""" Ollama Integration Service for BMC Hub Handles supplier invoice extraction using Ollama LLM with CVR matching """ import json import hashlib import logging import os from pathlib import Path from typing import Optional, Dict, List, Tuple from datetime import datetime import re from app.core.config import settings from app.core.database import execute_insert, execute_query, execute_update, execute_query_single logger = logging.getLogger(__name__) class OllamaService: """Service for extracting supplier invoice data using Ollama LLM""" def __init__(self): self.endpoint = settings.OLLAMA_ENDPOINT self.model = settings.OLLAMA_MODEL self.system_prompt = self._build_system_prompt() logger.info(f"🤖 Initialized OllamaService: {self.endpoint}, model={self.model}") def _build_system_prompt(self) -> str: """Build Danish system prompt for invoice extraction with CVR""" own_cvr = getattr(settings, 'OWN_CVR', '29522790') # BMC har to CVR numre – begge er VORES (køber), aldrig leverandør own_cvr_rule = ( f"4b. KRITISK - LEVERANDØR vs. MODTAGER:\n" f" - På en dansk faktura er LEVERANDØREN (vendor) det firma der HAR SENDT fakturaen.\n" f" De kendes på: firmalogo øverst, bankkonto/IBAN/Gironr. nedad, ingen 'Faktureres til' label.\n" f" - MODTAGEREN (os, buyer) kendes på: navnes under 'Faktureres til', 'Att.', 'Kundenr.', adresseblok med vores navn.\n" f" - BMC DENMARK APS og alle varianter af 'BMC' er ALDRIG leverandøren – det er os (modtageren).\n" f" - CVR {own_cvr} er VORES eget CVR. Sæt ALDRIG vendor_cvr til {own_cvr}.\n" f" - CVR 14416285 er også VORES CVR. Sæt ALDRIG vendor_cvr til 14416285.\n" f" - Ignorer 'SE/CVR-nr.' der hører til modtager-blokken – brug KUN afsenderens CVR som vendor_cvr.\n" ) return ("""Du er en ekspert i at læse og udtrække strukturerede data fra danske fakturaer, kreditnotaer og leverandørdokumenter. VIGTIGE REGLER: 1. Returner KUN gyldig JSON - ingen forklaring eller ekstra tekst 2. Hvis et felt ikke findes, sæt det til null 3. Beregn confidence baseret på hvor sikker du er på hvert felt (0.0-1.0) 4. Datoer skal være i format YYYY-MM-DD """ + own_cvr_rule + """5. DANSKE PRISFORMATER: - Tusind-separator kan være . (punkt) eller mellemrum: "5.965,18" eller "5 965,18" - Decimal-separator er , (komma): "1.234,56 kr" - I JSON output skal du bruge . (punkt) som decimal: 1234.56 - Eksempel: "5.965,18 kr" → 5965.18 i JSON - Eksempel: "1.234,56 DKK" → 1234.56 i JSON 6. CVR-nummer skal være 8 cifre uden mellemrum 7. Moms/VAT skal udtrækkes fra hver linje hvis muligt 8. DOKUMENTTYPE DETEKTION: - "invoice" = Almindelig faktura - "credit_note" = Kreditnota (refusion, tilbagebetaling, korrektion) - Kig efter ord som: "Kreditnota", "Credit Note", "Refusion", "Tilbagebetaling", "Godtgørelse" 9. BELØB OG FORTEGN (ABSOLUT KRITISK): - **ALMINDELIGE FAKTURAER**: Alle beløb skal være POSITIVE tal (total_amount > 0, line_total > 0) - **KREDITNOTAER**: Alle beløb skal være NEGATIVE tal (total_amount < 0, line_total < 0) - Hvis dokumentet siger "Faktura" → document_type: "invoice" → POSITIVE beløb - Hvis dokumentet siger "Kreditnota" → document_type: "credit_note" → NEGATIVE beløb JSON format skal være: { "document_type": "invoice" eller "credit_note", "invoice_number": "fakturanummer eller kreditnota nummer", "vendor_name": "leverandør firmanavn", "vendor_cvr": "12345678", "invoice_date": "YYYY-MM-DD", "due_date": "YYYY-MM-DD", "currency": "DKK", "total_amount": 1234.56 (NEGATIVT for kreditnotaer), "vat_amount": 123.45 (NEGATIVT for kreditnotaer), "original_invoice_reference": "reference til original faktura (kun for kreditnotaer)", "lines": [ { "line_number": 1, "description": "beskrivelse