import pandas as pd import os import re from tools.helper_functions import OUTPUT_FOLDER from sklearn.feature_extraction.text import TfidfVectorizer from sklearn.metrics.pairwise import cosine_similarity from typing import List, Tuple import gradio as gr from gradio import Progress from pathlib import Path from pymupdf import Document from tools.file_conversion import redact_whole_pymupdf_page, convert_annotation_data_to_dataframe import en_core_web_lg nlp = en_core_web_lg.load() similarity_threshold = 0.95 def combine_ocr_output_text(input_files:List[str], output_folder:str=OUTPUT_FOLDER): """ Combines text from multiple CSV files containing page and text columns. Groups text by file and page number, concatenating text within these groups. Args: input_files (list): List of paths to CSV files Returns: pd.DataFrame: Combined dataframe with columns [file, page, text] """ all_data = [] output_files = [] if isinstance(input_files, str): file_paths_list = [input_files] else: file_paths_list = input_files for file in file_paths_list: if isinstance(file, str): file_path = file else: file_path = file.name # Read CSV file df = pd.read_csv(file_path) # Ensure required columns exist if 'page' not in df.columns or 'text' not in df.columns: print(f"Warning: Skipping {file_path} - missing required columns 'page' and 'text'") continue df['text'] = df['text'].fillna('').astype(str) # Group by page and concatenate text grouped = df.groupby('page')['text'].apply(' '.join).reset_index() # Add filename column grouped['file'] = os.path.basename(file_path) all_data.append(grouped) if not all_data: raise ValueError("No valid CSV files were processed") # Combine all dataframes combined_df = pd.concat(all_data, ignore_index=True) # Reorder columns combined_df = combined_df[['file', 'page', 'text']] output_combined_file_path = output_folder + "combined_ocr_output_files.csv" combined_df.to_csv(output_combined_file_path, index=None) output_files.append(output_combined_file_path) return combined_df, output_files def process_data(df:pd.DataFrame, column:str): ''' Clean and stem text columns in a data frame ''' def _clean_text(raw_text): # Remove HTML tags clean = re.sub(r'<.*?>', '', raw_text) clean = ' '.join(clean.split()) # Join the cleaned words back into a string return clean # Function to apply lemmatisation and remove stopwords def _apply_lemmatization(text): doc = nlp(text) # Keep only alphabetic tokens and remove stopwords lemmatized_words = [token.lemma_ for token in doc if token.is_alpha and not token.is_stop] return ' '.join(lemmatized_words) df['text_clean'] = df[column].apply(_clean_text) df['text_clean'] = df['text_clean'].apply(_apply_lemmatization) return df def map_metadata_single_page(similarity_df:pd.DataFrame, metadata_source_df:pd.DataFrame, preview_length:int=200): """Helper to map metadata for single page results.""" metadata_df = metadata_source_df[['file', 'page', 'text']] results_df = similarity_df.merge(metadata_df, left_on='Page1_Index', right_index=True)\ .rename(columns={'file': 'Page1_File', 'page': 'Page1_Page', 'text': 'Page1_Text'}) results_df = results_df.merge(metadata_df, left_on='Page2_Index', right_index=True, suffixes=('_1', '_2'))\ .rename(columns={'file': 'Page2_File', 'page': 'Page2_Page', 'text': 'Page2_Text'}) results_df["Similarity_Score"] = results_df["Similarity_Score"].round(3) final_df = results_df[['Page1_File', 'Page1_Page', 'Page2_File', 'Page2_Page', 'Similarity_Score', 'Page1_Text', 'Page2_Text']] final_df = final_df.sort_values(["Page1_File", "Page1_Page", "Page2_File", "Page2_Page"]) final_df['Page1_Text'] = final_df['Page1_Text'].str[:preview_length] final_df['Page2_Text'] = final_df['Page2_Text'].str[:preview_length] return final_df def map_metadata_subdocument(subdocument_df:pd.DataFrame, metadata_source_df:pd.DataFrame, preview_length:int=200): """Helper to map metadata for subdocument results.""" metadata_df = metadata_source_df[['file', 'page', 'text']] subdocument_df = subdocument_df.merge(metadata_df, left_on='Page1_Start_Index', right_index=True)\ .rename(columns={'file': 'Page1_File', 'page': 'Page1_Start_Page', 'text': 'Page1_Text'}) subdocument_df = subdocument_df.merge(metadata_df[['page']], left_on='Page1_End_Index', right_index=True)\ .rename(columns={'page': 'Page1_End_Page'}) subdocument_df = subdocument_df.merge(metadata_df, left_on='Page2_Start_Index', right_index=True)\ .rename(columns={'file': 'Page2_File', 'page': 'Page2_Start_Page', 'text': 'Page2_Text'}) subdocument_df = subdocument_df.merge(metadata_df[['page']], left_on='Page2_End_Index', right_index=True)\ .rename(columns={'page': 'Page2_End_Page'}) cols = ['Page1_File', 'Page1_Start_Page', 'Page1_End_Page', 'Page2_File', 'Page2_Start_Page', 'Page2_End_Page', 'Match_Length', 'Page1_Text', 'Page2_Text'] # Add Avg_Similarity if it exists (it won't for greedy match unless we add it) if 'Avg_Similarity' in subdocument_df.columns: subdocument_df['Avg_Similarity'] = subdocument_df['Avg_Similarity'].round(3) cols.insert(7, 'Avg_Similarity') final_df = subdocument_df[cols] final_df = final_df.sort_values(['Page1_File', 'Page1_Start_Page', 'Page2_File', 'Page2_Start_Page']) final_df['Page1_Text'] = final_df['Page1_Text'].str[:preview_length] final_df['Page2_Text'] = final_df['Page2_Text'].str[:preview_length] return final_df def save_results_and_redaction_lists(final_df: pd.DataFrame, output_folder: str) -> list: """ Saves the main results DataFrame and generates per-file redaction lists. This function is extracted to be reusable. Args: final_df (pd.DataFrame): The DataFrame containing the final match results. output_folder (str): The folder to save the output files. Returns: list: A list of paths to all generated files. """ output_paths = [] output_folder_path = Path(output_folder) output_folder_path.mkdir(exist_ok=True) if final_df.empty: print("No matches to save.") return [] # 1. Save the main results DataFrame similarity_file_output_path = output_folder_path / 'page_similarity_results.csv' final_df.to_csv(similarity_file_output_path, index=False) output_paths.append(str(similarity_file_output_path)) print(f"Main results saved to {similarity_file_output_path}") # 2. Save per-file redaction lists # Use 'Page2_File' as the source of duplicate content grouping_col = 'Page2_File' if grouping_col not in final_df.columns: print("Warning: 'Page2_File' column not found. Cannot generate redaction lists.") return output_paths for redact_file, group in final_df.groupby(grouping_col): output_file_name_stem = Path(redact_file).stem output_file_path = output_folder_path / f"{output_file_name_stem}_pages_to_redact.csv" all_pages_to_redact = set() is_subdocument_match = 'Page2_Start_Page' in group.columns if is_subdocument_match: for _, row in group.iterrows(): pages_in_range = range(int(row['Page2_Start_Page']), int(row['Page2_End_Page']) + 1) all_pages_to_redact.update(pages_in_range) else: pages = group['Page2_Page'].unique() all_pages_to_redact.update(pages) if all_pages_to_redact: redaction_df = pd.DataFrame(sorted(list(all_pages_to_redact)), columns=['Page_to_Redact']) redaction_df.to_csv(output_file_path, header=False, index=False) output_paths.append(str(output_file_path)) print(f"Redaction list for {redact_file} saved to {output_file_path}") return output_paths def identify_similar_pages( df_combined: pd.DataFrame, similarity_threshold: float = 0.9, min_word_count: int = 10, min_consecutive_pages: int = 1, greedy_match: bool = False, # NEW parameter output_folder: str = OUTPUT_FOLDER, progress=Progress(track_tqdm=True) ) -> Tuple[pd.DataFrame, List[str], pd.DataFrame]: """ Identifies similar pages with three possible strategies: 1. Single Page: If greedy_match=False and min_consecutive_pages=1. 2. Fixed-Length Subdocument: If greedy_match=False and min_consecutive_pages > 1. 3. Greedy Consecutive Match: If greedy_match=True. """ output_paths = [] progress(0.1, desc="Processing and filtering text") df = process_data(df_combined, 'text') df['word_count'] = df['text_clean'].str.split().str.len().fillna(0) original_row_count = len(df) df_filtered = df[df['word_count'] >= min_word_count].copy() df_filtered.reset_index(drop=True, inplace=True) print(f"Filtered out {original_row_count - len(df_filtered)} pages with fewer than {min_word_count} words.") if len(df_filtered) < 2: return pd.DataFrame(), [], df_combined vectorizer = TfidfVectorizer() tfidf_matrix = vectorizer.fit_transform(df_filtered['text_clean']) progress(0.3, desc="Calculating text similarity") similarity_matrix = cosine_similarity(tfidf_matrix, dense_output=False) coo_matrix = similarity_matrix.tocoo() # Create a DataFrame of all individual page pairs above the threshold. # This is the base for all three matching strategies. similar_pages = [ (r, c, v) for r, c, v in zip(coo_matrix.row, coo_matrix.col, coo_matrix.data) if r < c and v >= similarity_threshold ] if not similar_pages: return pd.DataFrame(), [], df_combined base_similarity_df = pd.DataFrame(similar_pages, columns=['Page1_Index', 'Page2_Index', 'Similarity_Score']) progress(0.6, desc="Aggregating results based on matching strategy") if greedy_match: print("Finding matches using greedy consecutive strategy.") # A