A pipeline of well-chosen libraries.

mime = magic.from_file(path)
sample = pdf.pages[:3]
is_scanned = avg_chars(sample) < 50

Detects MIME type and samples the first three pages to decide whether the PDF has selectable text or needs OCR.

img = preprocess(image)   # grayscale,
                          # upscale, sharpen
data = pytesseract.image_to_data(img)
text, conf = reconstruct(data)

Renders pages at 300 DPI, preprocesses for contrast and runs Tesseract with per-word confidence scoring.

if ocr_conf < 65:
    text, conf = htr.transcribe(
        image, model=KRAKEN_MODEL)

Falls back to handwriting recognition on low-confidence pages. Configurable via HTR_KRAKEN_MODEL env var.

for page in pdf.pages:
    for tbl in page.extract_tables():
        tables.append(clean(tbl))

Uses pdfplumber's lattice-based table detection on native PDFs and python-docx for Word documents.

doc = nlp(text)
entities = [(e.text, e.label_)
            for e in doc.ents]

Small spaCy model (~50 MB) loaded at startup; recognises persons, organisations, locations, dates and money.

patterns = {"email": r"...",
            "phone": r"...",
            "amount": r"[$€£]\s*\d+..."}
results = {k: re.findall(p, text)}

Plain compiled regex. Picks up emails, phones, URLs, dates and multi-currency amounts — plus invoice-specific fields when applicable.

for p in paragraphs:
    if short(p) and (
       p.isupper() or p.endswith(":")):
        headings.append(p)

Cheap heading detection by paragraph shape — no ML required for what is essentially a typography problem.

tfidf = TfidfVectorizer(
    stop_words="english"
).fit_transform(sentences)
top = sents[scores.argsort()[-5:]]

Extractive summary: score sentences by sum of TF-IDF weights, return the top five in original order.

ident = LanguageIdentifier
    .from_modelstring(model,
                      norm_probs=True)
lang, p = ident.classify(text)

Normalised 0–1 confidence rather than raw log-probability, so the score you see in the UI is actually meaningful.