use-case-report-extractor/process_files_openai.py

import os
import json
import glob
import pandas as pd
from docx import Document
from docx.document import Document as _Document
from docx.oxml.text.paragraph import CT_P
from docx.oxml.table import CT_Tbl
from docx.table import _Cell, Table
from docx.text.paragraph import Paragraph
from openai import OpenAI
from dotenv import load_dotenv

# Carica le variabili d'ambiente
load_dotenv()

# Configurazione directory
INPUT_DIR = './input'
OUTPUT_DIR = './output'
TEMPLATE_FILE = './template/use_case_template.csv'

# Assicurati che le directory esistano
os.makedirs(OUTPUT_DIR, exist_ok=True)

# Inizializza client OpenAI
client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))

def iter_block_items(parent):
    """
    Yield each paragraph and table child within *parent*, in document order.
    Each returned value is an instance of either Table or Paragraph.
    """
    if isinstance(parent, _Document):
        parent_elm = parent.element.body
    elif isinstance(parent, _Cell):
        parent_elm = parent._tc
    else:
        raise ValueError("something's not right")

    for child in parent_elm.iterchildren():
        if isinstance(child, CT_P):
            yield Paragraph(child, parent)
        elif isinstance(child, CT_Tbl):
            yield Table(child, parent)

def read_docx_chunks(file_path, chunk_size=4000):
    """
    Legge il file .docx e restituisce un generatore di chunk di testo.
    Mantiene l'ordine di paragrafi e tabelle.
    chunk_size: numero approssimativo di caratteri per chunk.
    """
    try:
        doc = Document(file_path)
        current_chunk = []
        current_length = 0
        
        for block in iter_block_items(doc):
            text = ""
            if isinstance(block, Paragraph):
                text = block.text.strip()
                if text:
                    text += "\n"
            elif isinstance(block, Table):
                # Converti tabella in testo (markdown-like o pipe separated)
                for row in block.rows:
                    row_data = [cell.text.strip() for cell in row.cells]
                    text += "| " + " | ".join(row_data) + " |\n"
                text += "\n"
            
            if text:
                current_chunk.append(text)
                current_length += len(text)
            
            # Se superiamo la dimensione del chunk, yieldiamo
            if current_length >= chunk_size:
                yield "".join(current_chunk)
                current_chunk = []
                current_length = 0
        
        # Yield dell'ultimo chunk se presente
        if current_chunk:
            yield "".join(current_chunk)
            
    except Exception as e:
        print(f"Errore lettura {file_path}: {e}")
        yield None

def parse_markdown_table(text):
    """
    Analizza un testo contenente una tabella Markdown e restituisce un DataFrame pandas.
    """
    lines = text.split('\n')
    table_lines = [line.strip() for line in lines if line.strip().startswith('|')]
    
    if len(table_lines) < 2:
        return None
        
    # Trova la riga separatrice (es. |---|---|)
    separator_index = -1
    for i, line in enumerate(table_lines):
        # Rimuovi pipe e spazi per controllare se contiene solo trattini/due punti
        content = line.replace('|', '').replace(':', '').replace('-', '').strip()
        if not content: # Se vuota dopo aver rimosso i caratteri separatori, è una riga separatrice
            separator_index = i
            break
            
    if separator_index <= 0:
        return None
        
    def split_row(row_str):
        # Split semplice per pipe
        parts = row_str.split('|')
        # Rimuovi primo e ultimo elemento se vuoti (dovuti ai pipe esterni)
        if row_str.startswith('|'): parts.pop(0)
        if row_str.endswith('|'): parts.pop(-1)
        return [p.strip() for p in parts]

    headers = split_row(table_lines[separator_index - 1])
    data_rows = table_lines[separator_index + 1:]
    
    data = []
    for row in data_rows:
        cols = split_row(row)
        # Allinea colonne
        if len(cols) == len(headers):
            data.append(cols)
        elif len(cols) < len(headers):
            data.append(cols + [''] * (len(headers) - len(cols)))
        else:
            data.append(cols[:len(headers)])
            
    if not data:
        return None
        
    return pd.DataFrame(data, columns=headers)

def step1_process_pages():
    """Step 1: Processa pagine -> Markdown -> CSV (Algoritmico)"""
    print("\n--- INIZIO STEP 1: Word -> Markdown -> CSV ---")
    word_files = glob.glob(os.path.join(INPUT_DIR, "*.docx"))
    
    if not word_files:
        print("Nessun file .docx trovato in input.")
        return

    # Colonne attese per validazione (opzionale, ma utile per il prompt)
    expected_columns = ["ID", "Descrizione", "Risultato Atteso", "Risultato Ottenuto", "Stato"]

    for file_path in word_files:
        filename = os.path.basename(file_path)
        file_base_name = os.path.splitext(filename)[0]
        print(f"Elaborazione: {filename}...")
        