set of pairs for fast lookups of (page1_idx, page2_idx) valid_pairs_set = set(zip(base_similarity_df['Page1_Index'], base_similarity_df['Page2_Index'])) # Keep track of indices that have been used in a sequence consumed_indices_1 = set() consumed_indices_2 = set() all_sequences = [] # Iterate through all potential starting pairs, sorted for consistent results sorted_pairs = base_similarity_df.sort_values(['Page1_Index', 'Page2_Index']) for _, row in sorted_pairs.iterrows(): start_idx1, start_idx2 = int(row['Page1_Index']), int(row['Page2_Index']) # If this pair has already been consumed by a previous sequence, skip it if start_idx1 in consumed_indices_1 or start_idx2 in consumed_indices_2: continue # This is a new sequence, start expanding it current_sequence = [(start_idx1, start_idx2)] k = 1 while True: next_idx1 = start_idx1 + k next_idx2 = start_idx2 + k # Check if the next pair in the sequence is a valid match if (next_idx1, next_idx2) in valid_pairs_set and \ next_idx1 not in consumed_indices_1 and \ next_idx2 not in consumed_indices_2: current_sequence.append((next_idx1, next_idx2)) k += 1 else: # The sequence has ended break # Record the found sequence and mark all its pages as consumed sequence_indices_1 = [p[0] for p in current_sequence] sequence_indices_2 = [p[1] for p in current_sequence] all_sequences.append({ 'Page1_Start_Index': sequence_indices_1[0], 'Page1_End_Index': sequence_indices_1[-1], 'Page2_Start_Index': sequence_indices_2[0], 'Page2_End_Index': sequence_indices_2[-1], 'Match_Length': len(current_sequence) }) consumed_indices_1.update(sequence_indices_1) consumed_indices_2.update(sequence_indices_2) if not all_sequences: return pd.DataFrame(), [], df_combined subdocument_df = pd.DataFrame(all_sequences) # We can add back the average similarity if needed, but it requires more lookups. # For now, we'll omit it for simplicity in the greedy approach. # ... (The rest is metadata mapping, same as the subdocument case) elif min_consecutive_pages > 1: # --- STRATEGY 2: Fixed-Length Subdocument Matching --- print(f"Finding consecutive page matches (min_consecutive_pages > 1)") similarity_df = base_similarity_df.copy() similarity_df.sort_values(['Page1_Index', 'Page2_Index'], inplace=True) is_consecutive = (similarity_df['Page1_Index'].diff() == 1) & (similarity_df['Page2_Index'].diff() == 1) block_id = is_consecutive.eq(False).cumsum() grouped = similarity_df.groupby(block_id) agg_results = grouped.agg( Page1_Start_Index=('Page1_Index', 'first'), Page2_Start_Index=('Page2_Index', 'first'), Page1_End_Index=('Page1_Index', 'last'), Page2_End_Index=('Page2_Index', 'last'), Match_Length=('Page1_Index', 'size'), Avg_Similarity=('Similarity_Score', 'mean') ).reset_index(drop=True) subdocument_df = agg_results[agg_results['Match_Length'] >= min_consecutive_pages].copy() if subdocument_df.empty: return pd.DataFrame(), [], df_combined else: # --- STRATEGY 1: Single Page Matching --- print(f"Finding single page matches (min_consecutive_pages=1)") final_df = map_metadata_single_page(base_similarity_df, df_filtered) # The rest of the logic (saving files) is handled after this if/else block pass # The final_df is already prepared # --- Map metadata and format output --- # This block now handles the output for both subdocument strategies (2 and 3) if greedy_match or min_consecutive_pages > 1: final_df = map_metadata_subdocument(subdocument_df, df_filtered) progress(0.8, desc="Saving output files") output_paths = save_results_and_redaction_lists(final_df, output_folder) return final_df, output_paths, df_combined # ============================================================================== # GRADIO HELPER FUNCTIONS # ============================================================================== # full_data:pd.DataFrame, def handle_selection_and_preview(evt: gr.SelectData, results_df:pd.DataFrame, full_duplicate_data_by_file: dict): """ This single function handles a user selecting a row. It: 1. Determines the selected row index. 