Spaces:

Kevinshh
/

Preformu

Running

Preformu / layers /input_normalizer.py

Upload full project

aecf8ce verified 10 months ago

11.7 kB

"""
Input Normalization Layer.

This module is responsible for parsing and normalizing all forms of user input
into the Canonical JSON Schema (AnalysisRequest). It handles:
- Free-text input parsing
- File content extraction
- Image processing
- SMILES/CAS number extraction

Design Note:
This layer only EXTRACTS and NORMALIZES data. It does NOT make any
scientific judgments or predictions. All interpretation is delegated
to the Prompt Orchestration layer.
"""

import re
from typing import Optional, Tuple, List
from pathlib import Path

from schemas.canonical_schema import (
    AnalysisRequest,
    APIInput,
    ExcipientInput,
    StabilityData,
)
from utils.file_parsers import FileParser
from utils.image_processor import ImageProcessor


class InputNormalizer:
    """
    Normalizes user input into Canonical JSON format.
    
    This class serves as the entry point for all user input processing.
    It coordinates text parsing, file extraction, and image processing
    to produce a standardized AnalysisRequest.
    """
    
    # Regex patterns for chemical notation extraction
    SMILES_PATTERN = re.compile(
        r'(?:SMILES[:\s]*)?'  # Optional "SMILES:" prefix
        r'([A-Za-z0-9@+\-\[\]\(\)\\\/=#$%&*!.:]{10,})'  # SMILES string
    )
    
    CAS_PATTERN = re.compile(
        r'(?:CAS[:\s-]*)?'  # Optional "CAS:" prefix
        r'(\d{2,7}-\d{2}-\d)'  # CAS number format
    )
    
    MOLECULAR_FORMULA_PATTERN = re.compile(
        r'([A-Z][a-z]?\d*(?:[A-Z][a-z]?\d*)*)'  # Chemical formula
    )
    
    MOLECULAR_WEIGHT_PATTERN = re.compile(
        r'(?:MW|分子量|Molecular Weight)[:\s]*(\d+\.?\d*)'
    )
    
    def __init__(self):
        """Initialize the input normalizer."""
        self.file_parser = FileParser()
        self.image_processor = ImageProcessor()
    
    def normalize(
        self,
        text_input: Optional[str] = None,
        file_paths: Optional[List[str]] = None,
        image_paths: Optional[List[str]] = None,
    ) -> AnalysisRequest:
        """
        Normalize all inputs into a single AnalysisRequest.
        
        Args:
            text_input: Free-text description from user
            file_paths: Paths to uploaded documents (Word, Excel, PDF)
            image_paths: Paths to uploaded images (structure diagrams)
            
        Returns:
            AnalysisRequest: Normalized canonical representation
        """
        # Initialize containers
        api_info = APIInput()
        excipients: List[ExcipientInput] = []
        stability_data: Optional[StabilityData] = None
        
        # Process text input
        if text_input:
            api_info, excipients, stability_data = self._parse_text_input(text_input)
        
        # Process files (if any)
        if file_paths:
            file_api, file_excipients, file_stability = self._process_files(file_paths)
            api_info = self._merge_api_info(api_info, file_api)
            excipients.extend(file_excipients)
            if file_stability:
                stability_data = file_stability
        
        # Process images (if any)
        if image_paths:
            for image_path in image_paths:
                # Store image path for potential structure rendering
                if not api_info.structure_image_path:
                    api_info.structure_image_path = image_path
        
        # Build the analysis request
        return AnalysisRequest(
            api=api_info,
            excipients=excipients,
            stability_data=stability_data,
            analysis_focus=self._determine_analysis_focus(api_info, excipients),
        )
    
    def _parse_text_input(
        self, 
        text: str
    ) -> Tuple[APIInput, List[ExcipientInput], Optional[StabilityData]]:
        """
        Parse free-text input to extract API, excipient, and stability info.
        