        # Directory output
        output_subdir = os.path.join(OUTPUT_DIR, file_base_name)
        os.makedirs(output_subdir, exist_ok=True)
        
        chunk_count = 0
        for chunk in read_docx_chunks(file_path):
            if not chunk:
                continue
            
            chunk_count += 1
            print(f"  -> Elaborazione Pagina/Chunk {chunk_count}...")

            system_prompt = f"""
            Sei un esperto QA. Analizza il frammento di documento fornito.
            Estrai TUTTI i dati relativi a Use Case, Test Case, Scenari e Verifiche senza rielaborare il testo, il testo di Use Case, Test Case, Scenari e Verifiche deve rimanere IDENTICO a quello letto.
            
            Output richiesto:
            - Genera una TABELLA MARKDOWN valida.
            - La tabella DEVE avere ESATTAMENTE queste colonne: {', '.join(expected_columns)}.
            - Se un dato non è presente, lascia la cella vuota.
            - Se non trovi NESSUN dato rilevante, rispondi ESATTAMENTE con "NO_DATA".
            - NON aggiungere altro testo prima o dopo la tabella.
            """

            user_prompt = f"Frammento {chunk_count}:\n\n{chunk}"

            try:
                response = client.chat.completions.create(
                    model="gpt-4o",
                    messages=[
                        {"role": "system", "content": system_prompt},
                        {"role": "user", "content": user_prompt}
                    ],
                    temperature=0.1
                )
                
                content = response.choices[0].message.content.strip()
                
                # Pulisci markdown fences
                if content.startswith("```markdown"):
                    content = content.replace("```markdown", "").replace("```", "")
                elif content.startswith("```"):
                    content = content.replace("```", "")
                
                content = content.strip()

                if content and "NO_DATA" not in content:
                    # 1. Salva Markdown
                    md_path = os.path.join(output_subdir, f"chunk_{chunk_count}.md")
                    with open(md_path, 'w', encoding='utf-8') as f:
                        f.write(content)
                    
                    # 2. Converti in CSV algoritmicamente
                    df = parse_markdown_table(content)
                    if df is not None and not df.empty:
                        csv_path = os.path.join(output_subdir, f"chunk_{chunk_count}.csv")
                        df.to_csv(csv_path, index=False, encoding='utf-8')
                        # print(f"    -> Generato CSV: {csv_path}")
                    else:
                        print(f"    -> Warning: Impossibile parsare tabella in chunk {chunk_count}")
                else:
                    pass
                    
            except Exception as e:
                print(f"Errore chunk {chunk_count} di {filename}: {e}")

def step2_aggregate_csvs():
    """Step 2: Aggrega i frammenti CSV in un unico file finale per documento."""
    print("\n--- INIZIO STEP 2: Aggregazione CSV ---")
    
    subdirs = [d for d in glob.glob(os.path.join(OUTPUT_DIR, "*")) if os.path.isdir(d)]
    
    if not subdirs:
        print("Nessuna cartella di frammenti trovata in output.")
        return

    for fragments_dir in subdirs:
        doc_name = os.path.basename(fragments_dir)
        print(f"Aggregazione per: {doc_name}...")
        
        all_files = glob.glob(os.path.join(fragments_dir, "*.csv"))
        try:
            all_files.sort(key=lambda x: int(os.path.basename(x).split('_')[1].split('.')[0]))
        except:
            all_files.sort()

        if not all_files:
            print(f"  -> Nessun frammento CSV trovato in {fragments_dir}")
            continue
            
        combined_df = pd.concat((pd.read_csv(f) for f in all_files), ignore_index=True)
        
        output_csv_path = os.path.join(OUTPUT_DIR, f"{doc_name}.csv")
        combined_df.to_csv(output_csv_path, index=False, encoding='utf-8')
        print(f"  -> Creato CSV finale: {output_csv_path}")

def main():
    step1_process_pages()
    step2_aggregate_csvs()
    print("\nProcesso completato.")

if __name__ == "__main__":
    main()