Self-hosted, verifiable pipeline for extracting structured data (station,

date/time, price per litre, litres, discount, total) from German fuel
receipt photos: grayscale + mild bilateral filter + contrast stretch,
Tesseract OCR, then LLM structuring via a small local model (Qwen 2.5,
served by Ollama). Each step is independently runnable and inspectable
through numbered scripts against a shared tooling container.

tooling-container/Dockerfile

@@ -0,0 +1,20 @@
+FROM python:3.12-slim
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        tesseract-ocr \
+        tesseract-ocr-deu \
+    && rm -rf /var/lib/apt/lists/*
+
+RUN pip install --no-cache-dir \
+        opencv-python-headless==4.10.0.84 \
+        numpy==2.1.3 \
+        pillow==11.0.0 \
+        pytesseract==0.3.13 \
+        requests==2.32.3 \
+        odfpy==1.4.1
+
+COPY src/ /opt/app/src/
+RUN chmod -R a+rX /opt/app/src
+
+WORKDIR /data
+ENTRYPOINT []

tooling-container/VERSION

@@ -0,0 +1 @@
+0.1.0

tooling-container/src/compare_ground_truth.py

@@ -0,0 +1,94 @@
+#!/usr/bin/env python3
+"""Test-only tool: compare a directory of extracted.json results (produced
+by running 1_preprocess_image.sh/2_ocr_image.sh/3_extract_data.sh over a
+set of receipts) against the ground-truth spreadsheet, matching rows by
+date and reporting per-field accuracy."""
+import argparse
+import glob
+import json
+import os
+
+from odf.opendocument import load
+from odf.table import Table, TableRow, TableCell
+from odf.text import P
+
+
+def load_ground_truth(ods_path: str) -> dict:
+    doc = load(ods_path)
+    table = doc.spreadsheet.getElementsByType(Table)[0]
+    rows = table.getElementsByType(TableRow)
+    gt = {}
+    for r in rows[1:]:
+        cells = r.getElementsByType(TableCell)
+        vals = []
+        for c in cells:
+            ps = c.getElementsByType(P)
+            vals.append(" ".join(str(p) for p in ps))
+        if not vals or not vals[0]:
+            continue
+        date = vals[0]
+        try:
+            gt[date] = {
+                "time": vals[1],
+                "price_per_litre": float(vals[2].replace(",", ".")) if vals[2] else None,
+                "litres": float(vals[3].replace(",", ".")) if vals[3] else None,
+                "discount": float(vals[4].replace("€", "").replace(",", ".").strip()) if vals[4] else None,
+                "total_price": float(vals[5].replace("€", "").replace(",", ".").strip()) if vals[5] else None,
+            }
+        except (ValueError, IndexError):
+            continue
+    return gt
+
+
+def close(a, b, tol=0.05):
+    if a is None or b is None:
+        return a is None and b is None
+    return abs(abs(a) - abs(b)) <= tol
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("-gt", dest="ods_path", required=True, help="path to ground-truth .ods")
+    parser.add_argument("-results", dest="results_dir", required=True,
+                         help="directory containing one subdirectory per receipt, each with an extracted.json")
+    args = parser.parse_args()
+
+    gt = load_ground_truth(args.ods_path)
+
+    fields = ["price_per_litre", "litres", "discount", "total_price"]
+    field_ok = {k: 0 for k in fields}
+    field_total = {k: 0 for k in fields}
+    matched = 0
+    total = 0
+
+    for json_path in sorted(glob.glob(os.path.join(args.results_dir, "*", "extracted.json"))):
+        receipt = os.path.basename(os.path.dirname(json_path))
+        total += 1
+        with open(json_path, "r", encoding="utf-8") as f:
+            data = json.load(f)
+
+        truth = gt.get(data.get("date"))
+        if not truth:
+            print(f"{receipt}: NO DATE MATCH (extracted date={data.get('date')!r})")
+            continue
+
+        matched += 1
+        bad = []
+        for k in fields:
+            field_total[k] += 1
+            ok = close(data.get(k), truth[k])
+            if ok:
+                field_ok[k] += 1
+            else:
+                bad.append(k)
+        status = "OK" if not bad else f"WRONG FIELDS {bad}"
+        print(f"{receipt}: {status}")
+
+    print("\n=== SUMMARY ===")
+    print(f"Receipts matched to ground truth: {matched}/{total}")
+    for k in fields:
+        print(f"  {k}: {field_ok[k]}/{field_total[k]} correct")
+
+
+if __name__ == "__main__":
+    main()