        This uses pattern matching and heuristics to identify different
        types of information in the user's text.
        """
        api_info = APIInput()
        excipients: List[ExcipientInput] = []
        stability_data: Optional[StabilityData] = None
        
        # Extract SMILES
        smiles_match = self.SMILES_PATTERN.search(text)
        if smiles_match:
            api_info.smiles = smiles_match.group(1)
        
        # Extract CAS number
        cas_match = self.CAS_PATTERN.search(text)
        if cas_match:
            api_info.cas_number = cas_match.group(1)
        
        # Extract molecular weight
        mw_match = self.MOLECULAR_WEIGHT_PATTERN.search(text)
        if mw_match:
            try:
                api_info.molecular_weight = float(mw_match.group(1))
            except ValueError:
                pass
        
        # Parse for excipients (common patterns)
        excipient_patterns = [
            r'(?:辅料|excipient)[:\s]*([^\n,]+)',
            r'(?:和|与|with)\s*([^\n,]+?)(?:的|的相容性|compatibility)',
            r'(?:相容性|compatibility)[^与和]*(?:与|和|with)\s*([^\n,]+)',
        ]
        
        for pattern in excipient_patterns:
            matches = re.findall(pattern, text, re.IGNORECASE)
            for match in matches:
                excipient_name = match.strip()
                if excipient_name and len(excipient_name) > 1:
                    excipients.append(ExcipientInput(name=excipient_name))
        
        # Special handling for common excipient names
        common_excipients = self._extract_common_excipients(text)
        for exc_name in common_excipients:
            if not any(e.name == exc_name for e in excipients):
                excipients.append(ExcipientInput(name=exc_name))
        
        # Parse stability conditions if present
        stability_patterns = [
            r'(\d+[°℃]C?\s*/\s*\d+%\s*RH)',  # e.g., "40°C/75%RH"
            r'(加速|长期|中间)\s*(?:条件|试验)',
        ]
        
        for pattern in stability_patterns:
            match = re.search(pattern, text)
            if match:
                stability_data = StabilityData(
                    conditions=match.group(0),
                    observations=self._extract_stability_observations(text),
                )
                break
        
        # If we have compound info but no drug name, use SMILES as identifier
        if api_info.smiles and not api_info.name:
            api_info.name = f"Compound ({api_info.smiles[:20]}...)"
        
        return api_info, excipients, stability_data
    
    def _extract_common_excipients(self, text: str) -> List[str]:
        """
        Extract commonly known excipient names from text.
        
        This provides a fallback for users who may not use
        explicit "excipient:" labels.
        """
        # Common excipients in both Chinese and English
        excipient_keywords = {
            # Diluents
            "无水磷酸氢钙": "DCP Anhydrous",
            "磷酸氢钙": "Dibasic Calcium Phosphate",
            "DCP": "DCP",
            "乳糖": "Lactose",
            "微晶纤维素": "MCC",
            "MCC": "MCC",
            "淀粉": "Starch",
            "甘露醇": "Mannitol",
            
            # Binders
            "HPMC": "HPMC",
            "羟丙甲纤维素": "HPMC",
            "PVP": "PVP",
            "预胶化淀粉": "Pregelatinized Starch",
            
            # Lubricants
            "硬脂酸镁": "Magnesium Stearate",
            "滑石粉": "Talc",
            
            # Disintegrants
            "交联羧甲纤维素钠": "Croscarmellose Sodium",
            "交联PVP": "Crospovidone",
        }
        
        found = []
        text_lower = text.lower()
        
        for cn_name, en_name in excipient_keywords.items():
            if cn_name.lower() in text_lower or en_name.lower() in text_lower:
                found.append(cn_name)
        
        return found
    
    def _extract_stability_observations(self, text: str) -> Optional[str]:
        """Extract stability test observations from text."""
        # Look for observation-related content
        patterns = [
            r'(?:结果|result|观察)[:\s]*([^\n]+)',
            r'(?:发现|observed)[:\s]*([^\n]+)',
        ]
        
        for pattern in patterns:
            match = re.search(pattern, text, re.IGNORECASE)
            if match:
                return match.group(1).strip()
        
        return None
    
    def _process_files(
        self, 
        file_paths: List[str]
    ) -> Tuple[APIInput, List[ExcipientInput], Optional[StabilityData]]:
        """
        Process uploaded files and extract relevant information.
        