af varen/ydelsen", "quantity": antal_som_tal, "unit_price": pris_per_stk (NEGATIVT for kreditnotaer), "line_total": total_for_linjen (NEGATIVT for kreditnotaer), "vat_rate": 25.00, "vat_amount": moms_beløb (NEGATIVT for kreditnotaer), "confidence": 0.0_til_1.0 } ], "confidence": gennemsnits_confidence, "raw_text_snippet": "første 200 tegn fra dokumentet" } EKSEMPEL PÅ FAKTURA (POSITIVE BELØB): Input: "FAKTURA 2025-001\\nGlobalConnect A/S\\nCVR: 12345678\\n1 stk iPhone 16 @ 5.965,18 DKK\\nMoms (25%): 1.491,30 DKK\\nTotal: 7.456,48 DKK" Output: { "document_type": "invoice", "invoice_number": "2025-001", "vendor_name": "GlobalConnect A/S", "vendor_cvr": "12345678", "total_amount": 7456.48, "vat_amount": 1491.30, "lines": [{ "line_number": 1, "description": "iPhone 16", "quantity": 1, "unit_price": 5965.18, "line_total": 5965.18, "vat_rate": 25.00, "vat_amount": 1491.30, "confidence": 0.95 }], "confidence": 0.95 } EKSEMPEL PÅ KREDITNOTA (NEGATIVE BELØB): Input: "KREDITNOTA CN-2025-042\\nGlobalConnect A/S\\nCVR: 12345678\\nReference: Faktura 2025-001\\nTilbagebetaling:\\n1 stk iPhone 16 returneret @ -5.965,18 DKK\\nMoms (25%): -1.491,30 DKK\\nTotal: -7.456,48 DKK" Output: { "document_type": "credit_note", "invoice_number": "CN-2025-042", "vendor_name": "GlobalConnect A/S", "vendor_cvr": "12345678", "original_invoice_reference": "2025-001", "total_amount": -7456.48, "vat_amount": -1491.30, "lines": [{ "line_number": 1, "description": "iPhone 16 returneret", "quantity": 1, "unit_price": -5965.18, "line_total": -5965.18, "vat_rate": 25.00, "vat_amount": -1491.30, "confidence": 0.95 }], "confidence": 0.95 }""") async def extract_from_text(self, text: str) -> Dict: """ Extract structured invoice data from text using Ollama Args: text: Document text content Returns: Extracted data as dict with CVR, invoice number, amounts, etc. """ # No truncation - send full text to AI prompt = f"{self.system_prompt}\n\nNU SKAL DU UDTRÆKKE DATA FRA DENNE FAKTURA:\n{text}\n\nReturner kun gyldig JSON:" logger.info(f"🤖 Extracting invoice data from text (length: {len(text)})") try: import httpx # Detect if using qwen3 model (requires Chat API) use_chat_api = self.model.startswith('qwen3') async with httpx.AsyncClient(timeout=1000.0) as client: if use_chat_api: # qwen3 models use Chat API format logger.info(f"🤖 Using Chat API for {self.model}") response = await client.post( f"{self.endpoint}/api/chat", json={ "model": self.model, "messages": [ { "role": "system", "content": self.system_prompt }, { "role": "user", "content": f"NU SKAL DU UDTRÆKKE DATA FRA DENNE FAKTURA:\n{text}\n\nVIGTIGT: Dit svar skal STARTE med {{ og SLUTTE med }} - ingen forklaring før eller efter JSON!" } ], "stream": False, "format": "json", "think": False, "options": { "temperature": 0.1, "top_p": 0.9, "num_predict": 8000 } } ) else: # qwen2.5 and other models use Generate API format logger.info(f"🤖 Using Generate API for {self.model}") response = await client.post( f"{self.endpoint}/api/generate", json={ "model": self.model, "prompt": prompt, "stream": False, "options": { "temperature": 0.1, "top_p": 0.9, "num_predict": 8000 } } ) if response.status_code != 200: raise Exception(f"Ollama returned status {response.status_code}: {response.text}") result = response.json() # Extract response based on API type if use_chat_api: # qwen3 models sometimes put the actual response in "thinking" field raw_response = result.get("message", {}).get("content", "") thinking = result.get("message", {}).get("thinking", "") # If content is empty but thinking has data, try to extract JSON from thinking if not raw_response and thinking: logger.info(f"💭 