feat: Phase 5 – PDF/Bild-Beleg-Import (lokal, Tesseract-OCR)

Backend:
- services/ocr: PDF-Text via PyMuPDF, OCR (Tesseract) für Scans/Bilder,
  heuristischer Parser (Datum + Betrag je Zeile, CH/DE/US-Beträge).
- routers/documents: /documents/preview (Upload -> Text -> Kandidaten ->
  gemeinsame Prüf-/Buchungslogik des CSV-Imports). csv_import refaktoriert:
  detect_rows() von CSV und Beleg gemeinsam genutzt.
- requirements: PyMuPDF, pytesseract, pillow; install.sh installiert tesseract-ocr
  (+ deu). pytest: Heuristik-Parser und PDF-Vorschau.

Frontend:
- Import-Seite: CSV/PDF-Umschalter; PDF-Modus mit Beleg-Auslesen, extrahiertem
  Text zur Kontrolle und Best-Effort-Hinweis; gleiche Prüfansicht + Buchen.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

backend/app/main.py

@@ -24,6 +24,7 @@ from .routers import (
     budgets,
     categories,
     dashboard,
+    documents,
     imports,
     rates,
     settings as settings_router,
@@ -105,6 +106,7 @@ api.include_router(transactions.router)
 api.include_router(budgets.router)
 api.include_router(dashboard.router)
 api.include_router(imports.router)
+api.include_router(documents.router)
 api.include_router(rates.router)
 
 app.include_router(api)

backend/app/routers/documents.py

@@ -0,0 +1,34 @@
+"""PDF-/Bild-Beleg-Import (Phase 5, lokal). Nutzt die Prüf-/Buchungslogik des CSV-Imports."""
+
+from fastapi import APIRouter, Depends, File, Form, HTTPException, UploadFile
+from sqlalchemy.orm import Session
+
+from ..database import get_db
+from ..deps import get_current_user, get_owned
+from ..models import Account, User
+from ..services import csv_import, ocr
+
+router = APIRouter(prefix="/documents", tags=["documents"])
+
+
+@router.post("/preview")
+async def preview(
+    file: UploadFile = File(...),
+    account_id: int = Form(...),
+    db: Session = Depends(get_db),
+    user: User = Depends(get_current_user),
+) -> dict:
+    account = get_owned(db, Account, account_id, user)
+    data = await file.read()
+    try:
+        text = ocr.extract_text(data, file.filename or "", file.content_type or "")
+    except ocr.ExtractionError as exc:
+        raise HTTPException(status_code=400, detail=str(exc))
+    candidates = ocr.parse_transactions(text)
+    rows = csv_import.detect_rows(db, user, account, candidates)
+    return {
+        "rows": rows,
+        "text": text[:8000],
+        "count": len(rows),
+        "filename": file.filename or "",
+    }

backend/app/services/csv_import.py

@@ -153,16 +153,45 @@ def _apply_rules(rules: list[CategorizationRule], payee: str, purpose: str) -> i
     return None
 
 
-def build_preview(db: Session, user: User, account: Account, template: ImportTemplate, data: bytes) -> list[dict]:
+def parse_file_to_candidates(data: bytes, template: ImportTemplate) -> list[dict]:
+    """Zerlegt die Datei anhand des Templates in rohe Kandidaten-Zeilen."""
     cmap = json.loads(template.column_map or "{}")
 
     def col(field: str) -> int | None:
         idx = cmap.get(field)
         return idx if isinstance(idx, int) and idx >= 0 else None
 
-    rows = read_rows(data, template)
+    out: list[dict] = []
+    for r in read_rows(data, template):
+        def cell(field: str) -> str:
+            c = col(field)
+            return r[c].strip() if c is not None and c < len(r) else ""
 