        Delegates to FileParser for actual content extraction.
        """
        api_info = APIInput()
        excipients: List[ExcipientInput] = []
        stability_data: Optional[StabilityData] = None
        
        for file_path in file_paths:
            path = Path(file_path)
            
            if path.suffix.lower() in ['.docx', '.doc']:
                content = self.file_parser.parse_word(file_path)
            elif path.suffix.lower() in ['.xlsx', '.xls']:
                content = self.file_parser.parse_excel(file_path)
            elif path.suffix.lower() == '.pdf':
                content = self.file_parser.parse_pdf(file_path)
            else:
                continue
            
            # Parse the extracted content as text
            if content:
                file_api, file_exc, file_stab = self._parse_text_input(content)
                api_info = self._merge_api_info(api_info, file_api)
                excipients.extend(file_exc)
                if file_stab:
                    stability_data = file_stab
        
        return api_info, excipients, stability_data
    
    def _merge_api_info(self, base: APIInput, new: APIInput) -> APIInput:
        """Merge two APIInput objects, preferring non-None values."""
        return APIInput(
            name=new.name or base.name,
            smiles=new.smiles or base.smiles,
            structure_image_path=new.structure_image_path or base.structure_image_path,
            cas_number=new.cas_number or base.cas_number,
            molecular_formula=new.molecular_formula or base.molecular_formula,
            molecular_weight=new.molecular_weight or base.molecular_weight,
            additional_info=self._merge_text(base.additional_info, new.additional_info),
        )
    
    def _merge_text(self, text1: Optional[str], text2: Optional[str]) -> Optional[str]:
        """Merge two text strings."""
        if text1 and text2:
            return f"{text1}\n{text2}"
        return text1 or text2
    
    def _determine_analysis_focus(
        self, 
        api: APIInput, 
        excipients: List[ExcipientInput]
    ) -> List[str]:
        """
        Determine which analysis dimensions to focus on based on input.
        
        This helps the Prompt Orchestrator prioritize its analysis.
        """
        focus = []
        
        # Always analyze API structure if SMILES is provided
        if api.smiles:
            focus.append("api_structure")
        
        # Always analyze excipients if provided
        if excipients:
            focus.append("excipient_analysis")
            focus.append("compatibility")
        
        return focus if focus else ["api_structure", "excipient_analysis", "compatibility"]

	"""
	Input Normalization Layer.

	This module is responsible for parsing and normalizing all forms of user input
	into the Canonical JSON Schema (AnalysisRequest). It handles:
	- Free-text input parsing
	- File content extraction
	- Image processing
	- SMILES/CAS number extraction

	Design Note:
	This layer only EXTRACTS and NORMALIZES data. It does NOT make any
	scientific judgments or predictions. All interpretation is delegated
	to the Prompt Orchestration layer.
	"""

	import re
	from typing import Optional, Tuple, List
	from pathlib import Path

	from schemas.canonical_schema import (
	AnalysisRequest,
	APIInput,
	ExcipientInput,
	StabilityData,
	)
	from utils.file_parsers import FileParser
	from utils.image_processor import ImageProcessor


	class InputNormalizer:
	"""
	Normalizes user input into Canonical JSON format.

	This class serves as the entry point for all user input processing.
	It coordinates text parsing, file extraction, and image processing
	to produce a standardized AnalysisRequest.
	"""

	# Regex patterns for chemical notation extraction
	SMILES_PATTERN = re.compile(
	r'(?:SMILES[:\s]*)?' # Optional "SMILES:" prefix
	r'([A-Za-z0-9@+\-\[\]\(\)\\\/=#$%&*!.:]{10,})' # SMILES string
	)

	CAS_PATTERN = re.compile(
	r'(?:CAS[:\s-]*)?' # Optional "CAS:" prefix
	r'(\d{2,7}-\d{2}-\d)' # CAS number format
	)

	MOLECULAR_FORMULA_PATTERN = re.compile(
	r'([A-Z][a-z]?\d(?:[A-Z][a-z]?\d)*)' # Chemical formula
	)

	MOLECULAR_WEIGHT_PATTERN = re.compile(
	r'(?:MW\|分子量\|Molecular Weight)[:\s](\d+\.?\d)'
	)

	def __init__(self):
	"""Initialize the input normalizer."""
	self.file_parser = FileParser()
	self.image_processor = ImageProcessor()

	def normalize(
	self,
	text_input: Optional[str] = None,
	file_paths: Optional[List[str]] = None,
	image_paths: Optional[List[str]] = None,
	) -> AnalysisRequest:
	"""
	Normalize all inputs into a single AnalysisRequest.