2. Calls the show_page_previews function to get the text data. 3. Returns all the necessary outputs for the UI. """ # If the user deselects, the event might be None. if not evt: return None, None, None # Clear state and both preview panes # 1. Get the selected index selected_index = evt.index[0] # 2. Get the preview data page1_data, page2_data = show_page_previews(full_duplicate_data_by_file, results_df, evt) # 3. Return all three outputs in the correct order return selected_index, page1_data, page2_data def exclude_match(results_df:pd.DataFrame, selected_index:int, output_folder="./output/"): """ Removes a selected row from the results DataFrame, regenerates output files, and clears the text preview panes. """ if selected_index is None: gr.Warning("No match selected. Please click on a row in the table first.") # Return the original dataframe and update=False for the files return results_df, gr.update(), None, None if results_df.empty: gr.Warning("No duplicate page results found, nothing to exclude.") return results_df, gr.update(), None, None # Drop the selected row updated_df = results_df.drop(selected_index).reset_index(drop=True) # Recalculate all output files using the helper function new_output_paths = save_results_and_redaction_lists(updated_df, output_folder) gr.Info(f"Match at row {selected_index} excluded. Output files have been updated.") # Return the updated dataframe, the new file list, and clear the preview panes return updated_df, new_output_paths, None, None def run_duplicate_analysis(files:list[pd.DataFrame], threshold:float, min_words:int, min_consecutive:int, greedy_match:bool, preview_length:int=500, progress=gr.Progress(track_tqdm=True)): """ Wrapper function updated to include the 'greedy_match' boolean. """ if not files: gr.Warning("Please upload files to analyze.") return None, None, None progress(0, desc="Combining input files...") df_combined, _ = combine_ocr_output_text(files) if df_combined.empty: gr.Warning("No data found in the uploaded files.") return None, None, None # Call the main analysis function with the new parameter results_df, output_paths, full_df = identify_similar_pages( df_combined=df_combined, similarity_threshold=threshold, min_word_count=min_words, min_consecutive_pages=int(min_consecutive), greedy_match=greedy_match, progress=progress ) # Clip text to first 200 characters full_df['text'] = full_df['text'].str[:preview_length] # Preprocess full_data (without preview text) for fast access (run once) full_data_by_file = { file: df.sort_values('page').set_index('page') for file, df in full_df.drop(["text_clean"],axis=1).groupby('file') } if results_df.empty: gr.Info(f"No duplicate pages found, no results returned.") return results_df, output_paths, full_data_by_file # full_df, def show_page_previews(full_data_by_file: dict, results_df: pd.DataFrame, evt: gr.SelectData, preview_length:int=500): """ Optimized version using pre-partitioned and indexed full_data. Triggered when a user selects a row in the results DataFrame. """ if not full_data_by_file or results_df is None or not evt: return None, None selected_row = results_df.iloc[evt.index[0], :] is_subdocument_match = 'Page1_Start_Page' in selected_row if is_subdocument_match: file1, start1, end1 = selected_row['Page1_File'], selected_row['Page1_Start_Page'], selected_row['Page1_End_Page'] file2, start2, end2 = selected_row['Page2_File'], selected_row['Page2_Start_Page'], selected_row['Page2_End_Page'] page1_data = full_data_by_file[file1].loc[start1:end1, ['text']].reset_index() page2_data = full_data_by_file[file2].loc[start2:end2, ['text']].reset_index() else: file1, page1 = selected_row['Page1_File'], selected_row['Page1_Page'] file2, page2 = selected_row['Page2_File'], selected_row['Page2_Page'] page1_data = full_data_by_file[file1].loc[[page1], ['text']].reset_index() page2_data = full_data_by_file[file2].loc[[page2], ['text']].reset_index() page1_data['text'] = page1_data['text'].str[:preview_length] page2_data['text'] = page2_data['text'].str[:preview_length] return page1_data[['page', 'text']], page2_data[['page', 'text']] def apply_whole_page_redactions_from_list(duplicate_page_numbers_df:pd.DataFrame, doc_file_name_with_extension_textbox:str, review_file_state:pd.DataFrame, duplicate_output_paths:list[str], pymupdf_doc:object, page_sizes:list[dict], all_existing_annotations:list[dict]): ''' Take a list of suggested whole pages to redact and apply it to review file data currently available from an existing PDF under review ''' # Create a copy of annotations to avoid modifying the original all_annotations = all_existing_annotations.copy() if not pymupdf_doc: print("Warning: No document file currently under review. Please upload a document on the 'Review redactions' tab to apply whole page redactions.") raise Warning("No document file currently under review. Please upload a document on the 'Review redactions' tab to apply whole page redactions.") return review_file_state, all_annotations # Initialize list of pages to redact list_whole_pages_to_redact = [] # Get list of pages to redact from either dataframe or file if not duplicate_page_numbers_df.empty: list_whole_pages_to_redact = duplicate_page_numbers_df.iloc[:, 0].tolist() elif duplicate_output_paths: expected_duplicate_pages_to_redact_name = f"{doc_file_name_with_extension_textbox}" whole_pages_list = pd.DataFrame() # Initialize empty DataFrame for output_file in duplicate_output_paths: # Note: output_file.name might not be available if output_file is just a string path # If it's a Path object or similar, .name is fine. Otherwise, parse from string. file_name_from_path = output_file.split('/')[-1] if isinstance(output_file, str) else output_file.name if expected_duplicate_pages_to_redact_name in file_name_from_path: whole_pages_list = pd.read_csv(output_file, header=None) # Use output_file directly if it's a path break if not whole_pages_list.empty: list_whole_pages_to_redact = whole_pages_list.iloc[:, 0].tolist() # Convert to set to remove duplicates, then back to list list_whole_pages_to_redact = list(set(list_whole_pages_to_redact)) if not list_whole_pages_to_redact: # Assuming gr is defined (e.g., gradio) print("No relevant list of whole pages to redact found, returning inputs.") raise Warning("Warning: No relevant list of whole pages to redact found, returning inputs.") return review_file_state, all_existing_annotations new_annotations = [] # Process each page for redaction for page in list_whole_pages_to_redact: try: page_index = int(page) - 1 if page_index < 0 or page_index >= len(pymupdf_doc): print(f"Page {page} is out of bounds for a document with {len(pymupdf_doc)} pages, skipping.") continue pymupdf_page = pymupdf_doc[page_index] # Find the matching page size dictionary page_size = next((size for size in page_sizes if size["page"] == int(page)), None) if not page_size: print(f"Page {page} not found in page_sizes object, skipping.") continue rect_height = page_size["cropbox_height"] rect_width = page_size["cropbox_width"] image = page_size["image_path"] # This `image` likely represents the page identifier # Create the whole page redaction box annotation_box = redact_whole_pymupdf_page(rect_height, rect_width, pymupdf_page, border=0.005, redact_pdf=False) # Find existing annotation for this image/page current_page_existing_boxes_group = next((annot_group for annot_group in all_annotations if annot_group["image"] == image), None) new_annotation_group = { "image": image, "boxes": [annotation_box] } if current_page_existing_boxes_group: # Check if we already have a whole page redaction for this page if not any(box["label"] == "Whole page" for box in current_page_existing_boxes_group["boxes"]): current_page_existing_boxes_group["boxes"].append(annotation_box) else: # Optional: Print a message if a whole-page redaction already exists for this page print(f"Whole page redaction for page {page} already exists in annotations, skipping addition.") pass else: # Create new annotation entry all_annotations.append(new_annotation_group) new_annotations.append(new_annotation_group) except Exception as e: print(f"Error processing page {page}: {str(e)}") continue # Convert annotations to dataframe and combine with existing review file whole_page_review_file = convert_annotation_data_to_dataframe(new_annotations) # Ensure all required columns are present in both DataFrames before concat # This is a common point of error if DFs have different schemas expected_cols = ['image', 'page', 'label', 'color', 'xmin', 'ymin', 'xmax', 'ymax', 'text', 'id'] for col in expected_cols: if col not in review_file_state.columns: review_file_state[col] = None # Or an appropriate default value if col not in whole_page_review_file.columns: whole_page_review_file[col] = None review_file_out = pd.concat([review_file_state, whole_page_review_file], ignore_index=True) review_file_out = review_file_out.sort_values(by=["page", "ymin", "xmin"]) # --- Remove duplicate entries from the final DataFrame --- dedup_subset_cols = ['page', 'label', 'text', 'id'] # Ensure these columns exist before trying to use them as subset for drop_duplicates if all(col in review_file_out.columns for col in dedup_subset_cols): review_file_out = review_file_out.drop_duplicates( subset=dedup_subset_cols, keep='first' # Keep the first occurrence of a duplicate redaction ) else: print(f"Warning: Not all columns required for de-duplication ({dedup_subset_cols}) are present in review_file_out. Skipping specific de-duplication.") # You might want a fallback or to inspect what's missing review_file_out.to_csv(OUTPUT_FOLDER + "review_file_out_after_whole_page.csv") gr.Info("Successfully created whole page redactions. Go to the 'Review redactions' tab to see them.") return review_file_out, all_annotations