wp2txt 2.1.2 → 2.3.0

data/lib/wp2txt/metadata_index.rb

@@ -0,0 +1,738 @@
+# frozen_string_literal: true
+
+require "sqlite3"
+require "fileutils"
+require "digest"
+require "open3"
+require "parallel"
+require "time"
+require "json"
+require_relative "regex"
+require_relative "section_extractor"
+require_relative "version"
+
+module Wp2txt
+  # Local metadata index (Tier 1) built from a multistream dump.
+  # Stores per-page categories, section headings, redirects, and the category
+  # hierarchy in SQLite, enabling offline exhaustive queries such as
+  # "all articles in category X that have a Plot section" without any API access.
+  class MetadataIndex
+    SCHEMA_VERSION = 2
+    CACHE_SUFFIX = "_meta.sqlite3"
+    NS_ARTICLE = 0
+    NS_CATEGORY = 14
+
+    attr_reader :db_path
+
+    def initialize(db_path)
+      @db_path = db_path
+      @db = nil
+    end
+
+    # Default index location for a given multistream file (mirrors IndexCache naming)
+    def self.path_for(multistream_path, cache_dir: nil)
+      dir = cache_dir || File.expand_path("~/.wp2txt/cache")
+      basename = File.basename(multistream_path, ".*").sub(/\.xml\z/, "")
+      path_hash = Digest::MD5.hexdigest(multistream_path)[0, 8]
+      File.join(dir, "#{basename}_#{path_hash}#{CACHE_SUFFIX}")
+    end
+
+    # Normalize a page title the way MediaWiki treats titles:
+    # underscores to spaces, trimmed, first letter capitalized
+    def self.normalize_title(name)
+      n = name.to_s.tr("_", " ").strip.squeeze(" ")
+      return n if n.empty?
+
+      n[0].upcase + n[1..].to_s
+    end
+
+    # Normalize a category name (same MediaWiki title rules as normalize_title)
+    def self.normalize_category(name)
+      normalize_title(name)
+    end
+
+    # Built metadata DBs for one language found in a cache directory, most
+    # recently built first. Used for cross-dump ATTACH resolution (Corpus)
+    # and langlinks sanity checks (LanglinksImporter).
+    # @return [Array<Hash>] [{db_path:, dump_name:, built_at:, built_with:}]
+    def self.cached_candidates(lang, cache_dir: nil)
+      dir = cache_dir || File.expand_path("~/.wp2txt/cache")
+      Dir.glob(File.join(dir, "#{lang}wiki-*#{CACHE_SUFFIX}")).filter_map do |path|
+        meta = read_metadata_file(path)
+        next unless meta && meta[:schema_version].to_i == SCHEMA_VERSION && meta[:built_at]
+
+        { db_path: path, dump_name: meta[:dump_name], built_at: meta[:built_at],
+          built_with: meta[:wp2txt_version] }
+      end.sort_by { |c| c[:built_at].to_s }.reverse
+    end
+
+    # Light, read-only metadata table read for a DB file we do not manage
+    def self.read_metadata_file(path)
+      db = SQLite3::Database.new(path, readonly: true)
+      result = {}
+      db.execute("SELECT key, value FROM metadata") { |key, value| result[key.to_sym] = value }
+      result
+    rescue SQLite3::Exception
+      nil
+    ensure
+      db&.close
+    end
+
+    # Remove wiki markup from a heading ('''bold''', [[link|label]], HTML tags)
+    def self.clean_heading(text)
+      t = text.gsub(/'{2,}/, "")
+      t = t.gsub(/\[\[(?:[^\]|]*\|)?([^\]]*)\]\]/) { ::Regexp.last_match(1) }
+      t.gsub(/<[^>]+>/, "").strip
+    end
+
+    # Expand a section name to its full alias group (bidirectional):
+    # "Plot" => ["Plot", "Synopsis", ...]; "Synopsis" => same group
+    def self.expand_section_names(name, alias_file: nil)
+      aliases = SectionExtractor::DEFAULT_ALIASES
+      if alias_file
+        custom = SectionExtractor.load_aliases_from_file(alias_file)
+        aliases = aliases.merge(custom) unless custom.empty?
