runtime/jevlike/ingest.py

"""Document ingestion: PDFs, scans, photos, HTML and text -> clean text for SystemOne.

The model reads text only, so every other input is converted first:

    from jevlike.ingest import to_text, predict_file
    r = to_text("10-K.pdf")                 # IngestResult(text, pages, method_per_page, seconds, warnings, ...)
    r = to_text("receipt.jpg", ocr="rapidocr")
    r = to_text(["p1.png", "p2.png"])      # one document given as several page images
    r = to_text(pdf_bytes)                  # bytes: the kind is sniffed from the content
    res = predict_file(model, "10-K.pdf", {"topic": Choice(...), ...})   # ingest + SystemOne.predict

    python -m jevlike.ingest FILE [--ocr rapidocr] [--json] [--questions spec.json --checkpoint DIR]

What each input goes through
----------------------------
* **PDF**: the text layer of every page via PyMuPDF (content-stream order, which keeps the columns
  of multi-column layouts apart). Ruled tables found by pdfplumber are replaced by markdown rows
  (``| Revenue | 1,234 | 1,456 |``) so numbers stay next to their row and column headers. A page
  with too little extractable text (< ``min_chars`` non-space characters, mostly unmapped glyphs,
  or a page-sized image with only a stamp of text) is rendered at ``dpi`` (300) and OCR'd.
* **Images** (.png .jpg .jpeg .tif/.tiff (all frames) .webp .bmp .gif): OCR with tesseract
  (default) or RapidOCR (``ocr="rapidocr"``, PP-OCRv4 via onnxruntime). Before OCR
  (``preprocess=True``): EXIF orientation, grayscale, perspective correction of a photographed
  page (largest bright quadrilateral on a darker background), a 3x3 median filter when the paper
  shows pixel noise, illumination flattening when the background is uneven, deskew (+-5 deg), and
  upscaling when the text is small (x-height < ~20 px; tesseract only). RapidOCR's bundled Chinese
  PP-OCRv4 recognizer drops spaces between English words; set ``JEVLIKE_RAPIDOCR_REC`` to an
  English recognizer (``en_PP-OCRv3_rec_infer.onnx`` from the ``SWHL/RapidOCR`` HF repo, 9 MB).
  ``JEVLIKE_OCR_THREADS`` limits RapidOCR's onnxruntime threads.
* **HTML**: scripts, styles, navigation, headers/footers outside the article, sidebars, cookie and
  share widgets are dropped; the ``<article>`` / ``<main>`` content is kept as markdown-ish text
  (``#`` headings, paragraphs, ``-`` list items, ``|`` table rows).
* **Text / Markdown**: decoded (UTF-8, else cp1252/latin-1) and whitespace-normalized.

Layout handling shared by the text-layer and OCR paths: side-by-side narrow blocks whose lines
line up (a borderless table that OCR / the PDF writer emits column by column) are zipped back
into rows; cells on one line separated by a wide gap are joined with `` | ``; lines repeated at
the top or bottom of most pages (running headers, "Page 3 of 9") are removed from multi-page
documents. Cleanup: ligatures / soft hyphens / zero-width characters, de-hyphenation of words
split across lines, re-flowing of wrapped lines into paragraphs (table rows and list items are
kept as lines), OCR noise lines without any letter or digit, whitespace normalization.
"""
from __future__ import annotations

import argparse
import inspect
import io
import json
import os
import re
import shutil
import statistics
import subprocess
import sys
import time
import unicodedata
from dataclasses import asdict, dataclass, field
from pathlib import Path
from typing import Any, Mapping, Optional, Sequence, Union

import numpy as np

Source = Union[str, os.PathLike, bytes, bytearray]

IMAGE_EXT = {".png", ".jpg", ".jpeg", ".tif", ".tiff", ".webp", ".bmp", ".gif"}
HTML_EXT = {".html", ".htm", ".xhtml"}
TEXT_EXT = {".txt", ".md", ".markdown", ".text", ".csv", ".tsv", ".rst", ".log"}
KINDS = ("pdf", "image", "html", "text")
OCR_ENGINES = ("tesseract", "rapidocr")


@dataclass
class IngestResult:
    text: str                       # the whole document, pages separated by a blank line
    pages: list[str]                # text per page (one entry for html / text)
    method_per_page: list[str]      # "text", "text+tables(2)", "ocr:tesseract", "ocr:rapidocr", "html", "plain"
    seconds: float                  # wall time of the whole conversion
    warnings: list[str] = field(default_factory=list)
    kind: str = ""                  # "pdf" | "image" | "html" | "text"
    ocr_confidence: list[Optional[float]] = field(default_factory=list)  # mean word confidence 0-100 per page (None: not OCR'd / unknown)

    def to_dict(self, with_text: bool = True) -> dict:
        d = asdict(self)
        if not with_text:
            d.pop("text"), d.pop("pages")
            d["chars"] = len(self.text)
            d["n_pages"] = len(self.pages)
        return d


# ================================================================ public API

def to_text(src: "Source | Sequence[Source]", kind: Optional[str] = None, *, ocr: str = "tesseract",
            lang: str = "eng", dpi: int = 300, min_chars: int = 30, tables: bool = True,
            preprocess: bool = True, strip_headers: bool = True, max_pages: Optional[int] = None,
            ocr_timeout: float = 120.0) -> IngestResult:
    """Convert a document to text.

    ``src``: a path, raw bytes, or a list of them (the pages of one document, e.g. page scans).
    ``kind``: "pdf" | "image" | "html" | "text" (default: from the extension, else sniffed from
    the content). ``ocr``: "tesseract" or "rapidocr". ``dpi``: render resolution for PDF pages
    that need OCR. ``min_chars``: a PDF page with fewer non-space characters in its text layer is
    OCR'd. ``tables``: markdown rows for ruled PDF tables (pdfplumber). ``preprocess``: image
    cleanup before OCR (see module doc). ``strip_headers``: drop running headers/footers of
    multi-page documents. ``max_pages``: stop after this many pages (with a warning).
    """
    if ocr not in OCR_ENGINES:
        raise ValueError(f"ocr must be one of {OCR_ENGINES}, got {ocr!r}")
    if kind is not None and kind not in KINDS:
        raise ValueError(f"kind must be one of {KINDS}, got {kind!r}")
    t0 = time.perf_counter()
    opts = dict(ocr=ocr, lang=lang, dpi=dpi, min_chars=min_chars, tables=tables, preprocess=preprocess,
                max_pages=max_pages, ocr_timeout=ocr_timeout)
    srcs = list(src) if isinstance(src, (list, tuple)) else [src]
    if not srcs:
        raise ValueError("nothing to ingest")
    pages: list[str] = []
    methods: list[str] = []
    confs: list[Optional[float]] = []
    warnings: list[str] = []
    kinds: list[str] = []
    for s in srcs:
        data, name = _read(s)
        k = kind or detect_kind(data, name)
        kinds.append(k)
        if k == "pdf":
            p, m, c, w = _ingest_pdf(data, **opts)
        elif k == "image":
            p, m, c, w = _ingest_image(data, **opts)
        elif k == "html":
            p, m, c, w = [html_to_text(_decode(data))], ["html"], [None], []
        else:
            p, m, c, w = [normalize_text(_decode(data))], ["plain"], [None], []
        if len(srcs) > 1 and w:
            w = [f"{name or 'input'}: {x}" for x in w]
        pages += p
        methods += m
        confs += c
        warnings += w
        if max_pages and len(pages) >= max_pages:
            if len(pages) > max_pages or s is not srcs[-1]:
                warnings.append(f"stopped after max_pages={max_pages} pages")
            pages, methods, confs = pages[:max_pages], methods[:max_pages], confs[:max_pages]
            break
    if strip_headers and len(pages) >= 2:
        pages, removed = strip_repeated_lines(pages)
        if removed:
            warnings.append(f"removed {removed} repeated header/footer line(s)")
    text = "\n\n".join(p for p in pages if p.strip())
    if not text.strip():
        warnings.append("no text extracted")
    ocr_conf = [c for c in confs if c is not None]
    if ocr_conf and min(ocr_conf) < 60:
        warnings.append(f"low OCR confidence on some pages (min {min(ocr_conf):.0f}/100): check the text")
    k = kinds[0] if len(set(kinds)) == 1 else "mixed"
    return IngestResult(text, pages, methods, time.perf_counter() - t0, warnings, k, confs)