tooling-container/src/extract.py

@@ -0,0 +1,101 @@
+#!/usr/bin/env python3
+"""Step 3: structure raw OCR text from a fuel receipt into JSON via a small
+text-only LLM served by Ollama on the host (reached over --network host)."""
+import argparse
+import json
+import os
+import time
+
+import requests
+
+OLLAMA_URL = os.environ.get("OLLAMA_URL", "http://127.0.0.1:11434/api/chat")
+MODEL = os.environ.get("OLLAMA_MODEL", "qwen2.5:3b")
+
+SCHEMA = {
+    "type": "object",
+    "properties": {
+        "petrol_station": {"type": ["string", "null"]},
+        "transaction_number": {"type": ["string", "null"]},
+        "date": {"type": ["string", "null"], "description": "DD.MM.YYYY"},
+        "time": {"type": ["string", "null"], "description": "HH:MM:SS"},
+        "price_per_litre": {"type": ["number", "null"]},
+        "litres": {"type": ["number", "null"]},
+        "discount": {"type": ["number", "null"]},
+        "total_price": {"type": ["number", "null"]},
+    },
+    "required": [
+        "petrol_station", "transaction_number", "date", "time",
+        "price_per_litre", "litres", "discount", "total_price",
+    ],
+}
+
+PROMPT_TEMPLATE = """The following is raw OCR text (with some noise/garbage from a security background pattern) extracted from a German fuel/petrol receipt. Extract these fields:
+
+- petrol_station: the name of the petrol station / company (e.g. "Shell", "Aral", "JET")
+- transaction_number: the transaction number (look for "Transaktions-Nr." specifically, not "Beleg-Nr."), as a string
+- date: the receipt date, formatted DD.MM.YYYY
+- time: the receipt time, formatted HH:MM:SS (24h)
+- litres and price_per_litre: these two numbers appear together on one line, in this order:
+  "<litres> l <price_per_litre> EUR/l" (the "l" litre symbol is often misread as "1" or "|").
+  For example "23,50 l 1,749 EUR/l" means litres=23.50 and price_per_litre=1.749. These are
+  two different numbers on the same line - don't confuse them with each other or with the
+  total price/Gesamtbetrag.
+- discount: total discount amount in EUR as a positive decimal number (0 if none). Look for
+  lines with words like "Rabatt", "Bonus", "Tankbonus", "Nachlass" followed by a negative EUR
+  amount, e.g. "Tankbonus -0,15 EUR" means discount=0.15 (drop the minus sign).
+- total_price: the final total amount paid, in EUR, as a decimal number
+
+Ignore garbled/nonsense words - they are OCR noise, not real data. If a field is not present, use null.
+
+OCR TEXT:
+---
+{ocr_text}
+---
+
+Respond with ONLY the JSON object, no other text.
+"""
+
+
+def structure(ocr_text: str) -> dict:
+    payload = {
+        "model": MODEL,
+        "messages": [
+            {"role": "user", "content": PROMPT_TEMPLATE.format(ocr_text=ocr_text)},
+        ],
+        "format": SCHEMA,
+        "stream": False,
+        "options": {"temperature": 0, "num_ctx": 3072},
+    }
+
+    start = time.time()
+    resp = requests.post(OLLAMA_URL, json=payload, timeout=300)
+    resp.raise_for_status()
+    result = resp.json()
+    elapsed = time.time() - start
+
+    content = result["message"]["content"]
+    parsed = json.loads(content)
+    parsed["_elapsed_seconds"] = round(elapsed, 1)
+    return parsed
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("-in", dest="in_path", required=True)
+    parser.add_argument("-out", dest="out_dir", required=True)
+    args = parser.parse_args()
+
+    os.makedirs(args.out_dir, exist_ok=True)
+    with open(args.in_path, "r", encoding="utf-8") as f:
+        ocr_text = f.read()
+
+    data = structure(ocr_text)
+
+    out_path = os.path.join(args.out_dir, "extracted.json")
+    with open(out_path, "w", encoding="utf-8") as f:
+        json.dump(data, f, indent=2, ensure_ascii=False)
+    print(out_path)
+
+
+if __name__ == "__main__":
+    main()

