wikiembed.py

#!/usr/bin/env python3

import bz2
import json
import xml.etree.cElementTree as etree
import re
import sqlite3
from urllib.parse import quote
import mwparserfromhell
import requests
import os
import coremltools as ct
from transformers import AutoTokenizer
import numpy as np
from usearch.index import Index, Matches
import multiprocessing
from multiprocessing import Pool

def download_wiki_dump(lang, date):
    """Download the Wikipedia dump file."""
    local_filename = f"{lang}wiki-{date}-dump.xml.bz2"
    if os.path.isfile(local_filename):
        print(f"Using existing file: {local_filename}")
        return local_filename

    base_url = f"https://dumps.wikimedia.org/{lang}wiki/{date}/"
    dump_status_url = base_url + "dumpstatus.json"
    response = requests.get(dump_status_url)
    dump_info = json.loads(response.text)

    for filename, info in dump_info["jobs"]["articlesmultistreamdump"]["files"].items():
        if filename.endswith(".xml.bz2"):
            file_url = base_url + filename
            break
    else:
        raise ValueError("Couldn't find the multistream dump file")

    with requests.get(file_url, stream=True) as r:
        r.raise_for_status()
        with open(local_filename, 'wb') as f:
            for chunk in r.iter_content(chunk_size=8192):
                f.write(chunk)

    return local_filename

def parse_and_clean_wikicode(raw_content, parser, lang):
    """Strips formatting and unwanted sections from raw page content."""
    wikicode = parser.parse(raw_content)

    # Filters for magic words that are parser instructions -- e.g., __NOTOC__
    re_rm_magic = re.compile("__[A-Z]*__", flags=re.UNICODE)

    # Filters for file/image links.
    media_prefixes = "|".join(["File", "Image", "Media"]) # + MEDIA_ALIASES.get(lang, []))
    re_rm_wikilink = re.compile(f"^(?:{media_prefixes}):", flags=re.IGNORECASE | re.UNICODE)

    def rm_wikilink(obj):
        return bool(re_rm_wikilink.match(str(obj.title)))

    # Filters for references and tables
    def rm_tag(obj):
        return str(obj.tag) in {"ref", "table"}

    # Leave category links in-place but remove the category prefixes
    cat_prefixes = "|".join(["Category"]) # + CAT_ALIASES.get(language, []))
    re_clean_wikilink = re.compile(f"^(?:{cat_prefixes}):", flags=re.IGNORECASE | re.UNICODE)

    def is_category(obj):
        return bool(re_clean_wikilink.match(str(obj.title)))

    def clean_wikilink(obj):
        text = obj.__strip__()
        text = re.sub(re_clean_wikilink, "", text)
        obj.text = text

    def try_replace_obj(obj):
        try:
            clean_wikilink(obj)
        except ValueError:
            # For unknown reasons, objects are sometimes not found.
            pass

    def try_remove_obj(obj, section):
        try:
            section.remove(obj)
        except ValueError:
            # For unknown reasons, objects are sometimes not found.
            pass

    sections = []
    # Filter individual sections to clean.
    for section in wikicode.get_sections(flat=True, include_lead=True, include_headings=True):
        heading = section.filter_headings()[0].title.strip() if section.filter_headings() else "Lead"
        for obj in section.ifilter_wikilinks(recursive=True):
            if rm_wikilink(obj):
                try_remove_obj(obj, section)
            elif is_category(obj):
                try_replace_obj(obj)
        for obj in section.ifilter_tags(matches=rm_tag, recursive=True):
            try_remove_obj(obj, section)

        section_text = re.sub(re_rm_magic, "", section.strip_code().strip())
        if section_text.startswith(f"{heading}\n"):
            section_text = section_text[len(heading):].strip()

        sections.append((heading, section_text))
    return sections

def construct_url(title, lang):
    return f"https://{lang}.wikipedia.org/wiki/{quote(title)}"

