scripts/build_dataset.py

"""Turn crawled raw files into a clean dataset (pages.jsonl + chunks.jsonl).



Inputs : data/raw/*.md (crawled docx), data/files/*.pdf (crawled attachments)

Outputs: data/dataset/pages.jsonl, data/dataset/chunks.jsonl, data/dataset/stats.json

"""

import json
import re
import sys
from pathlib import Path

for _s in (sys.stdout, sys.stderr):
    try:
        _s.reconfigure(encoding="utf-8", errors="replace")
    except Exception:  # noqa: BLE001
        pass

ROOT = Path(__file__).resolve().parent.parent
RAW = ROOT / "data" / "raw"
FILES = ROOT / "data" / "files"
OUT = ROOT / "data" / "dataset"
TREE = ROOT / "data" / "wiki_tree.json"

FRONTMATTER = re.compile(r"^---\n(.*?)\n---\n", re.S)
FIGURE = re.compile(r"<figure[^>]*>.*?</figure>", re.S)
FIGURE_SRC = re.compile(r'<source\s+name="([^"]+)"')
RESIDUAL_TAG = re.compile(r"</?(?:title|sub-page-list|p|fragment|source|figure)[^>]*/?>", re.S)
HEADING = re.compile(r"^(#{1,6})\s+(.*)$", re.M)
WS3 = re.compile(r"\n{3,}")
SPACES = re.compile(r"[ \t]+\n")

SKIP_TITLES = {"欢迎使用知识库 / Welcome to Wiki", "管理员手册 / Administrator Manual", "成员手册 / Member Manual"}

# Whole branches that are not course material and would only pollute retrieval:
# Feishu's own product manual, the wiki's member/admin handbook, the 高考 recall papers
# that were parked in the same workspace, and the guestbook. Dropping them is recorded
# in the dataset card; nothing from a course is removed.
EXCLUDE_COURSES = {
    "管理员手册 / Administrator Manual",
    "成员手册 / Member Manual",
    "历年考生回忆版新高考卷",
    "期末休息区",
}


def parse_frontmatter(text):
    m = FRONTMATTER.match(text)
    meta = {}
    if m:
        for line in m.group(1).splitlines():
            if ":" in line:
                k, v = line.split(":", 1)
                meta[k.strip()] = v.strip()
        text = text[m.end():]
    return meta, text


def clean_markdown(text):
    text = RESIDUAL_TAG.sub("", text)
    text = FIGURE.sub(lambda m: "[附件: %s]" % (FIGURE_SRC.search(m.group(0)).group(1) if FIGURE_SRC.search(m.group(0)) else "附件"), text)
    text = re.sub(r"<[^>]{1,80}>", "", text)
    text = SPACES.sub("\n", text)
    text = WS3.sub("\n\n", text)
    return text.strip()


def split_sections(text):
    """Split markdown into (heading_path, body) sections."""
    sections = []
    current = [""]
    buf = []
    for line in text.splitlines():
        m = HEADING.match(line)
        if m:
            if buf:
                sections.append((list(current), "\n".join(buf).strip()))
                buf = []
            level = len(m.group(1))
            title = m.group(2).strip()
            current = current[: level - 1] + [title]
        else:
            buf.append(line)
    if buf:
        sections.append((list(current), "\n".join(buf).strip()))
    return [(h, b) for h, b in sections if b]


def chunk_text(text, size=700, overlap=100):
    text = text.strip()
    if len(text) <= size:
        return [text] if text else []
    chunks = []
    start = 0
    while start < len(text):
        end = start + size
        if end < len(text):
            cut = max(text.rfind("\n\n", start, end), text.rfind("\n", start, end), text.rfind("。", start, end))
            if cut > start + size // 2:
                end = cut + 1
        chunks.append(text[start:end].strip())
        if end >= len(text):
            break
        start = max(end - overlap, start + 1)
    return [c for c in chunks if c]


def is_low_information(chunk, title, course, section):
    """True for chunks whose text is nothing but the title/path repeated.



    The wiki stores a heading for every page, and the chunker sometimes keeps only that

    heading, which produces a "chunk" like "3.5➡️3.0 数据库系统 / DataBase System" repeated.

    Those match any query that mentions the course name and push real content down.

    """
    junk = chunk
    for s in (title, course, section or ""):
        for part in str(s).split("/"):
            part = part.strip()
            if len(part) > 3:
                junk = junk.replace(part, "")
    junk = re.sub(r"[\s\-—·・/|*#>\[\]()（）。，、：:]", "", junk)
    return len(junk) < 20


def pdf_to_text(path):
    try:
        from pypdf import PdfReader
    except ImportError:
        return None
    try:
        reader = PdfReader(str(path))
    except Exception:  # noqa: BLE001
        return None
    pages = []
    for i, page in enumerate(reader.pages):
        try:
            pages.append(page.extract_text() or "")
        except Exception:  # noqa: BLE001
            pages.append("")
    return "\n\n".join(pages).strip()


def main():
    OUT.mkdir(parents=True, exist_ok=True)
    tree = json.load(open(TREE, encoding="utf-8")) if TREE.exists() else []
    meta_by_token = {n.get("obj_token"): n for n in tree}
    meta_by_file = {n.get("file"): n for n in tree if n.get("file")}

    pages, chunks = [], []

