parse_on_hf.py

"""Parse all PDFs + PPTXs from balade/chatbot-assets dataset to markdown.
Runs on HF Jobs infra (better CPU than local laptop).

Expected mount: /data (read-only, from balade/chatbot-assets dataset)
Uploads results back to the same dataset via API."""

import os, sys, time, json, traceback
from pathlib import Path

HF_TOKEN = os.environ.get("HF_TOKEN")
if not HF_TOKEN:
    print("FATAL: HF_TOKEN not set", flush=True)
    sys.exit(1)

from huggingface_hub import HfApi

REPO = "balade/chatbot-assets"
DATA_DIR = Path("/data")
OUT_DIR = "parsed_assets"
api = HfApi(token=HF_TOKEN)

def log(msg):
    print(f"[{time.strftime('%H:%M:%S')}] {msg}", flush=True)

def upload_markdown(remote_path, md_content):
    if remote_path.lower().endswith(".pdf"):
        out_path = remote_path[:-4] + ".md"
    elif remote_path.lower().endswith(".pptx"):
        out_path = remote_path[:-5] + ".md"
    else:
        out_path = remote_path + ".md"
    out_path = f"{OUT_DIR}/{out_path}"
    api.upload_file(
        repo_id=REPO, repo_type="dataset",
        path_or_fileobj=md_content.encode("utf-8"),
        path_in_repo=out_path,
    )
    return out_path

def parse_text_pdf(pdf_path):
    import pymupdf4llm
    return pymupdf4llm.to_markdown(str(pdf_path))

def parse_image_pdf(pdf_path):
    import easyocr, fitz
    reader = easyocr.Reader(["id"], gpu=False)
    doc = fitz.open(pdf_path)
    pages, total = [], doc.page_count
    for i in range(total):
        page = doc[i]
        pix = page.get_pixmap(dpi=200)
        img = pix.tobytes("png")
        result = reader.readtext(img)
        text = "\n".join([r[1] for r in result])
        pages.append(f"## Page {i+1}\n\n{text}\n")
        if (i + 1) % 5 == 0 or i == total - 1:
            log(f"  OCR: {i+1}/{total} pages")
    doc.close()
    return "\n\n".join(pages)

def parse_pptx(pptx_path):
    from pptx import Presentation
    prs = Presentation(pptx_path)
    pages = []
    for i, slide in enumerate(prs.slides, 1):
        texts = []
        for shape in slide.shapes:
            if shape.has_text_frame:
                for para in shape.text_frame.paragraphs:
                    t = para.text.strip()
                    if t:
                        texts.append(t)
            if shape.has_table:
                rows = []
                for row in shape.table.rows:
                    cells = [cell.text.strip() for cell in row.cells]
                    rows.append(" | ".join(cells))
                texts.append("\n".join(rows))
        pages.append(f"## Slide {i}\n\n" + "\n\n".join(texts))
    return "\n\n".join(pages)

def classify_pdf(pdf_path):
    import fitz
    doc = fitz.open(pdf_path)
    chars = sum(len(page.get_text()) for page in doc)
    n = doc.page_count
    doc.close()
    return chars > 200, chars, n

def main():
    log("Starting parsing job on HF infra")
    log(f"Data dir: {DATA_DIR} (exists: {DATA_DIR.exists()})")

    if not DATA_DIR.exists():
        log(f"FATAL: {DATA_DIR} not mounted")
        sys.exit(1)

    # Walk all files
    all_files = sorted(DATA_DIR.rglob("*"))
    pdfs = [f for f in all_files if f.suffix.lower() == ".pdf"]
    pptxs = [f for f in all_files if f.suffix.lower() == ".pptx"]
    log(f"Found {len(pdfs)} PDFs + {len(pptxs)} PPTXs")

    # Check what's already parsed
    try:
        parsed_set = set()
        for item in api.list_repo_tree(repo_id=REPO, repo_type="dataset", path=OUT_DIR, recursive=True):
            parsed_set.add(item.path)
        log(f"Already parsed: {len(parsed_set)} files")
    except Exception as e:
        log(f"No existing parsed files: {e}")
        parsed_set = set()

    results = {"text": 0, "image": 0, "pptx": 0, "errors": [], "skipped": 0}
    timing = {"text": 0.0, "image": 0.0, "pptx": 0.0}

    for local_path in pdfs + pptxs:
        rel = str(local_path.relative_to(DATA_DIR))
        suf = local_path.suffix.lower()

        if suf == ".pdf":
            expected = f"{OUT_DIR}/{rel[:-4]}.md"
        else:
            expected = f"{OUT_DIR}/{rel[:-5]}.md"

        if expected in parsed_set:
            log(f"SKIP (exists): {rel}")
            results["skipped"] += 1
            continue

        log(f"\n{'='*60}")
        log(f"Processing: {rel}")

        t0 = time.time()
        try:
            if suf == ".pdf":
                is_text, chars, pages = classify_pdf(local_path)
                if is_text:
                    log(f"  TYPE: text PDF ({pages} pg, {chars:,} chars)")
                    md = parse_text_pdf(local_path)
                    timing["text"] += time.time() - t0
                    results["text"] += 1
                else:
                    log(f"  TYPE: image PDF ({pages} pg) - OCR...")
                    md = parse_image_pdf(local_path)
                    timing["image"] += time.time() - t0
                    results["image"] += 1
            else:
                log(f"  TYPE: PPTX")
                md = parse_pptx(local_path)
                timing["pptx"] += time.time() - t0
                results["pptx"] += 1

            out_path = upload_markdown(rel, md)
            log(f"  Done: {out_path} ({len(md):,} chars, {time.time()-t0:.1f}s)")

        except Exception as e:
            log(f"  ERROR: {e}")
            traceback.print_exc()
            results["errors"].append(rel)

    log(f"\n{'='*60}")
    log(f"RESULTS")
    log(f"  Text PDFs: {results['text']}")
    log(f"  Image PDFs (OCR): {results['image']}")
    log(f"  PPTX: {results['pptx']}")
    log(f"  Skipped: {results['skipped']}")
    log(f"  Errors: {len(results['errors'])}")
    for e in results["errors"]:
        log(f"    ✗ {e}")
    log(f"  Timing - text: {timing['text']:.1f}s, OCR: {timing['image']:.1f}s, PPTX: {timing['pptx']:.1f}s")

    api.upload_file(
        repo_id=REPO, repo_type="dataset",
        path_or_fileobj=json.dumps({
            "timestamp": time.strftime("%Y-%m-%dT%H:%M:%SZ"), **results
        }, indent=2).encode(),
        path_in_repo=f"{OUT_DIR}/_parse_report.json",
    )
    log("Report uploaded. Done!")

if __name__ == "__main__":
    main()