+      end
+
+      down = name.downcase
+      aliases.each do |canonical, list|
+        group = [canonical, *list]
+        return group if group.any? { |g| g.downcase == down }
+      end
+      [name]
+    end
+
+    # ------------------------------------------------------------------
+    # Status
+    # ------------------------------------------------------------------
+
+    # True if the index file exists and has a compatible schema
+    def built?
+      return false unless File.exist?(@db_path)
+
+      meta = read_metadata
+      !meta.nil? && meta[:schema_version].to_i == SCHEMA_VERSION && !meta[:built_at].nil?
+    rescue SQLite3::Exception
+      false
+    end
+
+    # True if the index was built from the given (unchanged) multistream file
+    def valid_for?(multistream_path)
+      return false unless built?
+      return false unless File.exist?(multistream_path)
+
+      meta = read_metadata
+      stat = File.stat(multistream_path)
+      meta[:source_size].to_i == stat.size && meta[:source_mtime].to_i == stat.mtime.to_i
+    end
+
+    def stats
+      return nil unless File.exist?(@db_path)
+
+      meta = read_metadata || {}
+      {
+        db_path: @db_path,
+        db_size: File.size(@db_path),
+        dump_name: meta[:dump_name],
+        built_at: meta[:built_at],
+        built_with: meta[:wp2txt_version],
+        page_count: count_scalar("SELECT COUNT(*) FROM pages"),
+        article_count: count_scalar("SELECT COUNT(*) FROM pages WHERE namespace = #{NS_ARTICLE} AND redirect_to IS NULL"),
+        category_count: count_scalar("SELECT COUNT(DISTINCT category) FROM page_categories"),
+        section_count: count_scalar("SELECT COUNT(*) FROM page_sections")
+      }
+    end
+
+    # Provenance of an imported langlinks table (nil when not imported).
+    # The langlinks table is an optional post-build addition (LanglinksImporter),
+    # so its absence does not affect built? or schema_version.
+    def langlinks_provenance
+      return nil unless File.exist?(@db_path)
+
+      meta = read_metadata
+      return nil unless meta && meta[:langlinks_imported_at]
+
+      { source: meta[:langlinks_source],
+        source_size: meta[:langlinks_source_size].to_i,
+        imported_at: meta[:langlinks_imported_at],
+        imported_with: meta[:langlinks_wp2txt_version],
+        lang_filter: meta[:langlinks_lang_filter],
+        row_count: meta[:langlinks_row_count].to_i,
+        skipped_invalid: meta[:langlinks_skipped_invalid].to_i }
+    end
+
+    def close
+      @db&.close
+      @db = nil
+    end
+
+    # ------------------------------------------------------------------
+    # Build API (used by MetadataIndexBuilder)
+    # ------------------------------------------------------------------
+
+    # Build into a sidecar file and atomically rename in finalize_build!, so a
+    # failed multi-hour rebuild never destroys a working index
+    def prepare_build!
+      FileUtils.mkdir_p(File.dirname(@db_path))
+      close
+      @build_path = "#{@db_path}.building"
+      FileUtils.rm_f([@build_path, "#{@build_path}-wal", "#{@build_path}-shm"])
+      db = open_db
+      db.execute(<<~SQL)
+        CREATE TABLE metadata (key TEXT PRIMARY KEY, value TEXT)
+      SQL
+      db.execute(<<~SQL)
+        CREATE TABLE pages (
+          page_id INTEGER PRIMARY KEY,
+          title TEXT,
+          namespace INTEGER,
+          redirect_to TEXT,
+          text_length INTEGER
+        )
+      SQL
+      db.execute("CREATE TABLE page_categories (page_id INTEGER, category TEXT)")
+      # ord: position of the section within the article, where 0 is the lead
+      # text before the first heading. Lead rows are NOT stored here (they have
+      # no heading), so ord starts at 1 — consistent with fts_map.ord in the
+      # full-text DB, where the lead IS stored as ord 0.