-    # Identifikatoren aller Konten des Nutzers -> Konto-ID (für Transfer-Erkennung).
+        d = parse_date(cell("date"), template.date_format)
+        if template.amount_mode == AmountMode.debit_credit:
+            debit = parse_amount(cell("debit"), template.decimal_sep)
+            credit = parse_amount(cell("credit"), template.decimal_sep)
+            amount = (credit or 0) - (abs(debit) if debit else 0)
+        elif template.amount_mode == AmountMode.abs_expense:
+            a = parse_amount(cell("amount"), template.decimal_sep)
+            amount = -abs(a) if a is not None else None
+        else:
+            amount = parse_amount(cell("amount"), template.decimal_sep)
+
+        out.append({
+            "date": d.isoformat() if d else None,
+            "amount_cents": amount,
+            "payee": cell("payee") or None,
+            "purpose": cell("purpose") or None,
+            "counterparty_iban": normalize_iban(cell("counterparty_iban")) or None,
+        })
+    return out
+
+
+def detect_rows(db: Session, user: User, account: Account, candidates: list[dict]) -> list[dict]:
+    """Reichert Kandidaten an: Typ, Auto-Kategorie, interner Transfer, Duplikat.
+
+    Wird von CSV- UND Beleg-Import genutzt (gleiche Prüf-/Buchungslogik)."""
     idents: dict[str, int] = {}
     for value_norm, acc_id in (
         db.query(AccountIdentifier.value_norm, AccountIdentifier.account_id)
@@ -177,39 +206,25 @@ def build_preview(db: Session, user: User, account: Account, template: ImportTem
         .order_by(CategorizationRule.priority.desc(), CategorizationRule.hits.desc())
         .all()
     )
-
     existing = {
         h[0]
         for h in db.query(Transaction.import_hash).filter(
             Transaction.owner_id == user.id, Transaction.import_hash.isnot(None)
         )
     }
-    seen_in_file: set[str] = set()
+    seen: set[str] = set()
 
     out: list[dict] = []
-    for i, r in enumerate(rows):
+    for i, cand in enumerate(candidates):
-        def cell(field: str) -> str:
+        d_iso = cand.get("date")
-            c = col(field)
+        amount = cand.get("amount_cents")
-            return r[c].strip() if c is not None and c < len(r) else ""
+        payee = cand.get("payee")
-
+        purpose = cand.get("purpose")
-        d = parse_date(cell("date"), template.date_format)
+        iban = cand.get("counterparty_iban")
-
+        d = date.fromisoformat(d_iso) if d_iso else None
-        if template.amount_mode == AmountMode.debit_credit:
-            debit = parse_amount(cell("debit"), template.decimal_sep)
-            credit = parse_amount(cell("credit"), template.decimal_sep)
-            amount = (credit or 0) - (abs(debit) if debit else 0)
-        elif template.amount_mode == AmountMode.abs_expense:
-            a = parse_amount(cell("amount"), template.decimal_sep)
-            amount = -abs(a) if a is not None else None
-        else:
-            amount = parse_amount(cell("amount"), template.decimal_sep)
-
-        payee = cell("payee")
-        purpose = cell("purpose")
-        iban = normalize_iban(cell("counterparty_iban"))
 
         valid = d is not None and amount is not None and amount != 0
-        ttype = "transfer" if False else ("expense" if (amount or 0) < 0 else "income")
+        ttype = "expense" if (amount or 0) < 0 else "income"
         category_id: int | None = None
         to_account_id: int | None = None
 
@@ -218,11 +233,11 @@ def build_preview(db: Session, user: User, account: Account, template: ImportTem
             ttype = "transfer"
             to_account_id = other
         else:
-            category_id = _apply_rules(rules, payee, purpose)
+            category_id = _apply_rules(rules, payee or "", purpose or "")
 
         h = import_hash(account.id, d, amount, payee)
-        is_dup = h in existing or h in seen_in_file
+        is_dup = h in existing or h in seen
-        seen_in_file.add(h)
+        seen.add(h)
 
         out.append({
             "index": i,
@@ -242,6 +257,10 @@ def build_preview(db: Session, user: User, account: Account, template: ImportTem
     return out
 
 
+def build_preview(db: Session, user: User, account: Account, template: ImportTemplate, data: bytes) -> list[dict]:
+    return detect_rows(db, user, account, parse_file_to_candidates(data, template))
+
+
 # --- Verbuchen -------------------------------------------------------------
 