def predict_file(model, path: "Source | Sequence[Source]", questions: Mapping[str, Any], *,
                 ingest_kwargs: Optional[dict] = None, **predict_kwargs) -> dict:
    """Ingest ``path`` with :func:`to_text` and answer ``questions`` with ``model.predict``.

    When the text is longer than the model's state budget (``max_len`` minus the question budget)
    and the caller did not choose ``long_state``, it is read in overlapping windows
    (``long_state="chunk"``) instead of being cut in the middle. The result is the usual predict
    output plus ``"ingest"``: {pages, method_per_page, seconds, warnings, chars, kind,
    ocr_confidence, long_state}.
    """
    ing = to_text(path, **(ingest_kwargs or {}))
    text = ing.text if ing.text.strip() else "(no text could be extracted from the document)"
    kw = dict(predict_kwargs)
    params = _params(model.predict)
    if "long_state" in params and "long_state" not in kw and _too_long(model, text):
        kw["long_state"] = "chunk"
    res = model.predict(text, questions, **kw)
    info = ing.to_dict(with_text=False)
    info["long_state"] = kw.get("long_state", getattr(model, "long_state", None))
    res["ingest"] = info
    return res


def detect_kind(data: bytes, name: Optional[str] = None) -> str:
    """"pdf" | "image" | "html" | "text" from the file extension, else from the leading bytes."""
    ext = Path(name).suffix.lower() if name else ""
    if ext == ".pdf":
        return "pdf"
    if ext in IMAGE_EXT:
        return "image"
    if ext in HTML_EXT:
        return "html"
    if ext in TEXT_EXT:
        return "text"
    head = bytes(data[:1024])
    if head.lstrip().startswith(b"%PDF") or b"%PDF-" in head[:1024]:
        return "pdf"
    if (head.startswith(b"\x89PNG") or head.startswith(b"\xff\xd8\xff") or head[:4] in (b"II*\x00", b"MM\x00*")
            or (head.startswith(b"RIFF") and head[8:12] == b"WEBP") or head.startswith(b"GIF8")
            or head.startswith(b"BM")):
        return "image"
    low = head.lower().lstrip()
    if low.startswith((b"<!doctype html", b"<html", b"<?xml")) or b"<html" in low or b"<body" in low:
        return "html"
    return "text"


# ================================================================ helpers: io

def _read(s: Source) -> tuple[bytes, Optional[str]]:
    if isinstance(s, (bytes, bytearray)):
        return bytes(s), None
    p = Path(s)
    return p.read_bytes(), p.name


def _decode(data: bytes) -> str:
    if data.startswith(b"\xef\xbb\xbf"):
        data = data[3:]
    for enc in ("utf-8", "cp1252"):
        try:
            return data.decode(enc)
        except UnicodeDecodeError:
            pass
    return data.decode("latin-1")


def _params(fn) -> set[str]:
    try:
        return set(inspect.signature(fn).parameters)
    except (TypeError, ValueError):
        return set()


def _too_long(model, text: str) -> bool:
    ser = getattr(model, "ser", None)
    tok, max_len = getattr(ser, "tok", None), getattr(model, "max_len", None)
    if tok is None or not max_len:
        return False
    budget = max_len - getattr(ser, "head_max_len", 0) - 2
    # cheap bound first: ~>=2 chars per token for English
    if len(text) < 2 * budget:
        return False
    return len(tok(text, add_special_tokens=False)["input_ids"]) > budget


# ================================================================ text cleanup

_LIGATURES = {"ff": "ff", "fi": "fi", "fl": "fl", "ffi": "ffi", "ffl": "ffl", "ſt": "st",
              "st": "st"}
_INVISIBLE = dict.fromkeys(map(ord, "­​‌‍⁠"), None)
_BULLET = re.compile(r"^\s*(?:[-*•▪●‣⁃◦]|\(?\d{1,3}[.)]|\(?[a-z][.)])\s+")
_ROW = re.compile(r"(^\s*\|)|( \| )")


def _is_row(line: str) -> bool:
    return bool(_ROW.search(line))


def normalize_text(s: str) -> str:
    """Whitespace/unicode normalization only (no re-flow): for text and markdown inputs."""
    s = s.replace("\r\n", "\n").replace("\r", "\n").translate(_INVISIBLE)
    for k, v in _LIGATURES.items():
        s = s.replace(k, v)
    s = s.replace(" ", " ").replace("\t", "    ")
    lines = [re.sub(r"[ \f\v]+$", "", ln) for ln in s.split("\n")]
    s = "\n".join(lines)
    s = re.sub(r"\n{3,}", "\n\n", s)
    return s.strip()


def clean_text(s: str, reflow: bool = True) -> str:
    """Cleanup for extracted / OCR'd text: unicode fixes, de-hyphenation, paragraph re-flow,
    removal of lines without any letter or digit (OCR specks), whitespace normalization."""
    s = s.replace("\r\n", "\n").replace("\r", "\n").translate(_INVISIBLE)
    for k, v in _LIGATURES.items():
        s = s.replace(k, v)
    s = s.replace(" ", " ").replace("\t", " ").replace("\f", "\n")
    lines = []
    for ln in s.split("\n"):
        ln = re.sub(r" {2,}", " ", ln).strip()
        if ln and not re.search(r"[^\W_]", ln):  # no letter/digit at all: rule, speck, "|", "~ ."
            continue
        lines.append(ln)
    s = "\n".join(lines)
    # de-hyphenate a word split across a line break: "infor-\nmation" -> "information"
    s = re.sub(r"([A-Za-z]{2,})[-‐‑]\n[ ]*([a-z]{2,})", r"\1\2", s)
    paras = re.split(r"\n\s*\n", s)
    out = []
    for p in paras:
        pl = [ln for ln in p.split("\n") if ln.strip()]
        if not pl:
            continue
        if not reflow:
            out.append("\n".join(pl))
            continue
        merged: list[str] = []
        for ln in pl:
            keep_break = (not merged or _is_row(ln) or _is_row(merged[-1]) or _BULLET.match(ln)
                          or ln.startswith("#") or merged[-1].startswith("#"))
            if keep_break:
                merged.append(ln)
            else:
                merged[-1] = merged[-1] + " " + ln
        out.append("\n".join(merged))
    return "\n\n".join(out).strip()