+      db.execute(<<~SQL)
+        CREATE TABLE page_sections (
+          page_id INTEGER,
+          heading TEXT,
+          level INTEGER,
+          -- ord: section position in the article; 0 = lead text (not stored in
+          -- this table), so headings start at 1. Same semantics as fts_map.ord.
+          ord INTEGER
+        )
+      SQL
+      db.execute("CREATE TABLE category_hierarchy (child TEXT, parent TEXT)")
+    end
+
+    # Insert one scanned batch: {pages:, categories:, sections:, hierarchy:}
+    def insert_batch(rows)
+      db = open_db
+      db.transaction do
+        stmt = db.prepare("INSERT OR IGNORE INTO pages (page_id, title, namespace, redirect_to, text_length) VALUES (?, ?, ?, ?, ?)")
+        rows[:pages].each { |r| stmt.execute(r) }
+        stmt.close
+
+        stmt = db.prepare("INSERT INTO page_categories (page_id, category) VALUES (?, ?)")
+        rows[:categories].each { |r| stmt.execute(r) }
+        stmt.close
+
+        stmt = db.prepare("INSERT INTO page_sections (page_id, heading, level, ord) VALUES (?, ?, ?, ?)")
+        rows[:sections].each { |r| stmt.execute(r) }
+        stmt.close
+
+        stmt = db.prepare("INSERT INTO category_hierarchy (child, parent) VALUES (?, ?)")
+        rows[:hierarchy].each { |r| stmt.execute(r) }
+        stmt.close
+      end
+    end
+
+    def finalize_build!(source_path)
+      db = open_db
+      db.execute("CREATE INDEX IF NOT EXISTS idx_pages_title ON pages(title)")
+      db.execute("CREATE INDEX IF NOT EXISTS idx_pc_category ON page_categories(category)")
+      db.execute("CREATE INDEX IF NOT EXISTS idx_pc_page ON page_categories(page_id)")
+      db.execute("CREATE INDEX IF NOT EXISTS idx_ps_heading ON page_sections(heading COLLATE NOCASE)")
+      db.execute("CREATE INDEX IF NOT EXISTS idx_ps_page ON page_sections(page_id)")
+      db.execute("CREATE INDEX IF NOT EXISTS idx_ch_parent ON category_hierarchy(parent)")
+
+      stat = File.stat(source_path)
+      dump_name = File.basename(source_path)[/\A[a-z0-9_\-]+?-\d{8}/] || File.basename(source_path)
+      save_metadata(
+        schema_version: SCHEMA_VERSION,
+        wp2txt_version: Wp2txt::VERSION,
+        source_path: source_path,
+        source_size: stat.size,
+        source_mtime: stat.mtime.to_i,
+        dump_name: dump_name,
+        built_at: Time.now.utc.iso8601
+      )
+      db.execute("ANALYZE")
+      close
+      if @build_path
+        File.rename(@build_path, @db_path)
+        FileUtils.rm_f(["#{@db_path}-wal", "#{@db_path}-shm"])
+        @build_path = nil
+      end
+    end
+
+    # ------------------------------------------------------------------
+    # Queries
+    # ------------------------------------------------------------------
+
+    # Find article titles matching the given filters.