Content empty, attempting to extract JSON from thinking field (length: {len(thinking)})") # Try to find JSON block in thinking text json_start = thinking.find('{') json_end = thinking.rfind('}') + 1 if json_start >= 0 and json_end > json_start: potential_json = thinking[json_start:json_end] logger.info(f"📦 Found potential JSON in thinking field (length: {len(potential_json)})") raw_response = potential_json else: logger.warning(f"⚠️ No JSON found in thinking field, using full thinking as fallback") raw_response = thinking elif thinking: logger.info(f"💭 Model thinking (length: {len(thinking)})") # DEBUG: Log full result structure logger.info(f"📊 Chat API result keys: {list(result.keys())}") logger.info(f"📊 Message keys: {list(result.get('message', {}).keys())}") else: raw_response = result.get("response", "") logger.info(f"✅ Ollama extraction completed (response length: {len(raw_response)})") # Parse JSON from response extraction = self._parse_json_response(raw_response) # CRITICAL: Fix amount signs based on document_type # LLM sometimes returns negative amounts for invoices - fix this! document_type = extraction.get('document_type', 'invoice') if document_type == 'invoice': # Normal invoices should have POSITIVE amounts if extraction.get('total_amount') and extraction['total_amount'] < 0: logger.warning(f"⚠️ Fixing negative total_amount for invoice: {extraction['total_amount']} → {abs(extraction['total_amount'])}") extraction['total_amount'] = abs(extraction['total_amount']) if extraction.get('vat_amount') and extraction['vat_amount'] < 0: extraction['vat_amount'] = abs(extraction['vat_amount']) # Fix line totals if 'lines' in extraction: for line in extraction['lines']: if line.get('unit_price') and line['unit_price'] < 0: line['unit_price'] = abs(line['unit_price']) if line.get('line_total') and line['line_total'] < 0: line['line_total'] = abs(line['line_total']) if line.get('vat_amount') and line['vat_amount'] < 0: line['vat_amount'] = abs(line['vat_amount']) elif document_type == 'credit_note': # Credit notes should have NEGATIVE amounts if extraction.get('total_amount') and extraction['total_amount'] > 0: logger.warning(f"⚠️ Fixing positive total_amount for credit_note: {extraction['total_amount']} → {-abs(extraction['total_amount'])}") extraction['total_amount'] = -abs(extraction['total_amount']) if extraction.get('vat_amount') and extraction['vat_amount'] > 0: extraction['vat_amount'] = -abs(extraction['vat_amount']) # Fix line totals if 'lines' in extraction: for line in extraction['lines']: if line.get('unit_price') and line['unit_price'] > 0: line['unit_price'] = -abs(line['unit_price']) if line.get('line_total') and line['line_total'] > 0: line['line_total'] = -abs(line['line_total']) if line.get('vat_amount') and line['vat_amount'] > 0: line['vat_amount'] = -abs(line['vat_amount']) # Add raw response for debugging extraction['_raw_llm_response'] = raw_response return extraction except Exception as e: error_msg = f"Ollama extraction failed: {str(e)}" logger.error(f"❌ {error_msg}") error_str = str(e).lower() if "timeout" in error_str: return { "error": f"Ollama timeout efter 1000 sekunder", "confidence": 0.0 } elif "connection" in error_str or "connect" in error_str: return { "error": f"Kan ikke forbinde til Ollama på {self.endpoint}", "confidence": 0.0 } else: return { "error": error_msg, "confidence": 0.0 } def _parse_json_response(self, response: str) -> Dict: """Parse JSON from LLM response with aggressive fallback strategies""" logger.info(f"🔍 Response length: {len(response)}, preview: {response[:200]}") # Find outermost JSON object start = response.find('{') end = response.rfind('}') + 1 if start < 0 or end <= start: logger.error("❌ No JSON object found in response") return self._extract_fields_with_regex(response) json_str = response[start:end] # Strategy 1: direct parse try: return json.loads(json_str) except json.JSONDecodeError: pass # Strategy 2: remove trailing commas before } or ] fixed = re.sub(r',(\s*[}\]])', r'\1', json_str) try: return json.loads(fixed) except json.JSONDecodeError: pass # Strategy 3: remove JS-style comments (// and /* */) fixed = re.sub(r'//[^\n]*', '', fixed) fixed = re.sub(r'/\*.*?\*/', '', fixed, flags=re.DOTALL) try: return json.loads(fixed) except json.JSONDecodeError: pass # Strategy 4: truncate at last valid closing brace # Walk backwards to find longest valid JSON prefix for i in range(len(fixed) - 1, start, -1): if fixed[i] == '}': candidate = fixed[start - start:i + 1] if start == 0 else fixed[:i + 1] # rebuild from inner start c2 = fixed[:i + 1] if start == 0 else json_str[:i - start + 1] try: data = json.loads(c2) logger.warning(f"⚠️ JSON truncated to position {i} — partial parse OK") return data except json.JSONDecodeError: continue break # Strategy 5: regex extraction of key fields (always succeeds with partial data) logger.warning("⚠️ All JSON strategies failed — using regex field extraction") return self._extract_fields_with_regex(response) def _extract_fields_with_regex(self, text: str) -> Dict: """Extract invoice fields from text using regex when JSON parsing fails""" def _find(pattern, default=None): m = re.search(pattern, text, re.IGNORECASE) return m.group(1).strip() if m else default def _find_num(pattern): m = re.search(pattern, text, re.IGNORECASE) if not m: return None val = m.group(1).replace('.', '').replace(',', '.') try: return float(val) except: return None result = { "document_type": _find(r'"document_type"\s*:\s*"([^"]+)"', 'invoice'), "invoice_number": _find(r'"invoice_number"\s*:\s*"?([^",\n}]+)"?'), "vendor_name": _find(r'"vendor_name"\s*:\s*"([^"]+)"'), "vendor_cvr": _find(r'"vendor_cvr"\s*:\s*"?(\d{8})"?'), "invoice_date": _find(r'"invoice_date"\s*:\s*"([^"]+)"'), "due_date": _find(r'"due_date"\s*:\s*"([^"]+)"'), "currency": _find(r'"currency"\s*:\s*"([^"]+)"', 'DKK'), "total_amount": _find_num(r'"total_amount"\s*:\s*([\d.,]+)'), "vat_amount": _find_num(r'"vat_amount"\s*:\s*([\d.,]+)'), "confidence": 0.5, "lines": [], "_partial": True, } logger.info(f"🔧 Regex extraction: vendor={result['vendor_name']}, cvr={result['vendor_cvr']}, total={result['total_amount']}") return result def calculate_file_checksum(self, file_path: Path) -> str: """Calculate SHA256 checksum of file for duplicate detection""" sha256 = hashlib.sha256() with open(file_path, 'rb') as f: while chunk := f.read(8192): sha256.update(chunk) checksum = sha256.hexdigest() logger.info(f"📋 Calculated checksum: {checksum[:16]}... for {file_path.name}") return checksum async def _extract_text_from_file(self, file_path: Path) -> str: """Extract text from PDF, image, or text file""" suffix = file_path.suffix.lower() try: if suffix == '.pdf': return await self._extract_text_from_pdf(file_path) elif suffix in ['.png', '.jpg', '.jpeg']: return await self._extract_text_from_image(file_path) elif suffix in ['.txt', '.csv']: with open(file_path, 'r', encoding='utf-8', errors='ignore') as f: return f.read() else: raise ValueError(f"Unsupported file type: {suffix}") except Exception as e: logger.error(f"❌ Text extraction failed for {file_path.name}: {e}") raise async def _extract_text_from_pdf(self, file_path: Path) -> str: """Extract text from PDF using pdfplumber (better table/layout support)""" try: import