441 lines
17 KiB
Python
441 lines
17 KiB
Python
#!/usr/bin/env python3
|
|
"""
|
|
extract_buyers_poc.py - POC: extrahiert Buyer-Information-Sheet-Felder aus PDFs.
|
|
|
|
Verbesserte Fassung mit:
|
|
- HTTP-Timeout (kein unendliches Haengen mehr bei Problem-PDFs)
|
|
- Seitenlimit gegen riesige Scans (z.B. SALADINO mit 59 Seiten)
|
|
- hoehere Vision-Aufloesung + Gemma-Vision-Budget fuer Checkbox-Erkennung
|
|
- Guided Decoding (JSON-Schema erzwungen) mit Loose-Parser als Fallback
|
|
- Wahl zwischen nur-oberste-Ebene und rekursiv (--recursive)
|
|
- Sortierung nach (aus Dateiname geparstem) Datum fuer Nachvollziehbarkeit
|
|
|
|
Ablauf pro PDF:
|
|
1. Text-Extraktion (pypdf). Genug Text -> Text-Pfad.
|
|
2. Zu wenig Text (Scan) -> erste Seiten als Bilder -> Vision-Pfad.
|
|
3. VLM/LLM extrahiert Zielfelder als JSON (Schema-guided).
|
|
4. Eine Zeile pro PDF -> buyers_raw.jsonl
|
|
5. Am Ende: Business-Kategorien -> kategorien_roh.txt
|
|
|
|
Zielfelder verteilen sich ueber ZWEI Seiten (Buyer Information Sheet +
|
|
Confidentiality Agreement). Fehlt eine Seite -> ihre Felder bleiben null.
|
|
|
|
Aufruf:
|
|
python extract_buyers_poc.py \
|
|
--src "/mnt/bizmatch-nas/AA Buyers NDA's/Buyers NDA's A-Z/S" \
|
|
--out ./poc_out \
|
|
--api http://192.168.100.160:8000/v1 \
|
|
--model "google/gemma-4-12b-it" \
|
|
--limit 150
|
|
"""
|
|
|
|
import os
|
|
import io
|
|
import sys
|
|
import json
|
|
import base64
|
|
import argparse
|
|
import glob
|
|
import re
|
|
import datetime
|
|
|
|
from openai import OpenAI
|
|
from pypdf import PdfReader
|
|
from pdf2image import convert_from_path
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# Zielschema (fuer Guided Decoding) + Prompt
|
|
# ---------------------------------------------------------------------------
|
|
|
|
JSON_SCHEMA = {
|
|
"type": "object",
|
|
"properties": {
|
|
"is_buyer_sheet": {"type": "boolean"},
|
|
"name_company": {"type": ["string", "null"]},
|
|
"prospective_buyer": {"type": ["string", "null"]},
|
|
"company": {"type": ["string", "null"]},
|
|
"phone": {"type": ["string", "null"]},
|
|
"cell": {"type": ["string", "null"]},
|
|
"email": {"type": ["string", "null"]},
|
|
"address": {"type": ["string", "null"]},
|
|
"state": {"type": ["string", "null"]},
|
|
"how_did_you_hear": {"type": ["string", "null"]},
|
|
"interested_in_updates": {"type": ["boolean", "null"]},
|
|
"types_of_business_raw": {"type": ["string", "null"]},
|
|
"types_of_business": {"type": "array", "items": {"type": "string"}},
|
|
"background_experience": {"type": ["string", "null"]},
|
|
"date_of_introduction": {"type": ["string", "null"]},
|
|
},
|
|
"required": [
|
|
"is_buyer_sheet", "name_company", "prospective_buyer", "company",
|
|
"phone", "cell", "email", "address", "state", "how_did_you_hear",
|
|
"interested_in_updates", "types_of_business_raw", "types_of_business",
|
|
"background_experience", "date_of_introduction"
|
|
],
|
|
}
|
|
|
|
SYSTEM_PROMPT = """Du bist ein praezises Datenextraktions-System fuer Formulare der Firma "BizMatch Business Brokerage".
|
|
|
|
Es gibt zwei relevante Seiten:
|
|
1. "BUYER INFORMATION SHEET" - enthaelt: NAME/COMPANY, PHONE/FAX/CELL, ADDRESS, EMAIL, "HOW DID YOU HEAR ABOUT US", die Frage "ARE YOU INTERESTED IN UPDATES ON NEW LISTINGS" mit ZWEI KAESTCHEN direkt vor den Woertern YES und NO, "TYPES OF BUSINESSES INTERESTED IN", "BACKGROUND / BUSINESS EXPERIENCE".