+    # @param category [String, nil] category name (without namespace prefix)
+    # @param depth [Integer] subcategory recursion depth (0 = exact category only)
+    # @param has_section [String, nil] single section heading (alias-aware by default)
+    # @param sections [Array<String>, nil] multiple headings (OR match, used as-is)
+    # @param alias_set [String, nil] saved alias set name used to expand headings
+    # @param use_aliases [Boolean] expand has_section via built-in alias groups
+    # @param alias_file [String, nil] custom alias YAML (merged with defaults)
+    # @param title_match [String, nil] substring match on title
+    # @param limit [Integer] max titles to return (0 = no limit)
+    # @param offset [Integer] result offset
+    # @return [Array<String>] matching titles ordered by page_id
+    def find_articles(category: nil, depth: 0, categories: nil, category_match: nil,
+                      has_section: nil, sections: nil, alias_set: nil,
+                      use_aliases: true, alias_file: nil, title_match: nil, limit: 0, offset: 0)
+      cte, where, params = build_article_query(
+        category: category, depth: depth, categories: categories, category_match: category_match,
+        has_section: has_section, sections: sections,
+        alias_set: alias_set, use_aliases: use_aliases, alias_file: alias_file, title_match: title_match
+      )
+      sql = +""
+      sql << "WITH RECURSIVE #{cte} " if cte
+      sql << "SELECT p.title FROM pages p WHERE #{where} ORDER BY p.page_id"
+      sql << " LIMIT #{limit.to_i}" if limit.to_i.positive?
+      sql << " OFFSET #{offset.to_i}" if offset.to_i.positive?
+
+      open_db.execute(sql, params).map { |row| row[0] }
+    end
+
+    # Count articles matching the same filters as find_articles
+    def count_articles(category: nil, depth: 0, categories: nil, category_match: nil,
+                       has_section: nil, sections: nil, alias_set: nil,
+                       use_aliases: true, alias_file: nil, title_match: nil)
+      cte, where, params = build_article_query(
+        category: category, depth: depth, categories: categories, category_match: category_match,
+        has_section: has_section, sections: sections,
+        alias_set: alias_set, use_aliases: use_aliases, alias_file: alias_file, title_match: title_match
+      )
+      sql = +""
+      sql << "WITH RECURSIVE #{cte} " if cte
+      sql << "SELECT COUNT(*) FROM pages p WHERE #{where}"
+
+      open_db.get_first_value(sql, params).to_i
+    end
+
+    # Categories of one article (by exact title)
+    # @return [Array<String>, nil] category names, nil if the title is unknown
+    def categories_of(title)
+      row = open_db.execute("SELECT page_id FROM pages WHERE title = ?", [title]).first
+      return nil unless row
+
+      open_db.execute(
+        "SELECT category FROM page_categories WHERE page_id = ? ORDER BY category", [row[0]]
+      ).map(&:first)
+    end
+
+    # Look up titles in pages (existence + redirect target), batched to keep
+    # the IN clause small. Used by Corpus#extract_corpus titles: resolution.
+    # @return [Hash] { title => redirect_to_or_nil } for the titles that exist
+    def redirect_map(titles)
+      result = {}
+      titles.each_slice(500) do |slice|
+        placeholders = slice.map { "?" }.join(",")
+        open_db.execute(
+          "SELECT title, redirect_to FROM pages WHERE title IN (#{placeholders})", slice
+        ).each do |title, redirect_to|
+          result[title] = redirect_to
+        end
+      end
+      result
+    end
+
+    # Subcategory tree starting at category, as [{name:, depth:}, ...] (BFS order)
+    def category_tree(category, depth: 2)
+      cat = self.class.normalize_category(category)
+      sql = <<~SQL
+        WITH RECURSIVE cat_tree(name, d) AS (
+          SELECT ?, 0
+          UNION
+          SELECT ch.child, ct.d + 1
+          FROM category_hierarchy ch JOIN cat_tree ct ON ch.parent = ct.name
+          WHERE ct.d < #{depth.to_i}
+        )
+        SELECT name, MIN(d) FROM cat_tree GROUP BY name ORDER BY MIN(d), name
+      SQL
+      open_db.execute(sql, [cat]).map { |name, d| { name: name, depth: d } }
+    end
+
+    # Section heading frequencies across articles, optionally scoped to a category
+    def section_stats(category: nil, depth: 0, top_n: 50)
+      conds = ["p.namespace = #{NS_ARTICLE}", "p.redirect_to IS NULL"]
+      params = []
+      cte = nil
+      if category
+        cte, cond, cat_params = category_condition(category, depth)
+        conds << cond
+        params.concat(cat_params)
+      end
+      sql = +""
+      sql << "WITH RECURSIVE #{cte} " if cte
+      sql << <<~SQL
+        SELECT ps.heading, COUNT(*) AS cnt
+        FROM page_sections ps JOIN pages p ON p.page_id = ps.page_id
+        WHERE #{conds.join(' AND ')}
+        GROUP BY ps.heading ORDER BY cnt DESC, ps.heading LIMIT #{top_n.to_i}
+      SQL
+      open_db.execute(sql, params)
+    end
+
+    # Article counts, average positions, and pairwise co-occurrence for a set of
+    # headings. Synonymous headings almost never co-occur in the same article,
+    # so a high co-occurrence ratio is evidence AGAINST treating them as aliases.
