+            result['glyph_capabilities'] = glyph_caps
+            
+            logger.warning(f"Using generic vision model fallback for unknown model '{model_name}'")
+        
+        return result
+    
+    # Check vision support
+    vision_support = capabilities.get('vision_support', False)
+    result['vision_support'] = vision_support
+    result['capabilities'] = capabilities
+    
+    if not vision_support:
+        result['warnings'].append(f"Model '{model_name}' does not support vision")
+        result['recommendations'].append("Use a vision-capable model for image processing")
+        return result
+    
+    result['compatible'] = True
+    
+    # Get vision-specific capabilities
+    vision_caps = get_vision_capabilities(model_name)
+    result['vision_capabilities'] = vision_caps
+    
+    # Check Glyph compatibility
+    glyph_caps = get_glyph_compression_capabilities(model_name)
+    result['glyph_compatible'] = glyph_caps.get('glyph_compatible', False)
+    result['glyph_capabilities'] = glyph_caps
+    
+    # Add specific recommendations based on capabilities
+    if not vision_caps.get('image_patch_size'):
+        result['warnings'].append("No image_patch_size specified, using generic fallback")
+        result['recommendations'].append("Add image_patch_size to model capabilities for better accuracy")
+    
+    if not vision_caps.get('max_image_tokens'):
+        result['warnings'].append("No max_image_tokens specified")
+        result['recommendations'].append("Add max_image_tokens to model capabilities")
+    
+    return result

abstractcore/assets/architecture_formats.json

@@ -308,6 +308,18 @@
       "tool_prefix": "<|tool_call|>",
       "patterns": ["glm-4.5", "glm-4.6", "glm-4.5-air"]
     },
+    "glm4v": {
+      "description": "Zhipu AI's GLM-4V multimodal architecture (June 2024)",
+      "message_format": "glm_special_tokens",
+      "system_prefix": "<|system|>\n",
+      "system_suffix": "\n",
+      "user_prefix": "<|user|>\n",
+      "user_suffix": "\n",
+      "assistant_prefix": "<|assistant|>\n",
+      "assistant_suffix": "\n",
+      "tool_format": "json",
+      "patterns": ["glm-4v", "glm4v", "glyph", "zai-org/glyph", "glm-4.1v"]
+    },
     "glm4": {
       "description": "Zhipu AI's GLM-4 architecture (June 2024)",
       "message_format": "im_start_end",
@@ -407,7 +419,8 @@
     "basic": "Simple role: content format",
     "human_assistant": "Human/Assistant format",
     "openai_chat": "OpenAI chat completion format",
-    "llama3_header": "LLaMA 3+ format with <|start_header_id|> and <|eot_id|>"
+    "llama3_header": "LLaMA 3+ format with <|start_header_id|> and <|eot_id|>",
+    "glm_special_tokens": "GLM format with <|system|>, <|user|>, <|assistant|> tokens"
   },
   "tool_formats": {
     "pythonic": "Python function call syntax: [func(arg=val)]",