|
|
2. "CONFIDENTIALITY AGREEMENT" - enthaelt: "Prospective Buyer / Broker / Contact" (der ECHTE Name der Person), "Company (if applicable)", "Date of Introduction".
|
|
|
|
Ignoriere juristischen Vertragstext, Brokerage Services, Dropbox Sign, Unterschriften und alles, was unten nicht verlangt wird. Wenn KEINE dieser beiden BizMatch-Seiten erkennbar ist, setze is_buyer_sheet=false und alle anderen Felder null bzw. [].
|
|
|
|
CHECKBOX-ANWEISUNG (sehr wichtig): Bei "ARE YOU INTERESTED IN UPDATES" stehen zwei kleine Quadrate, je eines direkt VOR "YES" und vor "NO". Untersuche genau, welches Quadrat eine Markierung enthaelt (Haken, X, Kreuz, ausgefuellt, handschriftlich). Nur YES markiert -> true. Nur NO markiert -> false. Keines klar markiert -> null.
|
|
|
|
REGELN fuer "types_of_business":
|
|
- Extrahiere NUR die Geschaefts-KATEGORIE, nicht Umsatz/Groesse/Zusatz.
|
|
- "High Volume Restaurants sales 5Mil" -> ["Restaurant"]
|
|
- "Oil field service company, trucking" -> ["Oil Field", "Trucking"]
|
|
- "any profitable business" -> ["Any"]
|
|
- Singular, knapp, englisch. Leer/unklar -> [].
|
|
|
|
REGELN allgemein:
|
|
- Fehlendes/leeres Feld -> null (bzw. [] fuer Listen).
|
|
- ERFINDE NICHTS. Nur was tatsaechlich im Dokument steht.
|
|
- Bei "date_of_introduction": nach Moeglichkeit Format YYYY-MM-DD, sonst Originaltext, sonst null.
|
|
- Antworte NUR mit dem JSON-Objekt.
|
|
"""
|
|
|
|
USER_TEXT_INTRO = "Hier ist das Dokument als extrahierter Text:\n\n"
|
|
USER_IMG_INTRO = ("Hier sind die ersten Seiten des gescannten Dokuments. Achte "
|
|
"besonders auf die angekreuzte YES/NO-Checkbox. Extrahiere die Felder:")
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# Datum aus Dateinamen parsen (fuer Sortierung + als Datums-Fallback)
|
|
# ---------------------------------------------------------------------------
|
|
|
|
def date_from_filename(path):
|
|
"""
|
|
Sucht ein Datum im Dateinamen. Typische Muster: 120623, 030924, 071923
|
|
(MMDDYY) sowie 2024-01-01 o.ae. Gibt (datetime.date | None).
|
|
"""
|
|
name = os.path.basename(path)
|
|
# ISO zuerst
|
|
m = re.search(r"(20\d{2})[-_](\d{2})[-_](\d{2})", name)
|
|
if m:
|
|
try:
|
|
return datetime.date(int(m.group(1)), int(m.group(2)), int(m.group(3)))
|
|
except ValueError:
|
|
pass
|
|
# 6-stellig MMDDYY (haeufigstes Muster hier)
|
|
for m in re.finditer(r"(?<!\d)(\d{2})(\d{2})(\d{2})(?!\d)", name):
|
|
mm, dd, yy = int(m.group(1)), int(m.group(2)), int(m.group(3))
|
|
year = 2000 + yy
|
|
try:
|
|
if 1 <= mm <= 12 and 1 <= dd <= 31:
|
|
return datetime.date(year, mm, dd)
|
|
except ValueError:
|
|
continue
|
|
return None
|
|
|
|
|
|
def _invert_date(iso):
|
|
"""Fuer absteigende Sortierung: Datum in einen Schluessel umkehren,
|
|
der bei aufsteigender Sortierung neueste zuerst liefert."""