+    # @param headings [Array<String>] headings to compare
+    # @param category [String, nil] optional category scope
+    # @param depth [Integer] category recursion depth
+    # @return [Hash] { headings: [{heading:, articles:, avg_position:}],
+    #                  pairs: [{a:, b:, both:, cooccurrence_ratio:}] }
+    def section_cooccurrence(headings, category: nil, depth: 0)
+      scope_cte = nil
+      scope_cond = "p.namespace = #{NS_ARTICLE} AND p.redirect_to IS NULL"
+      scope_params = []
+      if category
+        scope_cte, cond, scope_params = category_condition(category, depth)
+        scope_cond += " AND #{cond}"
+      end
+
+      db = open_db
+      with = scope_cte ? "WITH RECURSIVE #{scope_cte} " : ""
+      # Placeholders bind positionally: with a recursive CTE the category `?`
+      # sits inside the CTE (before any heading `?`); without one it sits
+      # inside scope_cond (after the heading `?`)
+      cte_params = scope_cte ? scope_params : []
+      cond_params = scope_cte ? [] : scope_params
+
+      heading_stats = headings.map do |h|
+        row = db.execute(
+          "#{with}SELECT COUNT(DISTINCT ps.page_id), AVG(ps.ord) FROM page_sections ps " \
+          "JOIN pages p ON p.page_id = ps.page_id " \
+          "WHERE ps.heading COLLATE NOCASE = ? AND #{scope_cond}",
+          cte_params + [h] + cond_params
+        ).first
+        { heading: h, articles: row[0].to_i, avg_position: row[1]&.round(2) }
+      end
+
+      counts = heading_stats.to_h { |s| [s[:heading], s[:articles]] }
+      pairs = headings.combination(2).map do |a, b|
+        both = db.get_first_value(
+          "#{with}SELECT COUNT(*) FROM (" \
+          "SELECT ps.page_id FROM page_sections ps JOIN pages p ON p.page_id = ps.page_id " \
+          "WHERE ps.heading COLLATE NOCASE = ? AND #{scope_cond} " \
+          "INTERSECT " \
+          "SELECT ps.page_id FROM page_sections ps JOIN pages p ON p.page_id = ps.page_id " \
+          "WHERE ps.heading COLLATE NOCASE = ? AND #{scope_cond})",
+          cte_params + [a] + cond_params + [b] + cond_params
+        ).to_i
+        min = [counts[a], counts[b]].min
+        { a: a, b: b, both: both, cooccurrence_ratio: min.positive? ? (both.to_f / min).round(3) : 0.0 }
+      end
+
+      { headings: heading_stats, pairs: pairs }
+    end
+
+    # ------------------------------------------------------------------
+    # Alias sets (LLM-generated, persisted per dump for reproducibility)
+    # ------------------------------------------------------------------
+
+    # Save a named alias set. groups is an array of heading groups, e.g.