 

backend/app/services/ocr.py

@@ -0,0 +1,127 @@
+"""PDF-/Bild-Belege lokal auslesen (kein Cloud-Dienst).
+
+- Digitale PDFs: Text direkt via PyMuPDF.
+- Gescannte PDFs / Bilder: OCR via Tesseract (pytesseract).
+Danach ein heuristischer Parser, der Zeilen mit Datum + Betrag findet. Das ist
+bewusst Best-Effort – das Ergebnis landet in derselben Prüfansicht wie der
+CSV-Import und wird dort korrigiert.
+"""
+
+import io
+import re
+from datetime import datetime
+from decimal import Decimal, InvalidOperation
+
+try:  # PyMuPDF (Text + Rendern)
+    import fitz  # type: ignore
+except Exception:  # noqa: BLE001
+    fitz = None
+
+try:  # OCR
+    import pytesseract  # type: ignore
+    from PIL import Image  # type: ignore
+except Exception:  # noqa: BLE001
+    pytesseract = None
+    Image = None
+
+
+class ExtractionError(RuntimeError):
+    pass
+
+
+def _image_to_text(data: bytes) -> str:
+    if pytesseract is None or Image is None:
+        raise ExtractionError("OCR nicht verfügbar (pytesseract/Pillow/Tesseract fehlen).")
+    img = Image.open(io.BytesIO(data))
+    return pytesseract.image_to_string(img, lang="deu+eng")
+
+
+def _pdf_to_text(data: bytes) -> str:
+    if fitz is None:
+        raise ExtractionError("PDF-Verarbeitung nicht verfügbar (PyMuPDF fehlt).")
+    parts: list[str] = []
+    with fitz.open(stream=data, filetype="pdf") as doc:
+        for page in doc:
+            txt = page.get_text()
+            if txt and txt.strip():
+                parts.append(txt)
+            elif pytesseract is not None and Image is not None:
+                pix = page.get_pixmap(dpi=200)
+                img = Image.open(io.BytesIO(pix.tobytes("png")))
+                parts.append(pytesseract.image_to_string(img, lang="deu+eng"))
+    return "\n".join(parts)
+
+
+def extract_text(data: bytes, filename: str, content_type: str) -> str:
+    name = (filename or "").lower()
+    if name.endswith(".pdf") or content_type == "application/pdf":
+        return _pdf_to_text(data)
+    return _image_to_text(data)
+
+
+# --- Heuristischer Zeilen-Parser -------------------------------------------
+
+DATE_RE = re.compile(r"\b(\d{4}-\d{2}-\d{2}|\d{1,2}[./]\d{1,2}[./]\d{2,4})\b")
+# Betrag: entweder mit Tausender-Gruppen (1'234.56 / 1.234,56) oder schlicht (3000.00).
+AMOUNT_RE = re.compile(r"[-+]?(?:\d{1,3}(?:['.,\s]\d{3})+|\d+)[.,]\d{2}(?!\d)")
+_DATE_FORMATS = ["%Y-%m-%d", "%d.%m.%Y", "%d.%m.%y", "%d/%m/%Y", "%d/%m/%y"]
+
+
+def _norm_date(token: str):
+    for fmt in _DATE_FORMATS:
+        try:
+            return datetime.strptime(token, fmt).date()
+        except ValueError:
+            continue
+    return None
+
+
+def _parse_amount_auto(s: str) -> int | None:
+    """Betrag mit automatisch erkanntem Dezimaltrenner (CH/DE/US) -> Cent."""
+    s = s.strip()
+    neg = s.startswith("-") or (s.startswith("(") and s.endswith(")"))
+    m = re.search(r"[.,](\d{2})$", s)
+    if not m:
+        return None
+    int_digits = re.sub(r"\D", "", s[: m.start()])
+    frac = m.group(1)
+    try:
+        val = Decimal(f"{int_digits or '0'}.{frac}")
+    except InvalidOperation:
+        return None
+    cents = int(val * 100)
+    return -cents if neg else cents
+
+
+def parse_transactions(text: str) -> list[dict]:
+    """Findet Zeilen mit Datum + Betrag und macht daraus Kandidaten-Buchungen."""
+    rows: list[dict] = []
+    for line in text.splitlines():
+        line = line.strip()
+        if len(line) < 6:
+            continue
+        dm = DATE_RE.search(line)
+        if not dm:
+            continue
+        d = _norm_date(dm.group(0))
+        if d is None:
+            continue
+        rest = f"{line[: dm.start()]} {line[dm.end():]}"
+        amounts = AMOUNT_RE.findall(rest)
+        if not amounts:
+            continue
+        amount = _parse_amount_auto(amounts[0])
+        if amount is None or amount == 0:
+            continue
+        desc = rest
+        for a in amounts:
+            desc = desc.replace(a, " ")
+        desc = re.sub(r"\s+", " ", desc).strip(" -–|;,.")
+        rows.append({
+            "date": d.isoformat(),
+            "amount_cents": amount,
+            "payee": desc[:120] or None,
+            "purpose": None,
+            "counterparty_iban": None,
+        })
+    return rows