def _hf_key(line: str) -> str:
    """Header/footer comparison key: lowercase letters and '#' for each digit run."""
    return re.sub(r"\d+", "#", re.sub(r"[^a-z0-9]+", "", line.lower()))


def _similar(a: str, b: str) -> bool:
    if a == b:
        return True
    if min(len(a), len(b)) < 6:
        return False
    import difflib

    return difflib.SequenceMatcher(None, a, b, autojunk=False).ratio() >= 0.8


def strip_repeated_lines(pages: list[str], edge: int = 2, min_frac: float = 0.5) -> tuple[list[str], int]:
    """Remove running headers / footers: a line among the first or last ``edge`` lines of a page
    that (digits masked, fuzzy-matched so OCR noise does not hide it) occurs at a page edge on
    >= ``min_frac`` of the pages (and on >= 2 pages)."""
    if len(pages) < 2:
        return pages, 0
    split = [p.split("\n") for p in pages]
    edges = []
    for lines in split:
        idx = [i for i, ln in enumerate(lines) if ln.strip()]
        e = sorted(set(idx[:edge] + idx[-edge:]))
        # headers/footers are short lines; long ones are (re-flowed) body text
        edges.append([(i, _hf_key(lines[i])) for i in e if 0 < len(_hf_key(lines[i])) and len(lines[i]) <= 100])
    need = max(2, int(np.ceil(min_frac * len(pages))))
    drop: list[set[int]] = [set() for _ in pages]
    for p, ed in enumerate(edges):
        for i, k in ed:
            hits = sum(any(_similar(k, k2) for _, k2 in edges[q]) for q in range(len(pages)) if q != p) + 1
            if hits >= need:
                drop[p].add(i)
        if len(drop[p]) >= sum(1 for ln in split[p] if ln.strip()):
            drop[p] = set()  # never empty a page (duplicate or near-empty pages)
    removed = sum(len(d) for d in drop)
    if not removed:
        return pages, 0
    out = []
    for lines, d in zip(split, drop):
        keep = [ln for i, ln in enumerate(lines) if i not in d]
        out.append(re.sub(r"\n{3,}", "\n\n", "\n".join(keep)).strip())
    return out, removed


# ================================================================ layout: lines -> text

@dataclass
class _Line:
    x0: float
    y0: float
    x1: float
    y1: float
    text: str
    par: int = 0     # paragraph id inside its block; a change starts a new paragraph
    blk: int = 0     # source block id (set when blocks are zipped)

    @property
    def h(self) -> float:
        return max(1.0, self.y1 - self.y0)

    @property
    def yc(self) -> float:
        return (self.y0 + self.y1) / 2


def _numeric(text: str) -> bool:
    t = re.sub(r"\s", "", text)
    return bool(t) and sum(ch.isdigit() or ch in "$%.,()-–—+/:€£" for ch in t) >= 0.6 * len(t)


_NUMBER = re.compile(r"^[(\-\u2013\u2212+$\u20ac\u00a3]*\d[\d.,]*%?\)?[%,;:]?$")


def _cell_breaks(texts: list[str], gaps: list[float], h: float, prose: bool = False) -> list[bool]:
    """Which horizontal gaps inside one text line separate table cells (joined with " | ").
    Inside a prose block only very wide gaps (> 4 word heights) count, so stretched justified
    lines stay prose. Elsewhere a gap must be wide (> 1.5) and then either very wide, next to a
    number token, or much wider than the line's narrowest gap."""
    if not gaps:
        return []
    gmin = min(gaps)
    out = []
    for j, g in enumerate(gaps):
        if prose:
            out.append(g > 4.0 * h)
            continue
        a, b = _NUMBER.match(texts[j].strip()), _NUMBER.match(texts[j + 1].strip())
        if a and b:  # two numbers side by side: adjacent numeric cells (padding ~1 word height)
            out.append(g > 1.0 * h)
            continue
        out.append(g > 1.5 * h and (g > 4.0 * h or bool(a or b) or (len(gaps) > 1 and g > 3.0 * max(gmin, 0.05 * h))))
    return out


def _nwords(text: str) -> int:
    return sum(1 for w in text.split() if w != "|")


def _prose_lines(word_counts: list[int]) -> bool:
    """A block of running text: median >= 5 words per line over >= 3 lines."""
    return len(word_counts) >= 3 and statistics.median(word_counts) >= 5


def _y_overlap(a0, a1, b0, b1) -> float:
    return max(0.0, min(a1, b1) - max(a0, b0))


def _bbox(lines: list[_Line]) -> tuple[float, float, float, float]:
    return (min(l.x0 for l in lines), min(l.y0 for l in lines), max(l.x1 for l in lines), max(l.y1 for l in lines))


def _aligned(a: list[_Line], b: list[_Line]) -> float:
    """Fraction of lines of the smaller block with a y-matching line in the other block."""
    small, big = (a, b) if len(a) <= len(b) else (b, a)
    hit = sum(any(abs(s.yc - t.yc) < 0.5 * min(s.h, t.h) for t in big) for s in small)
    return hit / max(1, len(small))


def _zip_blocks(blocks: list[list[_Line]], page_w: float) -> list[tuple[list[_Line], bool]]:
    """Merge runs of side-by-side blocks whose lines line up (table columns emitted one after
    another) into one table block. Returns [(lines, is_table)] in reading order."""
    out: list[tuple[list[_Line], bool]] = []
    group: list[list[_Line]] = []

    def narrow(b):
        return max(l.x1 - l.x0 for l in b) < 0.4 * page_w

    def prose(b):  # columns of running text (>= 5 words per line) are never zipped into rows
        return statistics.median(_nwords(l.text) for l in b) >= 5

    def flush():
        if not group:
            return
        if len(group) == 1:
            out.append((group[0], False))
        else:
            lines = []
            for i, b in enumerate(group):
                for l in b:
                    l.blk = i
                    lines.append(l)
            out.append((lines, True))
        group.clear()

    for b in blocks:
        if group:
            prev = group[-1]
            gx0, gy0, gx1, gy1 = _bbox([l for g in group for l in g])
            bx0, by0, bx1, by1 = _bbox(b)
            px0, py0, px1, py1 = _bbox(prev)
            hmin = min(statistics.median(l.h for l in b), statistics.median(l.h for l in prev))
            right = bx0 >= px1 - 0.5 * hmin
            ov = _y_overlap(gy0, gy1, by0, by1) / max(1.0, min(gy1 - gy0, by1 - by0))
            both_single = len(b) == 1 and len(prev) == 1
            numeric = sum(_numeric(l.text) for l in b) >= 0.6 * len(b)
            if right and ov >= 0.5 and _aligned(prev, b) >= 0.6 and not (prose(b) or prose(prev)) and (
                    both_single or numeric or (narrow(b) and narrow(prev))):
                group.append(b)
                continue
        flush()
        group.append(b)
    flush()
    return out