+    # [["あらすじ", "ストーリー", "物語"], ["脚注", "出典"]]
+    def save_alias_set(name, groups)
+      raise ArgumentError, "groups must be a non-empty array of arrays" unless groups.is_a?(Array) && !groups.empty? && groups.all? { |g| g.is_a?(Array) && !g.empty? }
+
+      ensure_alias_table
+      open_db.execute(
+        "INSERT OR REPLACE INTO alias_sets (name, groups, created_at) VALUES (?, ?, ?)",
+        [name, JSON.generate(groups), Time.now.utc.iso8601]
+      )
+      { name: name, groups: groups }
+    end
+
+    # @return [Hash, nil] { name:, groups:, created_at: } or nil if not found
+    def get_alias_set(name)
+      ensure_alias_table
+      row = open_db.execute("SELECT name, groups, created_at FROM alias_sets WHERE name = ?", [name]).first
+      return nil unless row
+
+      { name: row[0], groups: JSON.parse(row[1]), created_at: row[2] }
+    end
+
+    def list_alias_sets
+      ensure_alias_table
+      open_db.execute("SELECT name, groups, created_at FROM alias_sets ORDER BY name").map do |name, groups, created_at|
+        { name: name, group_count: JSON.parse(groups).size, created_at: created_at }
+      end
+    end
+
+    def delete_alias_set(name)
+      ensure_alias_table
+      open_db.execute("DELETE FROM alias_sets WHERE name = ?", [name])
+      nil
+    end
+
+    private
+
+    # Returns [cte_sql_or_nil, where_sql, params]
+    def build_article_query(category:, depth:, has_section:, categories: nil, category_match: nil,
+                            sections: nil, alias_set: nil,
+                            use_aliases: true, alias_file: nil, title_match: nil)
+      conds = ["p.namespace = #{NS_ARTICLE}", "p.redirect_to IS NULL"]
+      params = []
+      cte = nil
+
+      if category
+        cte, cond, cat_params = category_condition(category, depth)
+        conds << cond
+        params.concat(cat_params)
+      end
+
+      # AND-intersection of exact category memberships (no recursion)
+      Array(categories).each do |cat|
+        conds << "p.page_id IN (SELECT pc.page_id FROM page_categories pc WHERE pc.category = ?)"
+        params << self.class.normalize_category(cat)
+      end
+
+      # Substring match on category names, e.g. "アメリカ合衆国の%映画" style scoping
+      if category_match
+        conds << "p.page_id IN (SELECT pc.page_id FROM page_categories pc WHERE pc.category LIKE ? ESCAPE '\\')"
+        params << "%#{escape_like(category_match)}%"
+      end
+
+      names = resolve_section_names(has_section: has_section, sections: sections,
+                                    alias_set: alias_set, use_aliases: use_aliases, alias_file: alias_file)
+      if names
+        placeholders = names.map { "?" }.join(",")
+        conds << "p.page_id IN (SELECT ps.page_id FROM page_sections ps WHERE ps.heading COLLATE NOCASE IN (#{placeholders}))"
+        params.concat(names)
+      end
+
+      if title_match
+        conds << "p.title LIKE ? ESCAPE '\\'"
+        params << "%#{escape_like(title_match)}%"
+      end
+
+      [cte, conds.join(" AND "), params]
+    end
+
+    # Merge has_section / sections into one heading list, expanding through a
+    # saved alias set (if given) or the built-in alias groups (single name only).
+    def resolve_section_names(has_section:, sections:, alias_set:, use_aliases:, alias_file:)
+      names = Array(sections).compact
+      names += [has_section] if has_section
+      return nil if names.empty?
+
+      if alias_set
+        set = get_alias_set(alias_set)
+        raise ArgumentError, "alias set not found: #{alias_set}" unless set
+
+        names = names.flat_map { |n| expand_via_groups(n, set[:groups]) }
+      elsif use_aliases && names.size == 1 && sections.nil?