backend/requirements.txt

@@ -8,4 +8,9 @@ argon2-cffi==23.1.0
 python-multipart==0.0.20
 httpx==0.28.1
 python-dateutil==2.9.0.post0
+# PDF/Bild-Import (Phase 5): Text aus PDFs (PyMuPDF) und OCR (Tesseract) für Scans/Bilder.
+PyMuPDF==1.25.1
+pytesseract==0.3.13
+pillow==11.1.0
 pytest==8.3.4
+

backend/tests/test_documents.py

@@ -0,0 +1,55 @@
+import io
+
+import pytest
+
+fitz = pytest.importorskip("fitz")
+
+
+def _account(admin):
+    r = admin.post("/api/accounts", json={"name": "Beleg", "type": "girokonto", "currency": "EUR"})
+    assert r.status_code == 201
+    return r.json()
+
+
+def _make_pdf() -> bytes:
+    doc = fitz.open()
+    page = doc.new_page()
+    page.insert_text((50, 72), "15.03.2026   REWE Supermarkt Zuerich   -50.00")
+    page.insert_text((50, 96), "25.03.2026   Gehalt Firma AG   3000.00")
+    page.insert_text((50, 120), "Kein Datum hier, nur Text 12.34")
+    return doc.tobytes()
+
+
+def test_parse_transactions_heuristic():
+    from app.services import ocr
+
+    text = (
+        "Kontoauszug März 2026\n"
+        "15.03.2026 REWE Supermarkt -50,00\n"
+        "2026-03-25 Gehalt 3'000.00\n"
+        "Zwischensumme\n"
+    )
+    rows = ocr.parse_transactions(text)
+    assert len(rows) == 2
+    assert rows[0]["date"] == "2026-03-15"
+    assert rows[0]["amount_cents"] == -5000
+    assert "REWE" in rows[0]["payee"]
+    assert rows[1]["date"] == "2026-03-25"
+    assert rows[1]["amount_cents"] == 300000  # Schweizer Tausender-Apostroph
+
+
+def test_pdf_document_preview(admin):
+    acc = _account(admin)
+    pdf = _make_pdf()
+    r = admin.post(
+        "/api/documents/preview",
+        files={"file": ("beleg.pdf", io.BytesIO(pdf), "application/pdf")},
+        data={"account_id": str(acc["id"])},
+    )
+    assert r.status_code == 200, r.text
+    body = r.json()
+    assert body["count"] == 2
+    rows = body["rows"]
+    kinds = {(row["type"], row["amount_cents"]) for row in rows}
+    assert ("expense", -5000) in kinds
+    assert ("income", 300000) in kinds

deploy/install.sh

@@ -48,7 +48,8 @@ fi
 info "Installiere Systempakete (Python, Build-Tools, git, curl) …"
 export DEBIAN_FRONTEND=noninteractive
 apt-get update -y
-apt-get install -y python3 python3-venv python3-dev build-essential curl ca-certificates git
+apt-get install -y python3 python3-venv python3-dev build-essential curl ca-certificates git \
+  tesseract-ocr tesseract-ocr-deu
 