def _join_row(cells: list[_Line], table: bool, prose: bool = False) -> str:
    cells = sorted(cells, key=lambda l: l.x0)
    h = min(l.h for l in cells)
    gaps = [b.x0 - a.x1 for a, b in zip(cells, cells[1:])]
    # separate line boxes on one row: cells of a zipped table (different source blocks), or a
    # wide gap by the rule of _cell_breaks (boxes are measured in box heights, ~1.4 word heights)
    breaks = _cell_breaks([c.text for c in cells], gaps, 0.7 * h, prose=prose and not table)
    txt = cells[0].text
    for a, b, br in zip(cells, cells[1:], breaks):
        txt += (" | " if (table and a.blk != b.blk) or br else " ") + b.text
    return txt


def _rows(lines: list[_Line]) -> list[str]:
    """Cluster lines by vertical position into rows (for zipped table blocks)."""
    hmed = statistics.median(l.h for l in lines)
    rows: list[list[_Line]] = []
    for l in sorted(lines, key=lambda l: l.yc):
        if rows and abs(l.yc - statistics.mean(x.yc for x in rows[-1])) < 0.5 * hmed:
            rows[-1].append(l)
        else:
            rows.append([l])
    return [_join_row(r, True) for r in rows]


def _seq_text(lines: list[_Line]) -> str:
    """Lines in reading order; a line that continues the previous one on the same baseline (a
    table cell to its right) is appended to it; paragraph changes give a blank line."""
    out: list[str] = []
    cur: list[_Line] = []
    prose = _prose_lines([_nwords(l.text) for l in lines])
    for l in lines:
        if cur:
            last = cur[-1]
            same_row = (l.x0 >= last.x1 - 0.2 * l.h and
                        _y_overlap(last.y0, last.y1, l.y0, l.y1) >= 0.5 * min(last.h, l.h))
            if same_row:
                cur.append(l)
                continue
            out.append(_join_row(cur, False, prose))
            if l.par != last.par:
                out.append("")
        cur = [l]
    if cur:
        out.append(_join_row(cur, False, prose))
    return "\n".join(out)


def layout_text(blocks: list[list[_Line]], page_w: float) -> str:
    parts: list[str] = []
    for lines, is_table in _zip_blocks([b for b in blocks if b], page_w):
        p = "\n".join(_rows(lines)) if is_table else _seq_text(lines)
        if not p.strip():
            continue
        if parts and _is_row(parts[-1].rsplit("\n", 1)[-1]) and _is_row(p.split("\n", 1)[0]):
            parts[-1] += "\n" + p   # consecutive table rows stay one block
        else:
            parts.append(p)
    return "\n\n".join(parts)


def markdown_table(rows: list[list[Optional[str]]]) -> str:
    rows = [[re.sub(r"\s+", " ", c or "").strip() for c in r] for r in rows]
    rows = [r for r in rows if any(r)]
    if not rows:
        return ""
    n = max(len(r) for r in rows)
    rows = [r + [""] * (n - len(r)) for r in rows]
    # drop columns that are empty everywhere (merged-cell artefacts)
    keep = [j for j in range(n) if any(r[j] for r in rows)]
    rows = [[r[j] for j in keep] for r in rows]
    # no "| --- |" separator row: it carries no information for the model
    return "\n".join("| " + " | ".join(r) + " |" for r in rows)


# ================================================================ PDF

def _ingest_pdf(data: bytes, *, ocr, lang, dpi, min_chars, tables, preprocess, max_pages, ocr_timeout):
    import fitz  # PyMuPDF

    warnings: list[str] = []
    try:
        doc = fitz.open(stream=data, filetype="pdf")
    except Exception as e:
        raise ValueError(f"cannot open PDF: {type(e).__name__}: {e}") from None
    if doc.needs_pass and not doc.authenticate(""):
        raise ValueError("PDF is password-protected")
    plumber = None
    if tables:
        try:
            import pdfplumber
            plumber = pdfplumber.open(io.BytesIO(data))
        except Exception as e:  # pragma: no cover - broken/unsupported file for pdfplumber
            warnings.append(f"table detection unavailable: {type(e).__name__}")
    pages, methods, confs = [], [], []
    n = doc.page_count if not max_pages else min(doc.page_count, max_pages)
    try:
        for i in range(n):
            page = doc[i]
            ppage = plumber.pages[i] if plumber is not None and i < len(plumber.pages) else None
            text, n_tab = _pdf_page_text(page, ppage, warnings)
            reason = _needs_ocr(page, text, min_chars)
            if reason:
                img = _render(page, dpi)
                otext, conf = _ocr_image(img, ocr=ocr, lang=lang, dpi=dpi, preprocess=preprocess,
                                         timeout=ocr_timeout)
                if len(otext.strip()) >= len(text.strip()):
                    pages.append(otext)
                    methods.append(f"ocr:{ocr}")
                    confs.append(conf)
                    continue
            pages.append(text)
            methods.append(f"text+tables({n_tab})" if n_tab else "text")
            confs.append(None)
    finally:
        if plumber is not None:
            plumber.close()
    if max_pages and doc.page_count > max_pages:
        warnings.append(f"PDF has {doc.page_count} pages; read the first {max_pages}")
    return pages, methods, confs, warnings


def _needs_ocr(page, text: str, min_chars: int) -> Optional[str]:
    body = re.sub(r"\s", "", text)
    if len(body) < min_chars:
        return "little text"
    bad = body.count("�") + len(re.findall(r"\(cid:\d+\)", text)) * 6
    if bad > 0.2 * len(body):
        return "unmapped glyphs"
    if len(body) < 300:
        try:
            area = abs(page.rect)
            img_area = sum(abs(r) for info in page.get_images(full=True) for r in page.get_image_rects(info[0]))
            if area and img_area > 0.5 * area:
                return "image page with a thin text layer"
        except Exception:
            pass
    return None


def _render(page, dpi: int) -> np.ndarray:
    import fitz

    pix = page.get_pixmap(dpi=dpi, colorspace=fitz.csGRAY, alpha=False)
    a = np.frombuffer(pix.samples, dtype=np.uint8).reshape(pix.height, pix.stride)[:, : pix.width]
    return a.copy()