    # --- docx pages ---
    for md_path in sorted(RAW.glob("*.md")):
        raw = md_path.read_text(encoding="utf-8", errors="replace")
        meta, body = parse_frontmatter(raw)
        title = meta.get("title") or md_path.stem
        if title in SKIP_TITLES:
            continue
        node = meta_by_file.get(md_path.name) or meta_by_token.get(meta.get("obj_token")) or {}
        path_list = node.get("path") or (meta.get("path", "").split("/") if meta.get("path") else [title])
        course = path_list[0] if path_list else title
        if course in EXCLUDE_COURSES:
            continue
        body = clean_markdown(body)
        if not body:
            continue
        page_id = f"docx::{node.get('node_token') or md_path.stem}"
        pages.append(
            {
                "id": page_id,
                "title": title,
                "course": course,
                "path": "/".join(path_list),
                "url": meta.get("url") or node.get("url") or "",
                "source_type": "docx",
                "obj_token": meta.get("obj_token", ""),
                "n_chars": len(body),
                "text": body,
            }
        )
        for heading, sec in split_sections(body):
            for ci, ctext in enumerate(chunk_text(sec)):
                if len(ctext) < 20 or is_low_information(ctext, title, course, "/".join([h for h in heading if h])):
                    continue
                chunks.append(
                    {
                        "id": f"{page_id}::{len(chunks)}",
                        "page_id": page_id,
                        "title": title,
                        "course": course,
                        "path": "/".join(path_list),
                        "section": "/".join([h for h in heading if h]) or title,
                        "url": meta.get("url") or node.get("url") or "",
                        "source_type": "docx",
                        "n_chars": len(ctext),
                        "text": ctext,
                    }
                )

    # --- pdf attachments ---
    pdf_meta = {n.get("file_path"): n for n in tree if n.get("file_path")}
    for pdf_path in sorted(FILES.glob("*.pdf")):
        rel = str(pdf_path.relative_to(ROOT)).replace("\\", "/")
        node = pdf_meta.get(rel) or {}
        text = pdf_to_text(pdf_path)
        if not text or len(text) < 50:
            continue
        path_list = node.get("path") or [pdf_path.name]
        page_id = f"pdf::{pdf_path.stem}"
        pages.append(
            {
                "id": page_id,
                "title": node.get("title") or pdf_path.name,
                "course": path_list[0] if path_list else "附件",
                "path": "/".join(path_list),
                "url": node.get("url") or "",
                "source_type": "pdf",
                "obj_token": node.get("obj_token", ""),
                "n_chars": len(text),
                "text": text,
            }
        )
        for ctext in chunk_text(text, size=700, overlap=100):
            if len(ctext) < 20 or is_low_information(ctext, pdf_path.name, path_list[0] if path_list else "", pdf_path.name):
                continue
            chunks.append(
                {
                    "id": f"{page_id}::{len(chunks)}",
                    "page_id": page_id,
                    "title": node.get("title") or pdf_path.name,
                    "course": path_list[0] if path_list else "附件",
                    "path": "/".join(path_list),
                    "section": pdf_path.name,
                    "url": node.get("url") or "",
                    "source_type": "pdf",
                    "n_chars": len(ctext),
                    "text": ctext,
                }
            )

    def dump(name, rows):
        p = OUT / name
        with p.open("w", encoding="utf-8") as f:
            for r in rows:
                f.write(json.dumps(r, ensure_ascii=False) + "\n")
        return p

    dump("pages.jsonl", pages)
    dump("chunks.jsonl", chunks)

    # Node tree (flattened: path joined by " / ") for the HF dataset viewer.
    tree_rows = [
        {
            "node_token": n.get("node_token"),
            "title": n.get("title"),
            "obj_type": n.get("obj_type"),
            "path": " / ".join(n.get("path") or []),
            "course": (n.get("path") or [None])[0],
            "url": n.get("url", ""),
            "file": n.get("file", ""),
            "file_path": n.get("file_path", ""),
        }
        for n in tree
        if ((n.get("path") or [None])[0] not in EXCLUDE_COURSES)
    ]
    dump("tree.jsonl", tree_rows)

    stats = {
        "n_pages": len(pages),
        "n_chunks": len(chunks),
        "n_nodes": len(tree_rows),
        "excluded_courses": sorted(EXCLUDE_COURSES),
        "n_chars": sum(p["n_chars"] for p in pages),
        "courses": sorted({p["course"] for p in pages}),
        "source_types": {t: sum(1 for p in pages if p["source_type"] == t) for t in {p["source_type"] for p in pages}},
    }
    (OUT / "stats.json").write_text(json.dumps(stats, ensure_ascii=False, indent=2), encoding="utf-8")
    print(json.dumps(stats, ensure_ascii=False, indent=2))
    print(f"\n-> {OUT}")


if __name__ == "__main__":
    main()