|
|
# Ordinalzahl negieren -> groesseres Datum = kleinerer Schluessel
|
|
y, m, d = (int(x) for x in iso.split("-"))
|
|
return -datetime.date(y, m, d).toordinal()
|
|
|
|
|
|
def sort_key(path):
|
|
d = date_from_filename(path)
|
|
if d:
|
|
# Gruppe 0 (mit Datum), neueste zuerst
|
|
return (0, _invert_date(d.isoformat()), path)
|
|
# ohne erkanntes Datum: nach mtime (neueste zuerst), ans Ende
|
|
try:
|
|
mt = -int(os.path.getmtime(path))
|
|
except OSError:
|
|
mt = 0
|
|
return (1, mt, path)
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# PDF-Handling
|
|
# ---------------------------------------------------------------------------
|
|
|
|
def pdf_page_count(path):
|
|
try:
|
|
return len(PdfReader(path).pages)
|
|
except Exception:
|
|
return -1
|
|
|
|
|
|
def extract_text(pdf_path, max_pages=4):
|
|
try:
|
|
reader = PdfReader(pdf_path)
|
|
pages = reader.pages[:max_pages]
|
|
return "\n".join(p.extract_text() or "" for p in pages)
|
|
except Exception:
|
|
return ""
|
|
|
|
|
|
def pdf_to_base64_images(pdf_path, max_pages=2, max_dim=2200, quality=85, dpi=200):
|
|
"""
|
|
Erste max_pages Seiten -> Base64-JPEG.
|
|
Hoehere Aufloesung als zuvor (dpi=200, max_dim=2200), damit die kleinen
|
|
YES/NO-Kaestchen erkennbar bleiben. max_pages=2 reicht: Buyer-Felder
|
|
stehen auf den ersten Seiten, und es haelt die Vision-Token im Rahmen.
|
|
"""
|
|
try:
|
|
images = convert_from_path(pdf_path, first_page=1, last_page=max_pages, dpi=dpi)
|
|
except Exception as e:
|
|
print(f" [Bildkonvertierung-Fehler] {e}")
|
|
return []
|
|
out = []
|
|
for img in images:
|
|
w, h = img.size
|
|
scale = min(1.0, max_dim / max(w, h))
|
|
if scale < 1.0:
|
|
img = img.resize((int(w * scale), int(h * scale)))
|
|
buf = io.BytesIO()
|
|
img.convert("RGB").save(buf, format="JPEG", quality=quality)
|
|
out.append(base64.b64encode(buf.getvalue()).decode("utf-8"))
|
|
return out
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# LLM-Call (mit Guided Decoding + Vision-Budget)
|
|
# ---------------------------------------------------------------------------
|
|
|
|
def call_model(client, model, content_payload, vision_budget=None):
|
|
# llama.cpp erzwingt JSON ueber response_format mit json_schema.
|
|
# (vLLM-spezifisches guided_json / mm_processor_kwargs gibt es hier nicht.)
|
|
resp = client.chat.completions.create(
|
|
model=model,
|
|
messages=[
|
|
{"role": "system", "content": SYSTEM_PROMPT},
|
|
{"role": "user", "content": content_payload},
|
|
],
|
|
temperature=0.0,
|
|
max_tokens=1200,
|
|
response_format={
|
|
"type": "json_schema",
|
|
"json_schema": {"name": "buyer_sheet", "schema": JSON_SCHEMA},
|
|
},
|
|
)
|
|
return resp.choices[0].message.content
|
|
|
|
|
|
def parse_json_loose(txt):
|
|
if not txt:
|
|
return None
|
|
txt = txt.strip()
|
|
txt = re.sub(r"^```(?:json)?|```$", "", txt, flags=re.MULTILINE).strip()
|
|
start, end = txt.find("{"), txt.rfind("}")
|
|
if start == -1 or end == -1 or end < start:
|
|
return None
|
|
try:
|
|
return json.loads(txt[start:end + 1])
|
|
except json.JSONDecodeError:
|
|
return None
|
|
|
|
|
|
# Seitenlimits
|
|
TEXT_SCAN_THRESHOLD = 120 # < so viele Textzeichen -> als Scan behandeln
|
|
MAX_SCAN_PAGES_TOTAL = 12 # PDFs mit mehr Seiten gelten als "kein reines Sheet"
|
|
|
|
# Felder, bei denen wir Vision bevorzugen (handschriftlich/visuell auf dem Formular)
|
|
_PREFER_VISION = {
|
|
"prospective_buyer", "name_company", "company", "phone", "cell", "email",
|
|
"address", "state", "how_did_you_hear", "interested_in_updates",
|
|
"date_of_introduction",
|
|
}
|
|
# Felder, bei denen Text meist die sauberere (getippte) Quelle ist
|
|
_PREFER_TEXT = {"background_experience", "types_of_business", "types_of_business_raw"}
|
|
|
|
|
|
def _empty(v):
|
|
return v in (None, "", [], {})
|
|
|
|
|
|
def merge_records(text_rec, vision_rec):
|
|
"""
|
|
Fuehrt Text- und Vision-Extraktion desselben PDF zusammen.