def _pdf_page_text(page, ppage, warnings: list[str]) -> tuple[str, int]:
    tabs: list[tuple[tuple[float, float, float, float], str]] = []
    if ppage is not None and page.rotation == 0:
        try:
            for t in ppage.find_tables():
                rows = t.extract()
                if len(rows) >= 2 and max(len(r) for r in rows) >= 2:
                    md = markdown_table(rows)
                    if md:
                        tabs.append((t.bbox, md))
        except Exception as e:  # pragma: no cover
            warnings.append(f"table detection failed on page {page.number + 1}: {type(e).__name__}")
    d = page.get_text("dict", flags=0)  # no images; content-stream order
    blocks: list[list[_Line]] = []
    placed: set[int] = set()
    items: list[Any] = []   # list of block line-lists or ("table", md)
    for b in d.get("blocks", []):
        if b.get("type", 0) != 0:
            continue
        cur: list[_Line] = []
        blk_prose = _prose_lines([len("".join(sp.get("text", "") for sp in li.get("spans", [])).split())
                                  for li in b.get("lines", [])])
        for li in b.get("lines", []):
            spans = [s for s in li.get("spans", []) if s.get("text", "").strip()]
            if not spans:
                continue
            x0, y0, x1, y1 = li["bbox"]
            cx, cy = (x0 + x1) / 2, (y0 + y1) / 2
            ti = next((k for k, (bb, _) in enumerate(tabs) if bb[0] - 1 <= cx <= bb[2] + 1 and bb[1] - 1 <= cy <= bb[3] + 1), None)
            if ti is not None:
                if cur:
                    items.append(cur)
                    cur = []
                if ti not in placed:
                    placed.add(ti)
                    items.append(("table", tabs[ti][1]))
                continue
            txt = spans[0]["text"]
            sgaps = [s2["bbox"][0] - s1["bbox"][2] for s1, s2 in zip(spans, spans[1:])]
            size = max(1.0, statistics.median(sp.get("size", 10) for sp in spans))
            breaks = _cell_breaks([sp["text"] for sp in spans], sgaps, 0.7 * size, prose=blk_prose)
            for sp, g, br in zip(spans[1:], sgaps, breaks):
                sep = " | " if br else ("" if g < 0.15 * size else " ")
                txt += sep + sp["text"]
            cur.append(_Line(x0, y0, x1, y1, re.sub(r"\s+", " ", txt).strip()))
        if cur:
            items.append(cur)
    for ti, (_, md) in enumerate(tabs):
        if ti not in placed:
            items.append(("table", md))
    # layout: consecutive line blocks go through layout_text; tables are inserted verbatim
    parts: list[str] = []
    run: list[list[_Line]] = []
    for it in items:
        if isinstance(it, tuple):
            if run:
                parts.append(layout_text(run, page.rect.width))
                run = []
            parts.append(it[1])
        else:
            run.append(it)
    if run:
        parts.append(layout_text(run, page.rect.width))
    return clean_text("\n\n".join(parts)), len(tabs)


# ================================================================ images + OCR

def _ingest_image(data: bytes, *, ocr, lang, dpi, min_chars, tables, preprocess, max_pages, ocr_timeout):
    from PIL import Image, ImageOps

    Image.MAX_IMAGE_PIXELS = 200_000_000
    warnings: list[str] = []
    try:
        im = Image.open(io.BytesIO(data))
    except Exception as e:
        raise ValueError(f"cannot open image: {type(e).__name__}: {e}") from None
    n_frames = getattr(im, "n_frames", 1)
    pages, methods, confs = [], [], []
    for f in range(n_frames if not max_pages else min(n_frames, max_pages)):
        if n_frames > 1:
            im.seek(f)
        fr = ImageOps.exif_transpose(im) if f == 0 else im
        src_dpi = None
        try:
            src_dpi = float(fr.info.get("dpi", (0, 0))[0]) or None
        except Exception:
            pass
        gray = np.asarray(fr.convert("L"))
        text, conf = _ocr_image(gray, ocr=ocr, lang=lang, dpi=src_dpi, preprocess=preprocess, timeout=ocr_timeout)
        pages.append(text)
        methods.append(f"ocr:{ocr}")
        confs.append(conf)
    return pages, methods, confs, warnings


def _ocr_image(gray: np.ndarray, *, ocr: str, lang: str, dpi: Optional[float], preprocess: bool,
               timeout: float) -> tuple[str, Optional[float]]:
    if preprocess:
        gray, scale = prepare_image(gray, upscale=ocr == "tesseract")  # RapidOCR rescales internally
        dpi = (dpi or 300) * scale if dpi else None
    if ocr == "rapidocr":
        blocks, conf = _rapidocr_blocks(gray)
    else:
        blocks, conf = _tesseract_blocks(gray, lang=lang, dpi=dpi, timeout=timeout)
    return clean_text(layout_text(blocks, gray.shape[1])), conf


# ---------------------------------------------------------------- preprocessing

def prepare_image(gray: np.ndarray, upscale: bool = True) -> tuple[np.ndarray, float]:
    """Photo/scan cleanup before OCR: perspective correction, denoising, illumination flattening,
    deskew and (``upscale``) enlarging small text. Returns (image, scale factor applied)."""
    import cv2

    g = gray
    g = _dewarp_page(g)
    g = _denoise(g)
    g = _flatten_illumination(g)
    g = _deskew(g)
    xh = _text_height(g) if upscale else None
    scale = 1.0
    if xh and xh < 20:
        scale = min(3.0, 26.0 / xh, 6000.0 / max(g.shape))
        if scale > 1.15:
            g = cv2.resize(g, None, fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)
        else:
            scale = 1.0
    return g, scale


def _small(g: np.ndarray, target: int = 1000) -> tuple[np.ndarray, float]:
    import cv2

    f = target / max(g.shape)
    if f >= 1:
        return g, 1.0
    return cv2.resize(g, None, fx=f, fy=f, interpolation=cv2.INTER_AREA), f


def _quad_angles_ok(q: np.ndarray, tol: float = 25.0) -> bool:
    """Every interior angle of the quadrilateral within 90 +- tol degrees (a page seen in
    perspective; a region cut off by a shadow edge or a fold is rejected)."""
    for i in range(4):
        a, b, c = q[i - 1], q[i], q[(i + 1) % 4]
        v1, v2 = a - b, c - b
        cos = float(np.dot(v1, v2) / (np.linalg.norm(v1) * np.linalg.norm(v2) + 1e-9))
        if abs(np.degrees(np.arccos(np.clip(cos, -1, 1))) - 90) > tol:
            return False
    return True


def _order_quad(q: np.ndarray) -> np.ndarray:
    s, d = q.sum(1), np.diff(q, axis=1).ravel()
    return np.array([q[np.argmin(s)], q[np.argmin(d)], q[np.argmax(s)], q[np.argmax(d)]], np.float32)


def _page_quad(s: np.ndarray) -> Optional[np.ndarray]:
    """Largest plausible page quadrilateral in a (downscaled) photo, from several binarizations:
    page-vs-background edges (robust to shadows falling across the page) and global thresholds."""
    import cv2

    H, W = s.shape
    blur = cv2.GaussianBlur(s, (5, 5), 0)
    masks = []
    edges = cv2.Canny(blur, 30, 90)
    masks.append(cv2.dilate(edges, np.ones((5, 5), np.uint8)))
    otsu, _ = cv2.threshold(blur, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
    for t in (otsu, 0.75 * otsu, 0.55 * otsu):
        _, th = cv2.threshold(blur, t, 255, cv2.THRESH_BINARY)
        masks.append(cv2.morphologyEx(th, cv2.MORPH_CLOSE, np.ones((9, 9), np.uint8)))
    border = 0.006 * max(H, W)
    for m in masks:  # priority order: edges first, then thresholds from strict to loose
        cnts, _ = cv2.findContours(m, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
        for c in sorted(cnts, key=cv2.contourArea, reverse=True)[:2]:
            hull = cv2.convexHull(c)
            area = cv2.contourArea(hull)
            if area < 0.25 * H * W or area > 0.97 * H * W:
                continue
            peri = cv2.arcLength(hull, True)
            for eps in (0.02, 0.03, 0.05):
                ap = cv2.approxPolyDP(hull, eps * peri, True)
                if len(ap) == 4:
                    q = _order_quad(ap.reshape(4, 2).astype(np.float32))
                    inside = bool(((q[:, 0] > border) & (q[:, 0] < W - 1 - border) &
                                   (q[:, 1] > border) & (q[:, 1] < H - 1 - border)).all())
                    if inside and _quad_angles_ok(q) and cv2.contourArea(q) >= 0.9 * area:
                        return q  # a corner on the frame border = leaked into the background
                    break
    return None