	Args:
	text_input: Free-text description from user
	file_paths: Paths to uploaded documents (Word, Excel, PDF)
	image_paths: Paths to uploaded images (structure diagrams)

	Returns:
	AnalysisRequest: Normalized canonical representation
	"""
	# Initialize containers
	api_info = APIInput()
	excipients: List[ExcipientInput] = []
	stability_data: Optional[StabilityData] = None

	# Process text input
	if text_input:
	api_info, excipients, stability_data = self._parse_text_input(text_input)

	# Process files (if any)
	if file_paths:
	file_api, file_excipients, file_stability = self._process_files(file_paths)
	api_info = self._merge_api_info(api_info, file_api)
	excipients.extend(file_excipients)
	if file_stability:
	stability_data = file_stability

	# Process images (if any)
	if image_paths:
	for image_path in image_paths:
	# Store image path for potential structure rendering
	if not api_info.structure_image_path:
	api_info.structure_image_path = image_path

	# Build the analysis request
	return AnalysisRequest(
	api=api_info,
	excipients=excipients,
	stability_data=stability_data,
	analysis_focus=self._determine_analysis_focus(api_info, excipients),
	)

	def _parse_text_input(
	self,
	text: str
	) -> Tuple[APIInput, List[ExcipientInput], Optional[StabilityData]]:
	"""
	Parse free-text input to extract API, excipient, and stability info.

	This uses pattern matching and heuristics to identify different
	types of information in the user's text.
	"""
	api_info = APIInput()
	excipients: List[ExcipientInput] = []
	stability_data: Optional[StabilityData] = None

	# Extract SMILES
	smiles_match = self.SMILES_PATTERN.search(text)
	if smiles_match:
	api_info.smiles = smiles_match.group(1)

	# Extract CAS number
	cas_match = self.CAS_PATTERN.search(text)
	if cas_match:
	api_info.cas_number = cas_match.group(1)

	# Extract molecular weight
	mw_match = self.MOLECULAR_WEIGHT_PATTERN.search(text)
	if mw_match:
	try:
	api_info.molecular_weight = float(mw_match.group(1))
	except ValueError:
	pass

	# Parse for excipients (common patterns)
	excipient_patterns = [
	r'(?:辅料\|excipient)[:\s]*([^\n,]+)',
	r'(?:和\|与\|with)\s*([^\n,]+?)(?:的\|的相容性\|compatibility)',
	r'(?:相容性\|compatibility)[^与和](?:与\|和\|with)\s([^\n,]+)',
	]

	for pattern in excipient_patterns:
	matches = re.findall(pattern, text, re.IGNORECASE)
	for match in matches:
	excipient_name = match.strip()
	if excipient_name and len(excipient_name) > 1:
	excipients.append(ExcipientInput(name=excipient_name))

	# Special handling for common excipient names
	common_excipients = self._extract_common_excipients(text)
	for exc_name in common_excipients:
	if not any(e.name == exc_name for e in excipients):
	excipients.append(ExcipientInput(name=exc_name))

	# Parse stability conditions if present
	stability_patterns = [
	r'(\d+[°℃]C?\s/\s\d+%\s*RH)', # e.g., "40°C/75%RH"
	r'(加速\|长期\|中间)\s*(?:条件\|试验)',
	]

	for pattern in stability_patterns:
	match = re.search(pattern, text)
	if match:
	stability_data = StabilityData(
	conditions=match.group(0),
	observations=self._extract_stability_observations(text),
	)
	break

	# If we have compound info but no drug name, use SMILES as identifier
	if api_info.smiles and not api_info.name:
	api_info.name = f"Compound ({api_info.smiles[:20]}...)"

	return api_info, excipients, stability_data

	def _extract_common_excipients(self, text: str) -> List[str]:
	"""
	Extract commonly known excipient names from text.