 # --- Python-Version prüfen (3.11+ erforderlich) ------------------------------
 PY_VER="$(python3 -c 'import sys; print("%d.%d" % sys.version_info[:2])' 2>/dev/null || echo 0.0)"

docs/ROADMAP.md

@@ -28,8 +28,9 @@ Bank-Templates (Trennzeichen/Dezimal/Datum/Spaltenmapping/Vorzeichen), lernende
 Auto-Kategorisierung, Prüfansicht, Duplikat-Erkennung, interne-Transfer-Erkennung
 via IBAN, Batch-Undo. Vorgefertigte Bank-Templates folgen aus Nutzer-Beispielen.
 
-## Phase 5 – PDF/Bild-Import (Stretch)
+## Phase 5 – PDF/Bild-Import (Stretch)  ✅
-Lokal (pdfplumber/pypdf + Tesseract-OCR), speist dieselbe Prüfansicht.
+Lokal: PyMuPDF (PDF-Text) + Tesseract-OCR (Scans/Bilder), heuristischer
+Zeilen-Parser (Datum + Betrag), speist dieselbe Prüfansicht. Best-Effort.
 
 ## Querschnitt – Backup/Export
 DB-Backup-Download + Export je Entität (CSV/JSON); dokumentierte LXC-Sicherung.

web/src/api.js

@@ -198,4 +198,7 @@ export const api = {
   importCommit: (body) => request("/imports/commit", { method: "POST", body }),
   importBatches: () => request("/imports/batches"),
   undoImportBatch: (id) => request(`/imports/batches/${id}`, { method: "DELETE" }),
+
+  // PDF/Bild-Beleg
+  documentPreview: (formData) => request("/documents/preview", { method: "POST", formData }),
 };

web/src/finance.css

@@ -560,3 +560,14 @@ tr.row-invalid {
 tr.row-dup {
   opacity: 0.55;
 }
+.doc-text {
+  max-height: 240px;
+  overflow: auto;
+  font-size: 0.78rem;
+  white-space: pre-wrap;
+  color: var(--muted);
+  background: var(--surface-2);
+  padding: var(--sp-3);
+  border-radius: var(--radius-sm);
+  margin-top: var(--sp-2);
+}

web/src/locales/de.json

@@ -197,13 +197,18 @@
     "targetAccount": "Zielkonto"
   },
   "import": {
-    "title": "CSV-Import",
+    "title": "Import",
+    "modeCsv": "CSV",
+    "modeDoc": "PDF · Bild",
+    "readDoc": "Beleg auslesen",
+    "docHint": "PDF-/Bild-Erkennung ist Best-Effort — bitte jede Zeile prüfen und ggf. korrigieren.",
+    "extractedText": "Extrahierter Text (zur Kontrolle)",
     "step1": "1 · Konto & Datei",
     "step2": "2 · Format",
     "step3": "3 · Spalten zuordnen",
     "step4": "4 · Prüfen & buchen",
     "account": "Zielkonto",
-    "file": "CSV-Datei",
+    "file": "Datei",
     "preset": "Bank-Vorlage (vorgefertigt)",
     "template": "Eigene Vorlage",
     "templateManual": "Manuell einstellen",

web/src/locales/en.json

@@ -197,13 +197,18 @@
     "targetAccount": "Target account"
   },
   "import": {
-    "title": "CSV import",
+    "title": "Import",
+    "modeCsv": "CSV",
+    "modeDoc": "PDF · Image",
+    "readDoc": "Read document",
+    "docHint": "PDF/image recognition is best-effort — please check and fix every row.",
+    "extractedText": "Extracted text (for reference)",
     "step1": "1 · Account & file",
     "step2": "2 · Format",
     "step3": "3 · Map columns",
     "step4": "4 · Review & import",
     "account": "Target account",
-    "file": "CSV file",
+    "file": "File",
     "preset": "Bank preset",
     "template": "My template",
     "templateManual": "Configure manually",

web/src/pages/Import.jsx

@@ -41,11 +41,13 @@ export default function Import() {
   const [presets, setPresets] = useState([]);
   const [batches, setBatches] = useState([]);
 
