import os import time import tempfile import traceback from pathlib import Path from typing import List, Union, Optional, Dict, Any, Literal from fastapi import FastAPI, File, UploadFile, HTTPException, status from pydantic import BaseModel, Field from PIL import Image # --- IMPORTS --- from mdr_pdf_parser import ( MagicPDFProcessor, MDRStructuredBlock, MDRTextBlock, MDRTableBlock, MDRFormulaBlock, MDRFigureBlock, MDRTextKind, MDRTableFormat, MDRRectangle, MDRTextSpan, MDRExtractedTableFormat # For configuration ) # --- Configuration --- # Read from environment variables, falling back to defaults MODEL_DIR = os.environ.get("MDR_MODEL_DIR", "/models") # Default path inside container DEVICE = os.environ.get("MDR_DEVICE", "cuda") # Default to cuda, processor will fallback if needed TABLE_FORMAT_STR = os.environ.get("MDR_TABLE_FORMAT", "MARKDOWN") # Default table format # Convert table format string to Enum try: TABLE_FORMAT = MDRExtractedTableFormat[TABLE_FORMAT_STR.upper()] except KeyError: print(f"Warning: Invalid MDR_TABLE_FORMAT '{TABLE_FORMAT_STR}'. Defaulting to DISABLE.") TABLE_FORMAT = MDRExtractedTableFormat.DISABLE # --- Initialize Processor --- # This happens once when the service starts print("Initializing MagicPDFProcessor...") print(f"Model Directory: {MODEL_DIR}") print(f"Target Device: {DEVICE}") print(f"Table Format: {TABLE_FORMAT.name}") start_time = time.time() try: mdr_processor = MagicPDFProcessor( device=DEVICE, model_dir_path=MODEL_DIR, extract_table_format=TABLE_FORMAT, # Set debug_dir_path=None for production service debug_dir_path=None ) print(f"MagicPDFProcessor initialized successfully ({time.time() - start_time:.2f}s)") except Exception as e: print(f"FATAL ERROR: Failed to initialize MagicPDFProcessor during startup: {e}") print("Service cannot start.") traceback.print_exc() # Optionally exit or raise to prevent FastAPI from starting incorrectly mdr_processor = None # Ensure processor is None if init fails # --- API Models (Pydantic) --- # Define models for API input/output for validation and documentation class MDRPointModel(BaseModel): x: float y: float class MDRRectangleModel(BaseModel): lt: MDRPointModel rt: MDRPointModel lb: MDRPointModel rb: MDRPointModel @classmethod def from_mdr_rectangle(cls, rect: MDRRectangle): return cls( lt=MDRPointModel(x=rect.lt[0], y=rect.lt[1]), rt=MDRPointModel(x=rect.rt[0], y=rect.rt[1]), lb=MDRPointModel(x=rect.lb[0], y=rect.lb[1]), rb=MDRPointModel(x=rect.rb[0], y=rect.rb[1]), ) class MDRTextSpanModel(BaseModel): content: str rank: float rect: MDRRectangleModel @classmethod def from_mdr_text_span(cls, span: MDRTextSpan): return cls( content=span.content, rank=span.rank, rect=MDRRectangleModel.from_mdr_rectangle(span.rect) ) class MDRBasicBlockModel(BaseModel): block_type: str # To distinguish block types in the union rect: MDRRectangleModel texts: List[MDRTextSpanModel] = Field(default_factory=list) # Captions/footnotes font_size: float class MDRTextBlockModel(MDRBasicBlockModel): block_type: Literal["TextBlock"] = "TextBlock" kind: str # Use string representation of enum has_paragraph_indentation: bool last_line_touch_end: bool # Override texts field specifically for TextBlock texts: List[MDRTextSpanModel] # Text content itself @classmethod def from_mdr_text_block(cls, block: MDRTextBlock): return cls( rect=MDRRectangleModel.from_mdr_rectangle(block.rect), texts=[MDRTextSpanModel.from_mdr_text_span(span) for span in block.texts], font_size=block.font_size, kind=block.kind.name, # Convert enum to string name has_paragraph_indentation=block.has_paragraph_indentation, last_line_touch_end=block.last_line_touch_end ) class MDRTableBlockModel(MDRBasicBlockModel): block_type: Literal["TableBlock"] = "TableBlock" content: str format: str # Use string representation of enum # Omit 'image' field from API response @classmethod def from_mdr_table_block(cls, block: MDRTableBlock): return cls( rect=MDRRectangleModel.from_mdr_rectangle(block.rect), texts=[MDRTextSpanModel.from_mdr_text_span(span) for span in block.texts], # Captions font_size=block.font_size, content=block.content, format=block.format.name # Convert enum to string name ) class MDRFormulaBlockModel(MDRBasicBlockModel): block_type: Literal["FormulaBlock"] = "FormulaBlock" content: Optional[str] = None # Omit 'image' field from API response @classmethod def from_mdr_formula_block(cls, block: MDRFormulaBlock): return cls( rect=MDRRectangleModel.from_mdr_rectangle(block.rect), texts=[MDRTextSpanModel.from_mdr_text_span(span) for