#!/usr/bin/env python3 """ Kas Engine v3 — Native PDF analyse via Gemini met paginabereiken. Splits grote PDF's in stukken van max 50 pagina's, analyseer elk stuk, en syntheseer tot één eindrapport. """ import json, sys, os, requests, base64, time import sqlite3, fitz from datetime import datetime DB_PATH = os.path.join(os.path.dirname(__file__), 'hseq_kennisbank.db') GEMINI_API_KEY = os.environ.get('GEMINI_API_KEY') GEMINI_URL = 'https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:generateContent' MAX_PAGES_PER_CHUNK = 50 file_paths = json.loads(sys.argv[1]) context = sys.argv[2] if len(sys.argv) > 2 else '' language = sys.argv[3] if len(sys.argv) > 3 else 'NL' analysis_id = int(sys.argv[4]) if len(sys.argv) > 4 else 0 def update_db(status, result_text=None, error_msg=None): conn = sqlite3.connect(DB_PATH) c = conn.cursor() if error_msg: c.execute("UPDATE scanner_analyses SET status=?, error_msg=? WHERE id=?", (status, error_msg, analysis_id)) elif result_text: c.execute("UPDATE scanner_analyses SET status=?, result_json=?, completed_at=? WHERE id=?", (status, json.dumps({'text': result_text}), datetime.now().isoformat(), analysis_id)) else: c.execute("UPDATE scanner_analyses SET status=? WHERE id=?", (status, analysis_id)) conn.commit() conn.close() def gemini_call(prompt, pdf_base64, max_retries=3): """Enkele Gemini API call met retry.""" payload = { 'contents': [{ 'parts': [ {'text': prompt}, {'inline_data': {'mime_type': 'application/pdf', 'data': pdf_base64}} ] }], 'generationConfig': { 'temperature': 0.3, 'maxOutputTokens': 16384 } } for attempt in range(max_retries): try: resp = requests.post( f'{GEMINI_URL}?key={GEMINI_API_KEY}', json=payload, timeout=300 ) if resp.status_code == 429: wait = 30 * (attempt + 1) print(f" Rate limited, wacht {wait}s (poging {attempt+1})...") time.sleep(wait) continue if resp.status_code == 400: error_detail = resp.json() msg = error_detail.get('error', {}).get('message', str(error_detail))[:300] return None, f"Gemini 400: {msg}" resp.raise_for_status() result = resp.json() if 'candidates' in result: return result['candidates'][0]['content']['parts'][0]['text'], None else: return None, f"Gemini antwoord onverwacht: {str(result)[:300]}" except requests.exceptions.Timeout: return None, "Gemini time-out (5 min)" except Exception as e: if attempt < max_retries - 1: time.sleep(10) continue return None, f"Gemini fout: {str(e)[:300]}" return None, f"Gemini na {max_retries} pogingen niet beschikbaar" def split_pdf(filepath, max_pages=MAX_PAGES_PER_CHUNK): """Splits PDF in stukken van max_pages pagina's. Retourneert lijst van (chunk_pdf_path, page_range_str).""" doc = fitz.open(filepath) total = len(doc) chunks = [] for start in range(0, total, max_pages): end = min(start + max_pages - 1, total - 1) chunk_doc = fitz.open() chunk_doc.insert_pdf(doc, from_page=start, to_page=end) chunk_path = f'/tmp/scan_chunk_{analysis_id}_{start}.pdf' chunk_doc.save(chunk_path) chunk_doc.close() chunks.append((chunk_path, f'pagina {start+1}-{end+1}')) doc.close() return chunks def main(): if not file_paths: update_db('error', error_msg='Geen bestanden ontvangen') return filepath, filename = file_paths[0] if not os.path.isfile(filepath): update_db('error', error_msg=f'Bestand niet gevonden: {filename}') return if not GEMINI_API_KEY: update_db('error', error_msg='GEMINI_API_KEY niet geconfigureerd') return lang_label = 'Nederlands' if language == 'NL' else 'English' # Stap 1: PDF splitsen in chunks print(f"Splitsen {filename}...") chunks = split_pdf(filepath) total = len(chunks) print(f"{total} chunks van max {MAX_PAGES_PER_CHUNK} pagina's") # Stap 2: Elk chunk analyseren chunk_analyses = [] for i, (chunk_path, page_range) in enumerate(chunks): print(f"Analyse chunk {i+1}/{total} ({page_range})...") with open(chunk_path, 'rb') as f: pdf_b64 = base64.b64encode(f.read()).decode() print(f" Base64: {len(pdf_b64)/1024/1024:.1f} MB") chunk_prompt = ( f"Dit is DEEL {i+1}/{total} van document '{filename}' ({page_range}).\n\n" "Analyseer dit deel gedetailleerd:\n" "- Welke hoofdstukken/secties komen aan bod?\n" "- Welke HSEQ-aspecten (veiligheid, compliance, risico, milieu, kwaliteit)?\n" "- Specifieke eisen, normen, waarden, scenario's?\n" "- Tabellen, formules, berekeningen?\n" "- Actiepunten of verplichtingen?\n\n" "Wees feitelijk en gedetailleerd. Citeer specifieke waarden en paragrafen." ) result, error = gemini_call(chunk_prompt, pdf_b64) # Cleanup chunk file try: os.remove(chunk_path) except: pass if error: update_db('error', error_msg=f"Chunk {i+1} ({page_range}): {error}") print(f" FOUT: {error}") return chunk_analyses.append(f"=== DEEL {i+1}/{total} ({page_range}) ===\n{result}") print(f" OK: {len(result)} tekens") # Stap 3: Synthese tot eindrapport print("Synthese...") all_text = '\n\n'.join(chunk_analyses) synthese_prompt = ( f"Je hebt {total} delen van document '{filename}' geanalyseerd.\n" f"Hier zijn de deelanalyses:\n\n{all_text}\n\n" "Maak nu EEN samenvoegend eindrapport:\n\n" "## 1. Leeswijzer\nDocument, scope, doelgroep, wettelijk kader.\n\n" "## 2. Executive Summary\nKernboodschap in max 250 woorden.\n\n" "## 3. Documentstructuur & Kerninhoud\nAlle hoofdstukken met bevindingen.\n\n" "## 4. HSEQ Highlights Matrix\nMinimaal 10 rijen:\n" "| Thema | Risico | Compliance-eis | Actiepunt | Prioriteit |\n\n" "## 5. Normen, Waarden & Scenario's\n\n" "## 6. Synergieen & Hiaten\n\n" "## 7. Actielijst voor Inrichtinghouder\nGenummerd, met prioriteit.\n\n" "Gebruik ALLE deelanalyses. Specifiek met hoofdstuk/paragraaf verwijzingen. " "HSEQ-terminologie. Feitelijk. Geen AI-cliches." ) final_result, error = gemini_call(synthese_prompt, "") if error: # Fallback: geef de ruwe deelanalyses final_result = all_text update_db('completed', result_text=final_result) print(f"Analyse {analysis_id} voltooid: {len(final_result)} tekens") if __name__ == '__main__': main()