def process_wiki_dump(filepath, lang):
    """Process the Wikipedia dump and yield cleaned articles."""
    total_pages = 0
    main_namespace_pages = 0
    articles_with_content = 0
    redirects = 0
    total_sections = 0
    try:
        with bz2.BZ2File(filepath) as f:
            context = etree.iterparse(f, events=("end",))
            for _, elem in context:
                if not elem.tag.endswith("page"):
                    continue

                total_pages += 1

                namespace = elem.tag[:-4]
                title = elem.find(f"./{namespace}title").text
                ns = elem.find(f"./{namespace}ns").text
                id_ = elem.find(f"./{namespace}id").text

                if ns != "0":  # Skip non-main namespace
                    elem.clear()
                    continue

                main_namespace_pages += 1

                red_ = elem.find(f"./{namespace}redirect")
                if red_ is not None:
                    redirects += 1
                    elem.clear()
                    continue

                raw_content = elem.find(f"./{namespace}revision/{namespace}text").text
                elem.clear()
                if raw_content is None or raw_content.strip() == "":
                    continue

                articles_with_content += 1

                try:
                    sections = parse_and_clean_wikicode(raw_content, parser=mwparserfromhell, lang=lang)
                except (mwparserfromhell.parser.ParserError) as e:
                    print(f"💥 mwparserfromhell ParseError: {e}")
                    return

                # if not text or text.strip() == "":
                #     return
                if not sections or len(sections) == 0:
                    return

                sections = [(heading, text.strip()) for heading, text in sections if len(text.strip()) > 1 and heading != "References" and heading != text.strip()]

                total_sections += len(sections)

                url = construct_url(title, lang)

                yield {
                    "article_id": id_,
                    "url": url,
                    "title": title,
                    "sections": sections
                }

                elem.clear()

                if articles_with_content % 1000 == 0:
                    print(f"Processed article {articles_with_content}: {title} {id_}")
    finally:
        print(f"Total pages processed: {total_pages}")
        print(f"Pages in main namespace: {main_namespace_pages}")
        print(f"Redirects skipped: {redirects}")
        print(f"Articles with content: {articles_with_content}")

def tokenize(text):
    return tokenizer(
        text,
        add_special_tokens=True,
        max_length=512,
        padding='max_length',
        truncation=True,
        return_attention_mask=True,
        return_tensors='np'
    )

def tokenize_for_chunking(text):
    res = tokenizer(
        text,
        add_special_tokens=True,
        padding='do_not_pad',
        truncation=False,
        return_tensors='np'
    )
    return res['input_ids'].flatten()

def embed(text):
    result = tokenize(text)
    token_ids = result['input_ids'].astype(np.float32)
    mask = result['attention_mask'].astype(np.float32)
    # print(f"Tokens: {token_ids}")
    # print(f"Mask: {mask}")
    predictions = model.predict({"input_ids": token_ids, "attention_mask": mask})
    return predictions['embeddings'][0]

def build_section_header(section_name, article_title):
    return f"Page: {article_title}\nSection: {section_name}\n\n"

def lazypack_chunk4(text, seps=["\n\n", ".\n", "?\n", "\n", ". ", "? ", " "], max_length=512, overlap=64, tokenize_fn=None):
    def count_units(chunk):
        return len(tokenize_fn(chunk)) if tokenize_fn else len(chunk)

    chunks = [text]
    if count_units(text) <= max_length:
        return chunks

    for sep in seps:
        new_chunks = []
        for chunk in chunks:
            split_chunks = [sub_chunk.strip() for sub_chunk in chunk.split(sep)]
            packed_chunk = ""
            for split_chunk in split_chunks:
                if packed_chunk and count_units(packed_chunk + sep + split_chunk) <= (max_length - overlap):
                    packed_chunk += sep + split_chunk
                else:
                    if packed_chunk:
                        new_chunks.append(packed_chunk)
                    packed_chunk = split_chunk
            if packed_chunk:
                new_chunks.append(packed_chunk)
        # Retain the separator at the end of each sub-chunk for context
        chunks = [x + sep.strip() if sep.strip() and x != new_chunks[-1] else x for x in new_chunks]
        chunks = new_chunks
        if all(count_units(chunk) <= (max_length - overlap) for chunk in chunks):
            break