span in block.texts], # Captions font_size=block.font_size, content=block.content ) class MDRFigureBlockModel(MDRBasicBlockModel): block_type: Literal["FigureBlock"] = "FigureBlock" # Omit 'image' field from API response @classmethod def from_mdr_figure_block(cls, block: MDRFigureBlock): return cls( rect=MDRRectangleModel.from_mdr_rectangle(block.rect), texts=[MDRTextSpanModel.from_mdr_text_span(span) for span in block.texts], # Captions font_size=block.font_size ) # Union type for the response model MDRStructuredBlockModelAPI = Union[MDRTextBlockModel, MDRTableBlockModel, MDRFormulaBlockModel, MDRFigureBlockModel] # Renamed API Union type # --- FastAPI App --- app = FastAPI( title="MagicDataReadiness PDF Processor", description="API service to extract structured content from PDF files.", version="1.0.0" ) @app.on_event("startup") async def startup_event(): if mdr_processor is None: # This prevents the app from starting if initialization failed raise RuntimeError("MagicPDFProcessor failed to initialize. Service cannot start.") print("MagicDataReadiness Service is ready.") @app.get("/health") async def health_check(): """Simple health check endpoint.""" if mdr_processor is None: raise HTTPException(status_code=status.HTTP_503_SERVICE_UNAVAILABLE, detail="Processor not initialized") return {"status": "ok", "message": "MagicPDFProcessor is running."} @app.post("/process-pdf/", response_model=List[MDRStructuredBlockModelAPI], # Use the Union type summary="Process a PDF file", description="Upload a PDF file to extract structured blocks (text, tables, figures, formulas).") async def process_pdf_endpoint(file: UploadFile = File(..., description="The PDF file to process.")): """ Handles PDF file upload, processing, and returns extracted blocks. """ if mdr_processor is None: raise HTTPException(status_code=status.HTTP_503_SERVICE_UNAVAILABLE, detail="Processor not initialized") if not file.filename.lower().endswith(".pdf"): raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid file type. Please upload a PDF.") # Save uploaded file temporarily temp_pdf_path = "" # Initialize path try: # Create a temporary directory if it doesn't exist temp_dir = Path("./temp_uploads") temp_dir.mkdir(exist_ok=True) # Use a temporary file within the directory with tempfile.NamedTemporaryFile(delete=False, suffix=".pdf", dir=temp_dir) as temp_file: content = await file.read() temp_file.write(content) temp_pdf_path = temp_file.name print(f"Received file '{file.filename}', saved temporarily to '{temp_pdf_path}'") except Exception as e: raise HTTPException(status_code=status.HTTP_500_INTERNAL_SERVER_ERROR, detail=f"Failed to save uploaded file: {e}") extracted_blocks_api: List[MDRStructuredBlockModelAPI] = [] start_process_time = time.time() try: print(f"Processing '{temp_pdf_path}'...") # Process the document using the temporary file path # Note: process_document returns a generator, collect all blocks all_blocks = list(mdr_processor.process_document(pdf_input=temp_pdf_path)) print(f"Extracted {len(all_blocks)} raw blocks.") # Convert internal block types to API response models for block in all_blocks: if isinstance(block, MDRTextBlock): extracted_blocks_api.append(MDRTextBlockModel.from_mdr_text_block(block)) elif isinstance(block, MDRTableBlock): extracted_blocks_api.append(MDRTableBlockModel.from_mdr_table_block(block)) elif isinstance(block, MDRFormulaBlock): extracted_blocks_api.append(MDRFormulaBlockModel.from_mdr_formula_block(block)) elif isinstance(block, MDRFigureBlock): extracted_blocks_api.append(MDRFigureBlockModel.from_mdr_figure_block(block)) # Add more elif clauses if there are other block types process_time = time.time() - start_process_time print(f"Processing finished in {process_time:.2f}s. Returning {len(extracted_blocks_api)} blocks.") except Exception as e: print(f"ERROR during PDF processing: {e}") traceback.print_exc() raise HTTPException(status_code=status.HTTP_500_INTERNAL_SERVER_ERROR, detail=f"An error occurred during PDF processing: {e}") finally: # Clean up the temporary file if temp_pdf_path and os.path.exists(temp_pdf_path): try: os.remove(temp_pdf_path) print(f"Cleaned up temporary file: {temp_pdf_path}") except OSError as e: print(f"Warning: Could not remove temporary file {temp_pdf_path}: {e}") return extracted_blocks_api # Optional: Add root endpoint for basic info/docs link @app.get("/") async def read_root(): return { "message": "Welcome to the MagicDataReadiness PDF Processor API!", "docs_url": "/docs", "health_url": "/health" }