	This provides a fallback for users who may not use
	explicit "excipient:" labels.
	"""
	# Common excipients in both Chinese and English
	excipient_keywords = {
	# Diluents
	"无水磷酸氢钙": "DCP Anhydrous",
	"磷酸氢钙": "Dibasic Calcium Phosphate",
	"DCP": "DCP",
	"乳糖": "Lactose",
	"微晶纤维素": "MCC",
	"MCC": "MCC",
	"淀粉": "Starch",
	"甘露醇": "Mannitol",

	# Binders
	"HPMC": "HPMC",
	"羟丙甲纤维素": "HPMC",
	"PVP": "PVP",
	"预胶化淀粉": "Pregelatinized Starch",

	# Lubricants
	"硬脂酸镁": "Magnesium Stearate",
	"滑石粉": "Talc",

	# Disintegrants
	"交联羧甲纤维素钠": "Croscarmellose Sodium",
	"交联PVP": "Crospovidone",
	}

	found = []
	text_lower = text.lower()

	for cn_name, en_name in excipient_keywords.items():
	if cn_name.lower() in text_lower or en_name.lower() in text_lower:
	found.append(cn_name)

	return found

	def _extract_stability_observations(self, text: str) -> Optional[str]:
	"""Extract stability test observations from text."""
	# Look for observation-related content
	patterns = [
	r'(?:结果\|result\|观察)[:\s]*([^\n]+)',
	r'(?:发现\|observed)[:\s]*([^\n]+)',
	]

	for pattern in patterns:
	match = re.search(pattern, text, re.IGNORECASE)
	if match:
	return match.group(1).strip()

	return None

	def _process_files(
	self,
	file_paths: List[str]
	) -> Tuple[APIInput, List[ExcipientInput], Optional[StabilityData]]:
	"""
	Process uploaded files and extract relevant information.

	Delegates to FileParser for actual content extraction.
	"""
	api_info = APIInput()
	excipients: List[ExcipientInput] = []
	stability_data: Optional[StabilityData] = None

	for file_path in file_paths:
	path = Path(file_path)

	if path.suffix.lower() in ['.docx', '.doc']:
	content = self.file_parser.parse_word(file_path)
	elif path.suffix.lower() in ['.xlsx', '.xls']:
	content = self.file_parser.parse_excel(file_path)
	elif path.suffix.lower() == '.pdf':
	content = self.file_parser.parse_pdf(file_path)
	else:
	continue

	# Parse the extracted content as text
	if content:
	file_api, file_exc, file_stab = self._parse_text_input(content)
	api_info = self._merge_api_info(api_info, file_api)
	excipients.extend(file_exc)
	if file_stab:
	stability_data = file_stab

	return api_info, excipients, stability_data

	def _merge_api_info(self, base: APIInput, new: APIInput) -> APIInput:
	"""Merge two APIInput objects, preferring non-None values."""
	return APIInput(
	name=new.name or base.name,
	smiles=new.smiles or base.smiles,
	structure_image_path=new.structure_image_path or base.structure_image_path,
	cas_number=new.cas_number or base.cas_number,
	molecular_formula=new.molecular_formula or base.molecular_formula,
	molecular_weight=new.molecular_weight or base.molecular_weight,
	additional_info=self._merge_text(base.additional_info, new.additional_info),
	)

	def _merge_text(self, text1: Optional[str], text2: Optional[str]) -> Optional[str]:
	"""Merge two text strings."""
	if text1 and text2:
	return f"{text1}\n{text2}"
	return text1 or text2

	def _determine_analysis_focus(
	self,
	api: APIInput,
	excipients: List[ExcipientInput]
	) -> List[str]:
	"""
	Determine which analysis dimensions to focus on based on input.

	This helps the Prompt Orchestrator prioritize its analysis.
	"""
	focus = []

	# Always analyze API structure if SMILES is provided
	if api.smiles:
	focus.append("api_structure")

	# Always analyze excipients if provided
	if excipients:
	focus.append("excipient_analysis")
	focus.append("compatibility")

	return focus if focus else ["api_structure", "excipient_analysis", "compatibility"]