    final_chunks = []
    previous_end = ""
    for i in range(len(chunks)):
        current_chunk = chunks[i]
        if previous_end:
            # Calculate the overlap based on tokens or characters
            half_overlap_length = min(overlap, int(count_units(previous_end) / 2), int(count_units(current_chunk) / 2))
            # Ensure we do not split words in the overlap
            words = previous_end.split()
            overlap_text = ""
            overlap_count = 0
            for word in reversed(words):
                overlap_count += count_units(word + " ")
                if overlap_count > half_overlap_length:
                    break
                overlap_text = word + " " + overlap_text

            # Add the overlap text to the current chunk
            combined_chunk = (overlap_text + current_chunk).strip()
            if count_units(combined_chunk) > max_length:
                final_chunks.append(current_chunk)
            else:
                final_chunks.append(combined_chunk)
        else:
            final_chunks.append(current_chunk) # First chunk, no overlap
        previous_end = current_chunk

    return final_chunks

def initialize_worker():
    global model, tokenizer
    model = ct.models.CompiledMLModel('./msmarco_distilbert_base_tas_b_512_single_quantized.mlmodelc')
    tokenizer = AutoTokenizer.from_pretrained("sentence-transformers/msmarco-distilbert-base-tas-b")

def batch_embed(batch, db_name):
    conn = sqlite3.connect(db_name)
    placeholders = ', '.join(['?'] * len(batch))
    rows = conn.execute(f"SELECT * FROM article_sections WHERE id IN ({placeholders})", batch).fetchall()
    conn.close()

    print(f"🚜 starting embedding batch from {rows[0][0]} to {rows[-1][0]}...")

    keys = []
    vectors = []
    for row in rows:
        (section_id, article_id, url, title, seq_id, section, text) = row
        chunk_header = build_section_header(section, title)
        chunk_header_offset = len(tokenize_for_chunking(chunk_header))
        max_chunk_length = 512 - chunk_header_offset
        chunks = lazypack_chunk4(text, max_length=max_chunk_length, tokenize_fn=tokenize_for_chunking)
        for chunk in chunks:
            chunk_embedding = embed(chunk_header + chunk)
            keys.append(section_id)
            vectors.append(chunk_embedding)
    print(f"🚜 finished embedding batch from {rows[0][0]} to {rows[-1][0]}.")
    return keys, np.array(vectors)

def index_batches_parallel(db_name, index_name, precision, batch_size=2000, num_workers=3):
    conn = sqlite3.connect(db_name)
    total_sections = conn.execute("SELECT COUNT(1) FROM article_sections").fetchone()[0]
    conn.close()

    print(f"🔮 Beginning embedding {total_sections} sections...")

    batches = [(list(range(i, min(i + batch_size, total_sections))), db_name) for i in range(0, total_sections, batch_size)]

    with Pool(num_workers, initializer=initialize_worker) as pool:
        results = pool.starmap(batch_embed, batches)

    batch_count = len(results)
    print(f"🔮 Beginning indexing {total_sections} in {batch_count} batches...")
    allowed_precision = ['f16', 'f32'] # NOTE: do not allow i8; known issues (accuracy, performance), esp w inner product metric
    index_precision = precision if precision in allowed_precision else 'f32'
    index = Index(ndim=768, connectivity=16, metric="ip", multi=True, dtype=index_precision)
    for keys, vectors in results:
        index.add(keys=keys, vectors=vectors)

    index.save(index_name)
    index_size = len(index)
    print(f"🔮 Indexing complete with {index_size} records; saved to disk at `{index_name}`.")