tooling-container/src/ocr.py

@@ -0,0 +1,34 @@
+#!/usr/bin/env python3
+"""Step 2: run Tesseract OCR (German) over a preprocessed receipt image,
+writing the raw recognized text to the output directory for inspection
+before it is handed to the structuring step."""
+import argparse
+import os
+
+import pytesseract
+from PIL import Image
+
+
+def ocr(path: str, out_dir: str) -> str:
+    os.makedirs(out_dir, exist_ok=True)
+    img = Image.open(path)
+    text = pytesseract.image_to_string(img, lang="deu", config="--psm 3")
+
+    out_path = os.path.join(out_dir, "ocr.txt")
+    with open(out_path, "w", encoding="utf-8") as f:
+        f.write(text)
+    return out_path
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("-in", dest="in_path", required=True)
+    parser.add_argument("-out", dest="out_dir", required=True)
+    args = parser.parse_args()
+
+    out_path = ocr(args.in_path, args.out_dir)
+    print(out_path)
+
+
+if __name__ == "__main__":
+    main()

tooling-container/src/preprocess.py

@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+"""Step 1: preprocess a receipt photo for OCR.
+
+Applies a mild edge-preserving smoothing pass (to take the edge off the
+receipt's dotted security background pattern) followed by a global
+percentile-based contrast stretch. Every intermediate stage is written to
+the output directory (01_gray, 02_bilateral, 03_final) so the effect of
+each step can be inspected on its own before trusting the final image to
+OCR.
+
+Heavier processing (strong denoising, local/CLAHE contrast, unsharp
+masking) was tried and tested worse: each one amplifies the security
+background pattern into OCR noise rather than suppressing it. This
+lighter two-step version was chosen because, across a set of receipts
+that the plain-grayscale baseline OCR'd incorrectly, it recovered data
+the baseline missed and never made a previously-correct read worse.
+"""
+import argparse
+import os
+
+import cv2
+import numpy as np
+from PIL import Image, ImageOps
+
+
+def load_grayscale(path: str) -> np.ndarray:
+    img = Image.open(path)
+    img = ImageOps.exif_transpose(img).convert("L")
+    return np.array(img)
+
+
+def autocontrast(gray: np.ndarray, low_pct=2, high_pct=98) -> np.ndarray:
+    lo, hi = np.percentile(gray, (low_pct, high_pct))
+    stretched = (gray.astype(np.float32) - lo) * 255.0 / (hi - lo)
+    return np.clip(stretched, 0, 255).astype(np.uint8)
+
+
+def preprocess(path: str, out_dir: str) -> str:
+    os.makedirs(out_dir, exist_ok=True)
+
+    gray = load_grayscale(path)
+    cv2.imwrite(os.path.join(out_dir, "01_gray.png"), gray)
+
+    bilateral = cv2.bilateralFilter(gray, d=5, sigmaColor=50, sigmaSpace=50)
+    cv2.imwrite(os.path.join(out_dir, "02_bilateral.png"), bilateral)
+
+    final = autocontrast(bilateral)
+    final_path = os.path.join(out_dir, "03_final.png")
+    cv2.imwrite(final_path, final)
+
+    return final_path
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("-in", dest="in_path", required=True)
+    parser.add_argument("-out", dest="out_dir", required=True)
+    args = parser.parse_args()
+
+    final_path = preprocess(args.in_path, args.out_dir)
+    print(final_path)
+
+
+if __name__ == "__main__":
+    main()