File size: 5,909 Bytes
e5ee651 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 | """
Document Understanding OCR
Extract structured invoice fields from OCR text with transparent confidence checks.
"""
from pathlib import Path
import re
import pandas as pd
import plotly.express as px
import streamlit as st
st.set_page_config(page_title="Document Understanding OCR", page_icon="📄", layout="wide")
def load_shared_css() -> None:
current_dir = Path(__file__).resolve().parent
candidates = [
current_dir / "shared" / "styles.css",
current_dir.parent / "shared" / "styles.css",
]
css_path = next(path for path in candidates if path.exists())
st.markdown(f"<style>{css_path.read_text(encoding='utf-8')}</style>", unsafe_allow_html=True)
load_shared_css()
SAMPLE_TEXT = """INVOICE
Vendor: Northwind Robotics GmbH
Invoice No: INV-2026-0418
Date: 2026-04-18
Due Date: 2026-05-18
Bill To: Atlas Manufacturing
Description Qty Unit Price Amount
Vision sensor calibration 3 420.00 1260.00
Edge gateway support 2 310.00 620.00
Subtotal: 1880.00
Tax: 357.20
Total: 2237.20 EUR
"""
FIELD_PATTERNS = {
"vendor": [r"Vendor:\s*(.+)", r"From:\s*(.+)"],
"invoice_number": [r"Invoice\s*(?:No|#|Number):\s*([A-Z0-9\-]+)"],
"invoice_date": [r"Date:\s*(\d{4}-\d{2}-\d{2})", r"Date:\s*(\d{1,2}/\d{1,2}/\d{4})"],
"due_date": [r"Due Date:\s*(\d{4}-\d{2}-\d{2})", r"Payment Due:\s*(.+)"],
"customer": [r"Bill To:\s*(.+)", r"Customer:\s*(.+)"],
"subtotal": [r"Subtotal:\s*([0-9,.]+)"],
"tax": [r"Tax:\s*([0-9,.]+)"],
"total": [r"Total:\s*([0-9,.]+)\s*([A-Z]{3})?"],
}
def first_match(text: str, patterns):
for pattern in patterns:
match = re.search(pattern, text, re.IGNORECASE)
if match:
groups = [group for group in match.groups() if group]
return " ".join(groups).strip(), 0.92 if len(groups[0]) > 2 else 0.68
return "", 0.0
def extract_line_items(text: str) -> pd.DataFrame:
rows = []
for line in text.splitlines():
match = re.search(r"^(.+?)\s+(\d+)\s+([0-9,.]+)\s+([0-9,.]+)$", line.strip())
if match:
description, qty, unit_price, amount = match.groups()
rows.append({
"description": description.strip(),
"quantity": int(qty),
"unit_price": float(unit_price.replace(",", "")),
"amount": float(amount.replace(",", "")),
})
return pd.DataFrame(rows)
def extract_document(text: str):
fields = []
payload = {}
for name, patterns in FIELD_PATTERNS.items():
value, confidence = first_match(text, patterns)
payload[name] = value
fields.append({"field": name, "value": value or "not found", "confidence": confidence})
items = extract_line_items(text)
if not items.empty:
computed_total = items["amount"].sum()
payload["line_item_count"] = int(len(items))
payload["computed_line_total"] = round(float(computed_total), 2)
field_df = pd.DataFrame(fields)
completeness = round((field_df["confidence"] > 0).mean() * 100, 1)
average_confidence = round(field_df["confidence"].mean() * 100, 1)
return payload, field_df, items, completeness, average_confidence
st.markdown("""
<div class="hero">
<div class="hf-badge">Document AI</div>
<h1>📄 Document Understanding OCR</h1>
<p>Turn noisy OCR text into auditable invoice JSON, field confidence, and line-item checks.</p>
<div class="pill-row">
<span class="hf-chip">Schema extraction</span>
<span class="hf-chip">Confidence audit</span>
<span class="hf-chip">HF model-ready</span>
</div>
</div>
""", unsafe_allow_html=True)
with st.sidebar:
st.markdown("### Technique")
st.info("This Space demonstrates the post-OCR layer: regex/schema extraction, confidence scoring, validation, and export. A production Space can swap the text input for Donut, LayoutLM, or TrOCR inference on Hugging Face.")
min_confidence = st.slider("Review threshold", 0.0, 1.0, 0.70, 0.05)
left, right = st.columns([1.1, 0.9])
with left:
text = st.text_area("OCR text", value=SAMPLE_TEXT, height=360)
with right:
st.markdown("### What This Space Proves")
st.markdown("""
- Extracts structured fields from unstructured document text.
- Separates evidence and confidence instead of pretending OCR is perfect.
- Produces JSON that can feed a dataset, API, or human review queue.
""")
payload, field_df, items, completeness, average_confidence = extract_document(text)
metric_cols = st.columns(3)
metric_cols[0].metric("Field completeness", f"{completeness}%")
metric_cols[1].metric("Average confidence", f"{average_confidence}%")
metric_cols[2].metric("Line items", int(payload.get("line_item_count", 0)))
tab1, tab2, tab3 = st.tabs(["Extraction", "Validation", "Export"])
with tab1:
st.dataframe(field_df, use_container_width=True, hide_index=True)
if not items.empty:
st.markdown("### Line Items")
st.dataframe(items, use_container_width=True, hide_index=True)
with tab2:
fig = px.bar(
field_df,
x="field",
y="confidence",
color="confidence",
range_y=[0, 1],
color_continuous_scale=["#b8a9d9", "#ffad7a", "#7accff"],
title="Field-level confidence",
)
st.plotly_chart(fig, use_container_width=True)
review = field_df[field_df["confidence"] < min_confidence]
if review.empty:
st.success("All extracted fields are above the review threshold.")
else:
st.warning("Fields below threshold need human review.")
st.dataframe(review, use_container_width=True, hide_index=True)
with tab3:
st.json(payload)
st.download_button(
"Download JSON",
data=pd.Series(payload).to_json(indent=2),
file_name="document_extraction.json",
mime="application/json",
use_container_width=True,
)
|