#!/bin/bash
# OCRs image-only textbook PDFs into extracted-text/<board>/<class>/<name>.ocr.txt
# Language from filename suffix: *-en.pdf -> eng, *-ur.pdf -> urd (project tessdata).
# Renders pages at 300dpi grayscale with pdftoppm, OCRs pages in parallel, concatenates in order.
#
# Usage: ./ocr.sh <pdf> [<pdf> ...]      # specific files
#        ./ocr.sh --all-missing          # every corpus PDF with no extracted text yet
set -u
ROOT="$(cd "$(dirname "$0")/.." && pwd)"
export TESSDATA_PREFIX="$ROOT/scripts/tessdata"
OUT="$ROOT/extracted-text"
JOBS=8

ocr_one() {
  f="$1"
  rel=${f#"$ROOT"/}
  board=$(echo "$rel" | cut -d/ -f1); cdir=$(echo "$rel" | cut -d/ -f2)
  base=$(basename "$f" .pdf)
  case "$base" in
    # Sindhi is NOT Urdu. It uses an extended Arabic script with ~16 additional consonants
    # (ٻ ڀ ٺ ٽ ٿ ڏ ڊ ڍ ڌ ڪ ڳ ڱ ڻ ڇ ڄ ڃ), and the Urdu model has no glyphs for any of them —
    # it silently substitutes Urdu letters, producing text that looks plausible and is wrong.
    # A first pass on Sindh's Social Studies 6 returned ZERO Sindhi-specific letters.
    *-si) lang=snd ;;
    *-ur) lang=urd ;;
    *-un)
      # A file still marked unknown-medium would silently get the English model, which
      # produces unreadable noise for an Urdu book. Refuse rather than guess.
      echo "SKIP  $rel — medium is still 'un'. Run ./scripts/detect_medium.sh first."; return ;;
    *) lang=eng ;;
  esac
  mkdir -p "$OUT/$board/$cdir"
  txt="$OUT/$board/$cdir/$base.ocr.txt"
  [ -s "$txt" ] && { echo "SKIP (done): $txt"; return; }
  # Short template on purpose. BSD xargs caps a constructed -I argument at 255 bytes, and a
  # long book name pushed the two absolute page paths past it: xargs printed "command line
  # cannot be assembled, too long" to stderr and the loop below then wrote a 0-character file
  # that looked like a completed extraction. Running from inside the directory keeps the
  # substituted names to ~12 bytes each regardless of how the book is called.
  tmp=$(mktemp -d "${TMPDIR:-/tmp}/ocr.XXXXXX")
  echo "OCR [$lang] $rel ..."
  pdftoppm -gray -r 300 "$f" "$tmp/pg" 2>/dev/null
  ( cd "$tmp" && ls pg-*.p*m 2>/dev/null | sort | \
      xargs -P "$JOBS" -I{} tesseract "{}" "{}" -l "$lang" --psm 3 2>/dev/null )
  : > "$txt.part"
  for p in $(ls "$tmp"/pg-*.txt 2>/dev/null | sort); do
    pgnum=$(basename "$p" .txt | sed -E 's/\.p[gp]m$//; s/^pg-0*//')
    printf '\n===== PAGE %s =====\n' "$pgnum" >> "$txt.part"
    cat "$p" >> "$txt.part"
  done
  rm -rf "$tmp"
  chars=$(tr -d '[:space:]' < "$txt.part" | wc -c | tr -d ' ')
  # An empty result is a failure, not an extraction. Leaving a 0-byte .ocr.txt on disk makes
  # every later "is this book done?" check answer yes, which is how a silent xargs failure
  # turned into twenty books that looked processed and held nothing.
  if [ "$chars" -lt 200 ]; then
    rm -f "$txt.part"
    echo "  FAIL: produced only ${chars} chars — no text written, book still marked missing"
    return
  fi
  mv "$txt.part" "$txt"
  echo "  done: $chars chars -> $txt"
}

if [ "${1:-}" = "--all-missing" ]; then
  # other/ holds the scholarly library — published papers that already carry their own text
  # layer and are not part of the corpus. OCRing them wastes hours and produces nothing used.
  find "$ROOT" -name "*.pdf" -not -path "*/extracted-text/*" -not -path "*/other/*" -not -path "*/site/*" | sort | while read -r f; do
    base=$(basename "$f" .pdf)
    rel=${f#"$ROOT"/}; board=$(echo "$rel" | cut -d/ -f1); cdir=$(echo "$rel" | cut -d/ -f2)
    if [ ! -s "$OUT/$board/$cdir/$base.txt" ] && [ ! -s "$OUT/$board/$cdir/$base.ocr.txt" ]; then
      ocr_one "$f"
    fi
  done
else
  for f in "$@"; do ocr_one "$(cd "$(dirname "$f")" && pwd)/$(basename "$f")"; done
fi