def _dewarp_page(g: np.ndarray) -> np.ndarray:
    """Perspective-correct a photographed page: the largest plausible page quadrilateral
    (>= 25% of the frame, near-right angles) that is clearly brighter than its surroundings."""
    import cv2

    s, f = _small(g)
    quad = _page_quad(s)
    if quad is None:
        return g
    mask = np.zeros_like(s)
    cv2.fillConvexPoly(mask, quad.astype(np.int32), 255)
    inside, outside = s[mask > 0], s[mask == 0]
    if outside.size < 0.03 * s.size or float(np.median(inside)) - float(np.median(outside)) < 40:
        return g
    tl, tr, br, bl = quad / f
    w = int(max(np.linalg.norm(tr - tl), np.linalg.norm(br - bl)))
    h = int(max(np.linalg.norm(bl - tl), np.linalg.norm(br - tr)))
    if w < 100 or h < 100:
        return g
    M = cv2.getPerspectiveTransform(np.array([tl, tr, br, bl], np.float32),
                                    np.array([[0, 0], [w - 1, 0], [w - 1, h - 1], [0, h - 1]], np.float32))
    return cv2.warpPerspective(g, M, (w, h), flags=cv2.INTER_CUBIC, borderValue=255)


def _denoise(g: np.ndarray) -> np.ndarray:
    """3x3 median filter when the paper shows pixel noise / specks (noisy scans, grainy photos)."""
    import cv2

    med = cv2.medianBlur(g, 3)
    thr, _ = cv2.threshold(med, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
    bg = med > thr + 10
    if bg.mean() < 0.2:
        return g
    noise = float(np.std(g[bg].astype(np.float32) - med[bg]))
    return med if noise > 3.0 else g


def _flatten_illumination(g: np.ndarray) -> np.ndarray:
    """Divide by a smooth background estimate when the paper brightness varies across the page."""
    import cv2

    s, f = _small(g, 600)
    bg = cv2.medianBlur(cv2.morphologyEx(s, cv2.MORPH_CLOSE, np.ones((15, 15), np.uint8)), 31)
    lo, hi = np.percentile(bg, [5, 95])
    if hi - lo < 30:
        return g
    bg_full = cv2.resize(bg, (g.shape[1], g.shape[0]), interpolation=cv2.INTER_LINEAR).astype(np.float32)
    out = np.clip(g.astype(np.float32) / np.maximum(bg_full, 1) * 245.0, 0, 255)
    return out.astype(np.uint8)


def _deskew(g: np.ndarray, max_angle: float = 5.0) -> np.ndarray:
    """Projection-profile deskew: the angle that maximizes the variance of row ink sums."""
    import cv2

    s, _ = _small(g, 1200)
    _, th = cv2.threshold(s, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
    if th.mean() < 0.5:  # (almost) blank
        return g
    h, w = th.shape
    center = (w / 2, h / 2)

    def score(a):
        M = cv2.getRotationMatrix2D(center, a, 1.0)
        r = cv2.warpAffine(th, M, (w, h), flags=cv2.INTER_NEAREST, borderValue=0)
        return float(np.var(r.sum(1, dtype=np.float64)))

    angles = np.arange(-max_angle, max_angle + 1e-6, 0.5)
    best = max(angles, key=score)
    fine = np.arange(best - 0.4, best + 0.41, 0.1)
    best = float(max(fine, key=score))
    if abs(best) < 0.25:
        return g
    M = cv2.getRotationMatrix2D((g.shape[1] / 2, g.shape[0] / 2), best, 1.0)
    return cv2.warpAffine(g, M, (g.shape[1], g.shape[0]), flags=cv2.INTER_CUBIC, borderValue=255)


def _text_height(g: np.ndarray) -> Optional[float]:
    """Median height of character-sized connected components (~ x-height), in pixels."""
    import cv2

    _, th = cv2.threshold(g, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
    n, _, stats, _ = cv2.connectedComponentsWithStats(th, connectivity=8)
    if n < 20:
        return None
    hs = stats[1:, cv2.CC_STAT_HEIGHT]
    ws = stats[1:, cv2.CC_STAT_WIDTH]
    ok = (hs >= 3) & (hs <= 150) & (ws <= 3 * hs) & (ws >= 1)
    if ok.sum() < 20:
        return None
    return float(np.median(hs[ok]))


# ---------------------------------------------------------------- engines

def _tesseract_blocks(gray: np.ndarray, lang: str = "eng", dpi: Optional[float] = None, psm: int = 3,
                      timeout: float = 120.0) -> tuple[list[list[_Line]], Optional[float]]:
    import cv2

    exe = shutil.which("tesseract")
    if not exe:
        raise RuntimeError("tesseract is not installed (apt install tesseract-ocr), or use ocr='rapidocr'")
    ok, png = cv2.imencode(".png", gray)
    env = dict(os.environ)
    env.setdefault("OMP_THREAD_LIMIT", "1")
    cmd = [exe, "stdin", "stdout", "-l", lang, "--psm", str(psm)]
    if dpi:
        cmd += ["--dpi", str(int(round(dpi)))]
    else:
        cmd += ["--dpi", "300"]
    cmd.append("tsv")
    r = subprocess.run(cmd, input=png.tobytes(), capture_output=True, timeout=timeout, env=env)
    if r.returncode != 0:
        raise RuntimeError(f"tesseract failed: {r.stderr.decode(errors='replace')[:300]}")
    rows = r.stdout.decode("utf-8", errors="replace").splitlines()[1:]
    lines: dict[tuple[int, int, int], list[tuple[int, int, int, int, float, str]]] = {}
    order: list[tuple[int, int, int]] = []
    confs, weights = [], []
    for row in rows:
        f = row.split("\t")
        if len(f) < 12 or f[0] != "5":
            continue
        txt = f[11].strip()
        conf = float(f[10])
        if not txt or conf < 0:
            continue
        key = (int(f[2]), int(f[3]), int(f[4]))
        if key not in lines:
            lines[key] = []
            order.append(key)
        x, y, w, h = int(f[6]), int(f[7]), int(f[8]), int(f[9])
        lines[key].append((x, y, x + w, y + h, conf, txt))
        confs.append(conf)
        weights.append(len(txt))
    blocks: dict[int, list[_Line]] = {}
    border: list[int] = []
    wc: dict[int, list[int]] = {}
    for key in order:
        wc.setdefault(key[0], []).append(len(lines[key]))
    for key in order:
        ws = _fix_pipes(sorted(lines[key], key=lambda w: w[0]))
        if not ws:
            continue
        hs = sorted(w[3] - w[1] for w in ws)
        hmed = hs[len(hs) // 2]
        txt = ws[0][5]
        breaks = _cell_breaks([w[5] for w in ws], [b[0] - a[2] for a, b in zip(ws, ws[1:])], hmed,
                              prose=_prose_lines(wc[key[0]]))
        for b, br in zip(ws[1:], breaks):
            txt += (" | " if br else " ") + b[5]
        ln = _Line(min(w[0] for w in ws), min(w[1] for w in ws), max(w[2] for w in ws), max(w[3] for w in ws),
                   txt, par=key[1])
        if key[0] not in blocks:
            blocks[key[0]] = []
            border.append(key[0])
        blocks[key[0]].append(ln)
    conf = float(np.average(confs, weights=weights)) if confs else None
    return [blocks[b] for b in border], conf


def _fix_pipes(ws: list[tuple]) -> list[tuple]:
    """Tesseract reads a capital I as "|": a lone "|" followed at a normal word gap by a lowercase
    word becomes "I"; other lone bars (table rules) are dropped (cells are split by gaps anyway)."""
    out = []
    for j, w in enumerate(ws):
        if w[5] in ("|", "||", "|,"):
            nxt = ws[j + 1] if j + 1 < len(ws) else None
            h = w[3] - w[1]
            if nxt is not None and nxt[5][:1].islower() and nxt[0] - w[2] < 1.2 * h:
                out.append(w[:5] + ("I" + w[5][1:].replace("|", ""),))
            continue
        out.append(w)
    return out