pdfplumber all_text = [] with pdfplumber.open(file_path) as pdf: for page_num, page in enumerate(pdf.pages): # Strategy: Use regular text extraction (includes tables) # pdfplumber's extract_text() handles tables better than PyPDF2 page_text = page.extract_text(layout=True, x_tolerance=2, y_tolerance=2) if page_text: all_text.append(page_text) text = "\\n".join(all_text) logger.info(f"📄 Extracted {len(text)} chars from PDF with pdfplumber") return text except Exception as e: logger.error(f"❌ PDF extraction failed: {e}") raise async def _extract_text_from_image(self, file_path: Path) -> str: """Extract text from image using Tesseract OCR""" try: import pytesseract from PIL import Image image = Image.open(file_path) # Use Danish + English for OCR text = pytesseract.image_to_string(image, lang='dan+eng') logger.info(f"🖼️ Extracted {len(text)} chars from image via OCR") return text except Exception as e: logger.error(f"❌ OCR extraction failed: {e}") # Fallback to English only try: text = pytesseract.image_to_string(Image.open(file_path), lang='eng') logger.warning(f"⚠️ Fallback to English OCR: {len(text)} chars") return text except: raise def _get_mime_type(self, file_path: Path) -> str: """Get MIME type from file extension""" suffix = file_path.suffix.lower() mime_types = { '.pdf': 'application/pdf', '.png': 'image/png', '.jpg': 'image/jpeg', '.jpeg': 'image/jpeg', '.txt': 'text/plain', '.csv': 'text/csv' } return mime_types.get(suffix, 'application/octet-stream') async def quick_analysis_on_upload(self, pdf_text: str) -> Dict: """ Quick analysis when file is uploaded - extracts critical fields only: - CVR number (to match vendor) - Document type (invoice vs credit note) - Invoice/credit note number This runs BEFORE template matching for early vendor detection. Args: pdf_text: Extracted text from PDF Returns: Dict with cvr, document_type, document_number, vendor_id, vendor_name, is_own_invoice """ from app.core.config import settings logger.info("⚡ Running quick analysis on upload...") result = { "cvr": None, "document_type": None, # 'invoice' or 'credit_note' "document_number": None, "vendor_id": None, "vendor_name": None, "is_own_invoice": False # True if this is an outgoing invoice (BMC's own CVR) } # 1. FIND CVR NUMBER (8 digits) # Look for patterns like "CVR: 12345678", "CVR-nr.: 12345678", "CVR 12345678" # Important: Supplier invoices have BOTH buyer (BMC=29522790) and seller CVR # We need the SELLER's CVR (not BMC's own) cvr_patterns = [ r'CVR[:\-\s]*(\d{8})', r'CVR[:\-\s]*nr\.?\s*(\d{8})', r'CVR[:\-\s]*nummer\s*(\d{8})', r'SE[:\-\s]*(\d{8})', # SE = Svensk CVR, men også brugt i DK r'\b(\d{8})\b' # Fallback: any 8-digit number ] # Find ALL CVR numbers in document found_cvrs = [] for pattern in cvr_patterns: matches = re.finditer(pattern, pdf_text, re.IGNORECASE) for match in matches: cvr_candidate = match.group(1) # Validate it's a real CVR (starts with 1-4, not a random number) if cvr_candidate[0] in '1234' and cvr_candidate not in found_cvrs: found_cvrs.append(cvr_candidate) # Remove BMC's own CVR from list (buyer CVR, not seller) vendor_cvrs = [cvr for cvr in found_cvrs if cvr != settings.OWN_CVR] if settings.OWN_CVR in found_cvrs: # This is a proper invoice where BMC is the buyer if len(vendor_cvrs) > 0: # Found vendor CVR - use the first non-BMC CVR result['cvr'] = vendor_cvrs[0] logger.info(f"📋 Found vendor CVR: {vendor_cvrs[0]} (ignored BMC CVR: {settings.OWN_CVR})") # Try to match vendor vendor = self.match_vendor_by_cvr(vendor_cvrs[0]) if vendor: result['vendor_id'] = vendor['id'] result['vendor_name'] = vendor['name'] else: # Only BMC's CVR found = this is an outgoing