|
|
Grundregel: nicht-leerer Wert schlaegt leeren; bei Konflikt entscheidet
|
|
die bevorzugte Quelle je Feld (Vision fuer Handschrift, Text fuer Getipptes).
|
|
"""
|
|
if text_rec is None:
|
|
return vision_rec
|
|
if vision_rec is None:
|
|
return text_rec
|
|
merged = {}
|
|
keys = set(text_rec) | set(vision_rec)
|
|
for k in keys:
|
|
tv, vv = text_rec.get(k), vision_rec.get(k)
|
|
if _empty(tv) and _empty(vv):
|
|
merged[k] = tv if k in text_rec else vv
|
|
elif _empty(tv):
|
|
merged[k] = vv
|
|
elif _empty(vv):
|
|
merged[k] = tv
|
|
else:
|
|
# beide gefuellt -> bevorzugte Quelle
|
|
if k in _PREFER_VISION:
|
|
merged[k] = vv
|
|
elif k in _PREFER_TEXT:
|
|
merged[k] = tv
|
|
else:
|
|
merged[k] = vv # Default: Vision
|
|
return merged
|
|
|
|
|
|
def _extract_via_text(client, model, text):
|
|
payload = USER_TEXT_INTRO + text[:12000]
|
|
raw = call_model(client, model, payload)
|
|
return parse_json_loose(raw)
|
|
|
|
|
|
def _extract_via_vision(client, model, pdf_path, max_pages):
|
|
imgs = pdf_to_base64_images(pdf_path, max_pages=max_pages)
|
|
if not imgs:
|
|
return None
|
|
payload = [{"type": "text", "text": USER_IMG_INTRO}]
|
|
for b64 in imgs:
|
|
payload.append({"type": "image_url",
|
|
"image_url": {"url": f"data:image/jpeg;base64,{b64}"}})
|
|
raw = call_model(client, model, payload)
|
|
return parse_json_loose(raw)
|
|
|
|
|
|
def process_pdf(client, model, pdf_path, vision_budget=None):
|
|
"""
|
|
Hybrid-Strategie:
|
|
- reiner Scan (keine Textebene): nur Vision
|
|
- grosser Scan (> MAX_SCAN_PAGES_TOTAL): nur erste Seite Vision
|
|
- getipptes PDF mit Textebene: Text UND Vision, dann mergen
|
|
(Text bringt getippten Background, Vision die handschriftlichen Felder)
|
|
"""
|
|
n_pages = pdf_page_count(pdf_path)
|
|
text = extract_text(pdf_path)
|
|
has_text = len(text.strip()) >= TEXT_SCAN_THRESHOLD
|
|
|
|
truncated_note = None
|
|
try:
|
|
if not has_text and n_pages > MAX_SCAN_PAGES_TOTAL:
|
|
# grosser reiner Scan: nur erste Seite pruefen
|
|
truncated_note = f"grosser Scan ({n_pages} Seiten), nur erste Seite geprueft"
|
|
data = _extract_via_vision(client, model, pdf_path, max_pages=1)
|
|
mode = "vision"
|
|
elif not has_text:
|
|
# reiner Scan
|
|
data = _extract_via_vision(client, model, pdf_path, max_pages=2)
|
|
mode = "vision"
|
|
else:
|
|
# Textebene vorhanden -> HYBRID: beide Pfade, dann mergen
|
|
text_rec = _extract_via_text(client, model, text)
|
|
vision_rec = None
|
|
if n_pages <= MAX_SCAN_PAGES_TOTAL:
|
|
vision_rec = _extract_via_vision(client, model, pdf_path, max_pages=2)
|
|
data = merge_records(text_rec, vision_rec)
|
|
mode = "hybrid" if vision_rec is not None else "text"
|
|
except Exception as e:
|
|
return {"_error": f"API: {type(e).__name__}: {e}", "n_pages": n_pages}, "?"
|
|
|
|
if data is None:
|
|
return {"_error": "JSON-Parse fehlgeschlagen", "n_pages": n_pages}, mode
|
|
data["n_pages"] = n_pages
|
|
if truncated_note:
|
|
data["_note"] = truncated_note
|
|
return data, mode
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# Hauptlauf
|
|
# ---------------------------------------------------------------------------
|
|
|
|
def main():
|
|
ap = argparse.ArgumentParser()
|
|
ap.add_argument("--src", required=True)
|
|
ap.add_argument("--out", default="./poc_out")
|
|
ap.add_argument("--api", default="http://192.168.100.160:8000/v1")
|
|
ap.add_argument("--model", default="gemma-4-12b") # --alias des llama-server
|
|
ap.add_argument("--limit", type=int, default=150)
|
|
ap.add_argument("--recursive", action="store_true",
|
|
help="auch Unterordner durchsuchen (Default: nur oberste Ebene)")
|
|
ap.add_argument("--timeout", type=float, default=120,
|
|
help="HTTP-Timeout je Anfrage in Sekunden")
|
|
ap.add_argument("--vision-budget", type=int, default=0,
|
|
help="Gemma Vision-Token-Budget (0=aus/Default 280; sonst 70/140/280/560/1120). "
|
|
"ACHTUNG: hohe Werte koennen auf manchen ROCm-Builds das Bild zerstoeren.")