_RAPID = None


def _rapidocr_engine():
    global _RAPID
    if _RAPID is None:
        try:
            from rapidocr_onnxruntime import RapidOCR
        except ImportError as e:  # pragma: no cover
            raise RuntimeError("ocr='rapidocr' needs `pip install rapidocr_onnxruntime`") from e
        n = int(os.environ.get("JEVLIKE_OCR_THREADS", "0") or 0)
        # no dilation of the detector's text map: it merges adjacent lines of blurred text (photos)
        kw: dict[str, Any] = {"det_use_dilation": False}
        if n > 0:
            kw.update(intra_op_num_threads=n, inter_op_num_threads=1)
        rec = os.environ.get("JEVLIKE_RAPIDOCR_REC")
        if rec:
            kw["rec_model_path"] = rec
        _RAPID = RapidOCR(**kw)
    return _RAPID


def _rapidocr_blocks(gray: np.ndarray) -> tuple[list[list[_Line]], Optional[float]]:
    import cv2

    eng = _rapidocr_engine()
    # no 180-degree line classifier: pages are upright after deskew, and the classifier flips
    # (and then garbles or drops) some long text lines
    res, _ = eng(cv2.cvtColor(gray, cv2.COLOR_GRAY2BGR), use_cls=False)
    lines = []
    confs, weights = [], []
    for box, txt, score in res or []:
        txt = str(txt).strip()
        if not txt:
            continue
        pts = np.asarray(box, dtype=float)
        lines.append(_Line(pts[:, 0].min(), pts[:, 1].min(), pts[:, 0].max(), pts[:, 1].max(), txt))
        confs.append(float(score) * 100)
        weights.append(len(txt))
    conf = float(np.average(confs, weights=weights)) if confs else None
    if not lines:
        return [], conf
    return _xy_cut(lines), conf


def _gaps(iv: list[tuple[float, float]]) -> list[tuple[float, float]]:
    """Whitespace gaps between the union of 1-D intervals: [(width, midpoint)]."""
    iv = sorted(iv)
    out, end = [], iv[0][1]
    for a, b in iv[1:]:
        if a > end:
            out.append((a - end, (end + a) / 2))
        end = max(end, b)
    return out


def _interior(ls: list[_Line], a: float, b: float) -> bool:
    """A band [a, b] crossed by some boxes is a real gutter only if every crossing box has
    column text above and below it on both sides (a merged line inside two columns), not a
    full-width line at the top or bottom of the region (a title or a paragraph above a table:
    those are split off by a horizontal cut instead)."""
    cross = [l for l in ls if l.x0 < a and l.x1 > b]
    left = [l for l in ls if l.x1 <= a]
    right = [l for l in ls if l.x0 >= b]
    if not left or not right:
        return False
    for c in cross:
        for side in (left, right):
            if not (any(l.yc < c.y0 for l in side) and any(l.yc > c.y1 for l in side)):
                return False
    return True


def _xy_cut(lines: list[_Line], depth: int = 0) -> list[list[_Line]]:
    """Reading order for OCR line boxes without layout info (RapidOCR). Recursive XY-cut:
    split into columns at the widest vertical whitespace gap (>= 1 line height); otherwise split
    horizontally at the widest gap that lets one side split into columns (a full-width title
    above a two-column body, a paragraph above a table). Leaves are blocks read top-to-bottom,
    with a new paragraph at vertical gaps > 0.6 line height."""
    h = statistics.median(l.h for l in lines)

    def xgap(ls):
        """Widest vertical whitespace band; a band crossed by a few boxes (text merged across a
        narrow gutter by the detector) still counts when the rest of the region splits there."""
        if len(ls) < 2:
            return (0.0, None)
        g = _gaps([(l.x0, l.x1) for l in ls])
        best = max(g) if g else (0.0, None)
        tol = int(0.03 * len(ls))
        if tol >= 1:
            x0, x1 = int(min(l.x0 for l in ls)), int(max(l.x1 for l in ls)) + 1
            cov = np.zeros(x1 - x0 + 1, np.int32)
            for l in ls:
                cov[int(l.x0) - x0:int(l.x1) - x0 + 1] += 1
            low = cov <= tol
            run, start = 0, 0
            for i, v in enumerate(np.append(low, False)):
                if v:
                    if run == 0:
                        start = i
                    run += 1
                elif run:
                    if start > 0 and i < len(low) and run > best[0] and _interior(ls, x0 + start, x0 + i):
                        best = (float(run), x0 + start + run / 2)
                    run = 0
        return best

    def split(ls):
        """Column split of a region: a gutter >= 1 line height with running text (>= 4 words per
        box) on both sides. Table columns are not split: their rows are read across."""
        gx, px = xgap(ls)
        if gx < 1.0 * h or px is None:
            return None
        left = [l for l in ls if (l.x0 + l.x1) / 2 < px]
        right = [l for l in ls if (l.x0 + l.x1) / 2 >= px]
        if not left or not right:
            return None
        if min(statistics.median(_nwords(l.text) for l in side) for side in (left, right)) < 4:
            return None
        return left, right

    if len(lines) > 1 and depth < 60:
        cols = split(lines)
        if cols:
            return _xy_cut(cols[0], depth + 1) + _xy_cut(cols[1], depth + 1)
        for gy, py in sorted(_gaps([(l.y0, l.y1) for l in lines]), reverse=True):
            if gy < 0.1 * h:
                break
            top = [l for l in lines if l.yc < py]
            bot = [l for l in lines if l.yc >= py]
            if top and bot and (split(top) or split(bot)):
                return _xy_cut(top, depth + 1) + _xy_cut(bot, depth + 1)
    # OCR cell boxes on the same printed row can differ by a few pixels in y.
    # Cluster their baselines before sorting horizontally, or a number can be
    # emitted before its row label and break the table association.
    row_groups: list[list[_Line]] = []
    for line in sorted(lines, key=lambda l: l.yc):
        if row_groups and abs(line.yc - statistics.mean(x.yc for x in row_groups[-1])) < 0.5 * h:
            row_groups[-1].append(line)
        else:
            row_groups.append([line])
    leaf = [line for row in row_groups for line in sorted(row, key=lambda l: l.x0)]
    par, prev = 0, None
    for l in leaf:
        if prev is not None and l.y0 > prev.yc and (
                l.y0 - prev.y1 > 0.5 * min(l.h, prev.h) or max(l.h, prev.h) > 1.3 * min(l.h, prev.h)):
            par += 1
        l.par = par
        if prev is None or l.y1 > prev.y1:
            prev = l
    return [leaf]