+        names = self.class.expand_section_names(names.first, alias_file: alias_file)
+      end
+      names.uniq
+    end
+
+    # Bidirectional expansion through an array of heading groups
+    def expand_via_groups(name, groups)
+      down = name.downcase
+      groups.each do |group|
+        return group if group.any? { |g| g.downcase == down }
+      end
+      [name]
+    end
+
+    def ensure_alias_table
+      open_db.execute(<<~SQL)
+        CREATE TABLE IF NOT EXISTS alias_sets (
+          name TEXT PRIMARY KEY,
+          groups TEXT,
+          created_at TEXT
+        )
+      SQL
+    end
+
+    # Returns [cte_sql_or_nil, condition_sql, params] for a category filter
+    def category_condition(category, depth)
+      cat = self.class.normalize_category(category)
+      if depth.to_i.positive?
+        cte = <<~SQL.strip
+          cat_tree(name, d) AS (
+            SELECT ?, 0
+            UNION
+            SELECT ch.child, ct.d + 1
+            FROM category_hierarchy ch JOIN cat_tree ct ON ch.parent = ct.name
+            WHERE ct.d < #{depth.to_i}
+          )
+        SQL
+        cond = "p.page_id IN (SELECT pc.page_id FROM page_categories pc WHERE pc.category IN (SELECT name FROM cat_tree))"
+        [cte, cond, [cat]]
+      else
+        [nil, "p.page_id IN (SELECT pc.page_id FROM page_categories pc WHERE pc.category = ?)", [cat]]
+      end
+    end
+
+    def escape_like(str)
+      str.gsub(/[\\%_]/) { |c| "\\#{c}" }
+    end
+
+    def open_db
+      return @db if @db
+
+      @db = SQLite3::Database.new(@build_path || @db_path)
+      @db.busy_timeout = 5000
+      @db.execute("PRAGMA journal_mode = WAL")
+      @db.execute("PRAGMA synchronous = NORMAL")
+      @db.execute("PRAGMA cache_size = -64000")
+      @db
+    end
+
+    def count_scalar(sql)
+      open_db.get_first_value(sql).to_i
+    rescue SQLite3::Exception
+      0
+    end
+
+    def read_metadata
+      db = open_db
+      result = {}
+      db.execute("SELECT key, value FROM metadata") do |key, value|
+        result[key.to_sym] = value
+      end
+      result
+    rescue SQLite3::Exception
+      nil
+    end
+
+    def save_metadata(hash)
+      db = open_db
+      stmt = db.prepare("INSERT OR REPLACE INTO metadata (key, value) VALUES (?, ?)")
+      hash.each { |k, v| stmt.execute([k.to_s, v.to_s]) }
+      stmt.close
+    end
+  end
+
+  # Builds a MetadataIndex by scanning all streams of a multistream dump in parallel.
+  # Workers decompress and regex-scan streams; the parent process owns the sole
+  # SQLite connection and inserts each batch from the Parallel `finish` hook.
+  class MetadataIndexBuilder
+    STREAMS_PER_BATCH = 50
+
+    PAGE_BLOCK_REGEX = %r{<page>(.*?)</page>}m
+    TITLE_REGEX = %r{<title>([^<]*)</title>}
+    NS_REGEX = %r{<ns>(-?\d+)</ns>}
+    ID_REGEX = %r{<id>(\d+)</id>}
+    TEXT_REGEX = %r{<text[^>]*>(.*)</text>}m
+    HEADING_REGEX = /\A(={2,6})[ \t]*(.+?)[ \t]*={2,6}[ \t]*\z/
+    HTML_COMMENT_REGEX = /<!--.*?-->/m
+
+    def initialize(multistream_path, stream_offsets, db_path:, num_processes: 4)
+      @multistream_path = multistream_path
+      @stream_offsets = stream_offsets
+      @db_path = db_path
+      @num_processes = num_processes
+    end
+
+    # Build the index. Yields (batches_done, batches_total) after each batch.
+    # @return [MetadataIndex] the built index
+    def build(&progress)
+      index = MetadataIndex.new(@db_path)
+      index.prepare_build!