|
|
args = ap.parse_args()
|
|
|
|
os.makedirs(args.out, exist_ok=True)
|
|
jsonl_path = os.path.join(args.out, "buyers_raw.jsonl")
|
|
cats_path = os.path.join(args.out, "kategorien_roh.txt")
|
|
|
|
# Timeout am Client verhindert unendliches Haengen
|
|
client = OpenAI(base_url=args.api, api_key="test-key", timeout=args.timeout, max_retries=1)
|
|
|
|
if args.recursive:
|
|
pdfs = glob.glob(os.path.join(args.src, "**", "*.pdf"), recursive=True)
|
|
else:
|
|
pdfs = glob.glob(os.path.join(args.src, "*.pdf"))
|
|
|
|
# nach (Dateiname-)Datum sortieren
|
|
pdfs = sorted(pdfs, key=sort_key)
|
|
if args.limit:
|
|
pdfs = pdfs[:args.limit]
|
|
print(f"{len(pdfs)} PDFs im POC-Umfang (sortiert nach Datum, "
|
|
f"{'rekursiv' if args.recursive else 'nur oberste Ebene'}).\n")
|
|
|
|
all_categories = {}
|
|
n_ok = n_err = n_text = n_vision = n_hybrid = n_nosheet = 0
|
|
|
|
with open(jsonl_path, "w", encoding="utf-8") as jf:
|
|
for i, pdf in enumerate(pdfs, 1):
|
|
rel = os.path.relpath(pdf, args.src)
|
|
fdate = date_from_filename(pdf)
|
|
print(f"[{i}/{len(pdfs)}] {fdate or '????-??-??'} {rel}")
|
|
try:
|
|
data, mode = process_pdf(client, args.model, pdf, args.vision_budget)
|
|
except KeyboardInterrupt:
|
|
print("\nAbbruch durch Nutzer. Bisheriges ist gesichert.")
|
|
break
|
|
except Exception as e:
|
|
data, mode = {"_error": f"unerwartet: {type(e).__name__}: {e}"}, "?"
|
|
|
|
n_text += (mode == "text")
|
|
n_vision += (mode == "vision")
|
|
n_hybrid += (mode == "hybrid")
|
|
record = {"source_file": rel,
|
|
"file_date": fdate.isoformat() if fdate else None,
|
|
"extraction_mode": mode, **data}
|
|
jf.write(json.dumps(record, ensure_ascii=False) + "\n")
|
|
jf.flush()
|
|
|
|
if "_error" in data:
|
|
n_err += 1
|
|
print(f" FEHLER: {data['_error']}")
|
|
continue
|
|
n_ok += 1
|
|
if data.get("is_buyer_sheet") is False:
|
|
n_nosheet += 1
|
|
for cat in (data.get("types_of_business") or []):
|
|
key = cat.strip().lower()
|
|
if key:
|
|
all_categories[key] = all_categories.get(key, 0) + 1
|
|
|
|
with open(cats_path, "w", encoding="utf-8") as cf:
|
|
for cat, cnt in sorted(all_categories.items(), key=lambda x: -x[1]):
|
|
cf.write(f"{cnt:4d} {cat}\n")
|
|
|
|
print("\n=== POC-Zusammenfassung ===")
|
|
print(f" verarbeitet: {n_ok + n_err}")
|
|
print(f" erfolgreich: {n_ok}")
|
|
print(f" Fehler: {n_err}")
|
|
print(f" Text-Pfad: {n_text}")
|
|
print(f" Vision-Pfad: {n_vision}")
|
|
print(f" Hybrid-Pfad: {n_hybrid}")
|
|
print(f" kein Buyer-Sheet: {n_nosheet}")
|
|
print(f" distinkte Kategorien (roh): {len(all_categories)}")
|
|
print(f"\n JSONL: {jsonl_path}")
|
|
print(f" Kategorien: {cats_path}")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main() |