invoice result['is_own_invoice'] = True result['cvr'] = settings.OWN_CVR logger.warning(f"⚠️ OUTGOING INVOICE: Only BMC CVR found") elif len(vendor_cvrs) > 0: # No BMC CVR, but other CVR found - use first one result['cvr'] = vendor_cvrs[0] logger.info(f"📋 Found CVR: {vendor_cvrs[0]}") vendor = self.match_vendor_by_cvr(vendor_cvrs[0]) if vendor: result['vendor_id'] = vendor['id'] result['vendor_name'] = vendor['name'] # 2. DETECT DOCUMENT TYPE (Invoice vs Credit Note) credit_keywords = [ 'kreditnota', 'credit note', 'creditnote', 'kreditfaktura', 'refusion', 'tilbagebetaling', 'godtgørelse', 'tilbageførsel' ] text_lower = pdf_text.lower() is_credit_note = any(keyword in text_lower for keyword in credit_keywords) if is_credit_note: result['document_type'] = 'credit_note' logger.info("📄 Document type: CREDIT NOTE") else: result['document_type'] = 'invoice' logger.info("📄 Document type: INVOICE") # 3. EXTRACT DOCUMENT NUMBER # For invoices: "Faktura nr.", "Invoice number:", "Fakturanr." # For credit notes: "Kreditnota nr.", "Credit note number:" if result['document_type'] == 'credit_note': number_patterns = [ r'kreditnota\s*(?:nr\.?|nummer)[:\s]*(\S+)', r'credit\s*note\s*(?:no\.?|number)[:\s]*(\S+)', r'kreditfaktura\s*(?:nr\.?|nummer)[:\s]*(\S+)', ] else: number_patterns = [ r'faktura\s*(?:nr\.?|nummer)[:\s]*(\S+)', r'invoice\s*(?:no\.?|number)[:\s]*(\S+)', r'fakturanr\.?\s*[:\s]*(\S+)', ] for pattern in number_patterns: match = re.search(pattern, pdf_text, re.IGNORECASE) if match: result['document_number'] = match.group(1).strip() logger.info(f"🔢 Document number: {result['document_number']}") break logger.info(f"✅ Quick analysis complete: CVR={result['cvr']}, Type={result['document_type']}, Number={result['document_number']}, Vendor={result['vendor_name']}") return result def match_vendor_by_cvr(self, vendor_cvr: Optional[str]) -> Optional[Dict]: """ Match vendor from database using CVR number Args: vendor_cvr: CVR number from extraction Returns: Vendor dict if found, None otherwise """ if not vendor_cvr: return None # Clean CVR (remove spaces, dashes) cvr_clean = re.sub(r'[^0-9]', '', vendor_cvr) if len(cvr_clean) != 8: logger.warning(f"⚠️ Invalid CVR format: {vendor_cvr} (cleaned: {cvr_clean})") return None # Search vendors table vendor = execute_query_single( "SELECT * FROM vendors WHERE cvr_number = %s", (cvr_clean,)) if vendor: logger.info(f"✅ Matched vendor: {vendor['name']} (CVR: {cvr_clean})") return vendor else: logger.info(f"⚠️ No vendor found with CVR: {cvr_clean}") return None async def generate_summary(self, text: str) -> str: """ Generate a short summary of the text using Ollama """ if not text: return "" system_prompt = "Du er en hjælpsom assistent, der laver korte, præcise resuméer på dansk." user_prompt = f"Lav et kort resumé (max 50 ord) af følgende tekst:\n\n{text}" try: import aiohttp logger.info(f"🧠 Generating summary with Ollama ({self.model})...") async with aiohttp.ClientSession() as session: payload = { "model": self.model, "prompt": system_prompt + "\n\n" + user_prompt, "stream": False, "options": {"temperature": 0.3} } async with session.post(f"{self.endpoint}/api/generate", json=payload, timeout=60.0) as response: if response.status == 200: data = await response.json() summary = data.get("response", "").strip() logger.info("✅ Summary generated") return summary else: error_text = await response.text() logger.error(f"❌ Ollama error: {error_text}") return "Kunne ikke generere resumé (API fejl)." except Exception as e: logger.error(f"❌ Ollama summary failed: {e}") return f"Ingen resumé (Fejl: {str(e)})" # Global instance ollama_service = OllamaService()