+      # Close before Parallel forks workers so children do not inherit a
+      # writable SQLite connection; the finish hook reopens it lazily in
+      # the parent, which is the only process that ever writes.
+      index.close
+
+      pairs = @stream_offsets.zip(@stream_offsets[1..].to_a + [nil])
+      batches = pairs.each_slice(STREAMS_PER_BATCH).to_a
+      done = 0
+
+      Parallel.map(
+        batches,
+        in_processes: @num_processes,
+        finish: lambda { |_item, _idx, result|
+          index.insert_batch(result)
+          done += 1
+          progress&.call(done, batches.size)
+        }
+      ) do |batch|
+        self.class.scan_batch(@multistream_path, batch)
+      end
+
+      index.finalize_build!(@multistream_path)
+      index
+    end
+
+    # Scan a batch of [offset, next_offset] stream pairs.
+    # Runs inside worker processes: must not touch SQLite.
+    def self.scan_batch(multistream_path, offset_pairs)
+      out = { pages: [], categories: [], sections: [], hierarchy: [] }
+      File.open(multistream_path, "rb") do |f|
+        offset_pairs.each do |offset, next_offset|
+          f.seek(offset)
+          data = next_offset ? f.read(next_offset - offset) : f.read
+          xml = decompress_bz2(data)
+          xml.scan(PAGE_BLOCK_REGEX) { scan_page(::Regexp.last_match(1), out) }
+        end
+      end
+      out
+    end
+
+    def self.decompress_bz2(data)
+      stdout, status = Open3.capture2("bzcat", stdin_data: data)
+      raise "bzcat failed (exit #{status.exitstatus})" unless status.success?
+
+      stdout.force_encoding(Encoding::UTF_8)
+    end
+
+    def self.scan_page(block, out)
+      title = block[TITLE_REGEX, 1]
+      return unless title && !title.empty?
+
+      page_id = block[ID_REGEX, 1]&.to_i
+      return unless page_id
+
+      title = unescape_xml(title)
+      ns = (block[NS_REGEX, 1] || "0").to_i
+      text = block[TEXT_REGEX, 1] || ""
+      text = unescape_xml(text)
+      # Strip HTML comments before scanning, matching what the Article parser
+      # does for the FTS index: a trailing comment after a heading's closing
+      # `==` must not hide the heading, and commented-out [[Category:]] links
+      # must not be indexed as real categories
+      text = text.gsub(HTML_COMMENT_REGEX, "")
+
+      redirect_to = nil
+      if (m = REDIRECT_REGEX.match(text))
+        redirect_to = m[1].split(/[#|]/).first.to_s.strip
+        redirect_to = nil if redirect_to.empty?
+      end
+
+      out[:pages] << [page_id, title, ns, redirect_to, text.length]
+
+      categories = text.scan(CATEGORY_REGEX)
+                       .map { |c| MetadataIndex.normalize_category(c[0]) }
+                       .reject(&:empty?).uniq
+      if ns == MetadataIndex::NS_CATEGORY
+        child = MetadataIndex.normalize_category(title.sub(/\A[^:]+:/, ""))
+        categories.each { |c| out[:hierarchy] << [child, c] } unless child.empty?
+      else
+        categories.each { |c| out[:categories] << [page_id, c] }
+      end
+
+      # ord 0 is the lead text (not stored here — it has no heading); headings
+      # start at 1 so ord aligns with fts_map.ord in the full-text DB
+      ord = 0
+      text.each_line do |line|
+        l = line.chomp
+        next unless l.start_with?("==")
+
+        hm = HEADING_REGEX.match(l)
+        next unless hm
+
+        ord += 1
+        heading = MetadataIndex.clean_heading(hm[2])
+        next if heading.empty?
+
+        out[:sections] << [page_id, heading, hm[1].length, ord]
+      end
+    end
+
+    def self.unescape_xml(str)
+      str.gsub("&lt;", "<").gsub("&gt;", ">").gsub("&quot;", '"').gsub("&amp;", "&")
+    end
+  end
+end