def save_wiki_dump(lang, date):
    db_name = f"wikipedia_{lang}_{date}.db"
    db_exists = os.path.exists(db_name)
    if db_exists:
        print(f"📦 Found existing database `{db_name}`. Skipping download & clean.")
    else:
        print(f"☁️ Loading Wikipedia articles in {lang} from {date}...")
        dump_filepath = download_wiki_dump(lang, date)
        print(f"🔍 Downloaded dump file. Cleaning... {dump_filepath}")
        articles = process_wiki_dump(dump_filepath, lang)

        n_articles = articles.__sizeof__()
        if n_articles == 0:
            print("💩 No articles found!")
            print(f"   - the data> {articles}")
            exit(1)

        print(f"📦 Found {n_articles} articles! Importing into sqlite `{db_name}`, embedding and indexing...")
        conn = sqlite3.connect(db_name)
        cursor = conn.cursor()

        cursor.execute("""
            CREATE TABLE IF NOT EXISTS article_sections (
                id INTEGER PRIMARY KEY AUTOINCREMENT,
                article_id INTEGER,
                url TEXT,
                title TEXT,
                sequence_id INTEGER,
                section_name TEXT,
                text TEXT
            );
        """)

        batch_size = 1000
        batch = []
        total_articles = 0
        total_sections = 0

        for article in articles:
            article_id = article["article_id"]
            url = article["url"]
            title = article["title"]
            for seq_id, (section, text) in enumerate(article["sections"]):
                batch.append((article_id, url, title, seq_id, section, text))
                total_sections += 1
            total_articles += 1

            if len(batch) >= batch_size:
                print(f"⚙️ Processing articles up to {total_articles}...")
                cursor.executemany("""
                    INSERT OR REPLACE INTO article_sections (article_id, url, title, sequence_id, section_name, text)
                    VALUES (?, ?, ?, ?, ?, ?);
                """, batch)
                conn.commit()

                print(f"⚙️ Processed {total_articles} articles with {total_sections} sections...")
                batch = []

        if batch:
            print(f"⚙️ Processing articles up to {total_articles}...")
            cursor.executemany("""
                INSERT OR REPLACE INTO article_sections (article_id, url, title, sequence_id, section_name, text)
                VALUES (?, ?, ?, ?, ?, ?);
            """, batch)
            conn.commit()

        conn.execute("VACUUM;")
        count = conn.execute("SELECT COUNT(1) FROM article_sections;").fetchone()[0]
        conn.close()

        print(f"🌎 Initial processing complete! Total sections in database: {count}")#, Index Size: {index_size}")
    # end clean & save wikidump

    index_precision = "f16"
    index_path = f"./{lang}wiki-{date}.{index_precision}.index"
    index_batches_parallel(db_name, index_path, index_precision)

    index = Index.restore(index_path, view=True)
    conn = sqlite3.connect(db_name)
    cursor = conn.cursor()

    query = "What is the capital of AUS?"
    print(f"🔎 testing search... '{query}'")
    qembed = embed(query)
    res: Matches = index.search(qembed, 5)
    print(f"  - Results:")
    for result in res:
        (title, section, text) = cursor.execute("SELECT title, section_name, text FROM article_sections WHERE id = ?;", (f"{result.key}",)).fetchone()
        snippet = text[:280].replace("\n", " ")
        print(f"    - Key: {result.key} | Distance: {result.distance} | Excerpt from '{title}', '{section}': {snippet}")

    conn.close()

if __name__ == "__main__":
    model = ct.models.CompiledMLModel('./msmarco_distilbert_base_tas_b_512_single_quantized.mlmodelc')
    tokenizer = AutoTokenizer.from_pretrained("sentence-transformers/msmarco-distilbert-base-tas-b")
    lang = "simple"
    date = "20240720"
    save_wiki_dump(lang, date)