# ================================================================ HTML

_DROP_TAGS = ["script", "style", "noscript", "iframe", "svg", "canvas", "form", "button", "nav", "aside",
              "template", "select", "input", "object", "embed", "dialog", "menu"]
_BOILER = re.compile(r"(^|[-_\s])(nav|navbar|menu|footer|sidebar|side-bar|cookie|consent|subscribe|newsletter|"
                     r"share|sharing|social|breadcrumb|advert|ads?|promo|related|recommend|comments?|popup|modal|"
                     r"banner|skip-link|signup|paywall)([-_\s]|$)", re.I)
_BLOCK_TAGS = ["h1", "h2", "h3", "h4", "h5", "h6", "p", "li", "table", "pre", "blockquote", "dt", "dd",
               "figcaption", "caption", "address"]


def html_to_text(html: str) -> str:
    """Main content of an HTML page as markdown-ish text (headings, paragraphs, lists, tables)."""
    from bs4 import BeautifulSoup

    try:
        soup = BeautifulSoup(html, "lxml")
    except Exception:
        soup = BeautifulSoup(html, "html.parser")
    title = soup.title.get_text(" ", strip=True) if soup.title else ""
    for t in soup(_DROP_TAGS):
        t.decompose()
    for t in soup.find_all(["header", "footer"]):
        if not t.find_parent(["article", "main"]):
            t.decompose()
    for t in soup.find_all(True):
        if t.attrs is None or t.name in ("html", "body", "article", "main"):
            continue
        ident = " ".join([t.get("id") or ""] + list(t.get("class") or []))
        role = (t.get("role") or "").lower()
        hidden = t.get("aria-hidden") == "true" or "display:none" in (t.get("style") or "").replace(" ", "")
        if hidden or role in ("navigation", "banner", "contentinfo", "complementary", "search") or (
                ident.strip() and _BOILER.search(ident)):
            t.decompose()
    arts = soup.find_all("article")
    root = (max(arts, key=lambda a: len(a.get_text())) if arts else None) or soup.find("main") or soup.body or soup
    out: list[str] = []
    for el in root.find_all(_BLOCK_TAGS):
        if el.find_parent(_BLOCK_TAGS) is not None:  # its text is emitted with the enclosing block
            continue
        if el.name == "table":
            rows = [[c.get_text(" ", strip=True) for c in tr.find_all(["th", "td"])] for tr in el.find_all("tr")]
            md = markdown_table(rows)
            if md:
                out.append(md)
            continue
        txt = el.get_text(" ", strip=True)
        txt = re.sub(r"\s+", " ", txt).strip()
        if not txt:
            continue
        if el.name in ("h1", "h2", "h3", "h4", "h5", "h6"):
            out.append("#" * int(el.name[1]) + " " + txt)
        elif el.name == "li":
            out.append("- " + txt)
        elif el.name == "pre":
            out.append(el.get_text().strip("\n"))
        else:
            out.append(txt)
    body = "\n\n".join(out)
    full = re.sub(r"\s+", " ", root.get_text(" ", strip=True))
    if len(re.sub(r"\s", "", body)) < 0.3 * len(re.sub(r"\s", "", full)):
        body = root.get_text("\n", strip=True)  # div soup without block tags
    if title and not re.search(r"^#\s", body, re.M) and title not in body[:500]:
        body = "# " + title + "\n\n" + body
    # join consecutive list items into one block
    body = re.sub(r"(^- .*)\n\n(?=- )", r"\1\n", body, flags=re.M)
    body = re.sub(r"(^- .*)\n\n(?=- )", r"\1\n", body, flags=re.M)
    return normalize_text(body)


# ================================================================ questions from JSON + CLI

def questions_from_json(spec: Mapping[str, Any]) -> dict:
    """Wire-format questions ({key: {"type", "instructions", "criteria"}}, as the HTTP API takes;
    a top-level {"questions": {...}} wrapper is accepted) -> public Choice / Score / Noul."""
    from jevlike.types import Choice, Noul, Score

    spec = spec.get("questions", spec) if isinstance(spec, Mapping) else spec
    out = {}
    for key, q in spec.items():
        t = q.get("type")
        if t == "choice":
            out[key] = Choice(q["instructions"], q["criteria"])
        elif t == "score":
            c = q["criteria"]
            levels = [f"{k}: {d}" if d else k for k, d in c.items()] if isinstance(c, dict) else list(c)
            out[key] = Score(q["instructions"], levels)
        elif t == "noul":
            out[key] = Noul(q["instructions"])
        else:
            raise ValueError(f"question {key!r}: unknown type {t!r}")
    return out


def main(argv=None) -> int:
    ap = argparse.ArgumentParser(description="Convert a document (PDF, image, HTML, text) to text and optionally "
                                             "answer SystemOne questions about it.")
    ap.add_argument("files", nargs="+", help="file(s); several files = pages of one document")
    ap.add_argument("--kind", choices=KINDS, default=None)
    ap.add_argument("--ocr", choices=OCR_ENGINES, default="tesseract")
    ap.add_argument("--lang", default="eng", help="tesseract language(s), e.g. eng+deu")
    ap.add_argument("--dpi", type=int, default=300)
    ap.add_argument("--min-chars", type=int, default=30)
    ap.add_argument("--no-tables", action="store_true")
    ap.add_argument("--no-preprocess", action="store_true")
    ap.add_argument("--keep-headers", action="store_true", help="do not strip repeated headers/footers")
    ap.add_argument("--max-pages", type=int, default=None)
    ap.add_argument("--json", action="store_true", help="print the IngestResult (and answers) as JSON")
    ap.add_argument("--questions", default=None, help="questions JSON (HTTP API format) to answer about the document")
    ap.add_argument("--checkpoint", default=None, help="checkpoint dir for --questions (e.g. checkpoints/jevlike-large-nf)")
    ap.add_argument("--device", default=None)
    args = ap.parse_args(argv)

    kw = dict(kind=args.kind, ocr=args.ocr, lang=args.lang, dpi=args.dpi, min_chars=args.min_chars,
              tables=not args.no_tables, preprocess=not args.no_preprocess, strip_headers=not args.keep_headers,
              max_pages=args.max_pages)
    src = args.files if len(args.files) > 1 else args.files[0]
    if args.questions:
        if not args.checkpoint:
            ap.error("--questions needs --checkpoint")
        from jevlike.predict import SystemOne

        qs = questions_from_json(json.loads(Path(args.questions).read_text()))
        model = SystemOne.load(args.checkpoint, device=args.device)
        res = predict_file(model, src, qs, ingest_kwargs=kw)
        print(json.dumps(res, indent=1, ensure_ascii=False, default=float))
        return 0
    r = to_text(src, **kw)
    if args.json:
        print(json.dumps(r.to_dict(), indent=1, ensure_ascii=False))
    else:
        sys.stdout.write(r.text + "\n")
        info = ", ".join(f"{m}" for m in r.method_per_page)
        print(f"[{r.kind}: {len(r.pages)} page(s) [{info}], {len(r.text)} chars, {r.seconds:.2f}s]"
              + "".join(f"\n[warning] {w}" for w in r.warnings), file=sys.stderr)
    return 0


if __name__ == "__main__":
    sys.exit(main())