+  const [mode, setMode] = useState("csv"); // "csv" | "doc"
   const [accountId, setAccountId] = useState("");
   const [file, setFile] = useState(null);
   const [config, setConfig] = useState(defaultConfig());
   const [raw, setRaw] = useState(null);
   const [rows, setRows] = useState(null);
+  const [docText, setDocText] = useState("");
   const [busy, setBusy] = useState(false);
   const [saveName, setSaveName] = useState("");
 
@@ -117,6 +119,19 @@ export default function Import() {
     } catch (e) { toast(e.message, "error"); } finally { setBusy(false); }
   }
 
+  async function loadDoc() {
+    if (!file || !accountId) { toast(t("import.needFileAccount"), "error"); return; }
+    setBusy(true);
+    try {
+      const fd = new FormData();
+      fd.append("file", file);
+      fd.append("account_id", accountId);
+      const res = await api.documentPreview(fd);
+      setRows(res.rows.map((r) => ({ ...r })));
+      setDocText(res.text || "");
+    } catch (e) { toast(e.message, "error"); } finally { setBusy(false); }
+  }
+
   async function commit() {
     setBusy(true);
     try {
@@ -168,7 +183,15 @@ export default function Import() {
 
   return (
     <div>
-      <h1>{t("import.title")}</h1>
+      <div className="page-head">
+        <h1>{t("import.title")}</h1>
+        <div className="seg" style={{ width: 240, marginBottom: 0 }}>
+          <button type="button" className={`seg-btn${mode === "csv" ? " active" : ""}`}
+                  onClick={() => { setMode("csv"); setRows(null); setRaw(null); setDocText(""); }}>{t("import.modeCsv")}</button>
+          <button type="button" className={`seg-btn${mode === "doc" ? " active" : ""}`}
+                  onClick={() => { setMode("doc"); setRows(null); setRaw(null); setDocText(""); }}>{t("import.modeDoc")}</button>
+        </div>
+      </div>
 
       {/* Schritt 1: Konto + Datei */}
       <div className="card">
@@ -182,11 +205,20 @@ export default function Import() {
           </label>
           <label className="grow">
             <span>{t("import.file")}</span>
-            <input ref={fileRef} type="file" accept=".csv,text/csv,text/plain"
+            <input ref={fileRef} type="file"
-                   onChange={(e) => { setFile(e.target.files?.[0] || null); setRaw(null); setRows(null); }} />
+                   accept={mode === "doc" ? ".pdf,application/pdf,image/*" : ".csv,text/csv,text/plain"}
+                   onChange={(e) => { setFile(e.target.files?.[0] || null); setRaw(null); setRows(null); setDocText(""); }} />
           </label>
         </div>
-        <div className="row">
+        {mode === "doc" && (
+          <>
+            <button className="btn primary" onClick={loadDoc} disabled={busy || !file}>
+              <Icon name="search" size={15} /> {t("import.readDoc")}
+            </button>
+            <p className="muted small" style={{ marginTop: "var(--sp-2)" }}>{t("import.docHint")}</p>
+          </>
+        )}
+        <div className="row" style={{ display: mode === "csv" ? undefined : "none" }}>
           {presets.length > 0 && (
             <label className="grow">
               <span>{t("import.preset")}</span>
@@ -208,7 +240,8 @@ export default function Import() {
         </div>
       </div>
 
-      {/* Schritt 2: Format */}
+      {/* Schritt 2: Format (nur CSV) */}
+      {mode === "csv" && (
       <div className="card">
         <h2>{t("import.step2")}</h2>
         <div className="row">
@@ -266,6 +299,17 @@ export default function Import() {
           <Icon name="search" size={15} /> {t("import.readFile")}
         </button>
       </div>
+      )}
+
+      {/* Extrahierter Text (PDF/Bild) */}
+      {mode === "doc" && docText && (
+        <div className="card">
+          <details>
+            <summary className="muted">{t("import.extractedText")}</summary>
+            <pre className="doc-text">{docText}</pre>
+          </details>
+        </div>
+      )}
 
       {/* Schritt 3: Spaltenzuordnung */}
       {raw && (