Files
zim-utils/src/reader.js
ztimson 7d376c90b4
All checks were successful
Publish Library / Build NPM Project (push) Successful in 17s
Publish Library / Tag Version (push) Successful in 10s
Create title index for zims missing them
2026-08-24 13:08:17 -04:00

475 lines
18 KiB
JavaScript

'use strict';
import fs from 'node:fs';
import {createHash} from 'node:crypto';
import {decompressPool} from './decompress.js';
import {fuzzyMatch, titleFromUrl, kiwixSearch} from './utils.js';
const HEADER_SIZE = 80;
const NS_CONTENT = 'C';
const NS_METADATA = 'M';
const TITLE_SENTINEL = 0xffffffffffffffffn; // Indicator -> ZIM v6+ archives with no title
const DEFAULT_CLUSTER_CACHE_MAX = 32;
const DEFAULT_CLUSTER_TTL = 60_000;
// --- Fallback search index (only used for archives with no native title listing) ---
const INDEX_SUFFIX = '.searchidx.bin';
const INDEX_MAGIC = 'ZXI1';
const INDEX_HEADER_SIZE = 24; // magic(4) + staleness key(16) + recordCount(4)
const INDEX_RECORD_SIZE = 20; // keyOff(4) keyLen(2) urlOff(4) urlLen(2) titleOff(4) titleLen(2) mimetype(2)
/** Native, dependency-light reader for .zim archives. Supports zstd & LZMA cluster compression. */
export class ZimReader {
#fd = null;
#header = null;
#mimeTypes = [];
#hasTitleListing = false;
#clusterCache = new Map(); // clusterNumber -> {data, extended, timer}
#pending = new Map(); // clusterNumber -> Promise, dedupes concurrent misses
#clusterCacheMax;
#clusterTTL;
#indexPath;
#indexReady = null; // Promise, awaited by search() before using the fallback index
#indexFd = null; // open fd for the fallback index, once loaded/built
#indexRecordCount = 0;
#indexTableStart = 0;
get articleCount() { return this.#header?.articleCount ?? 0; }
get mediaCount() { return this.#header?.clusterCount ?? 0; }
/** @param {{clusterCacheMax?: number, clusterTTL?: number}} [opts] clusterTTL in ms; 0/null disables idle eviction. */
constructor(path, {clusterCacheMax = DEFAULT_CLUSTER_CACHE_MAX, clusterTTL = DEFAULT_CLUSTER_TTL} = {}) {
this.path = path;
this.#clusterCacheMax = clusterCacheMax;
this.#clusterTTL = clusterTTL;
this.#indexPath = `${path}${INDEX_SUFFIX}`;
}
/** Binary search the URL pointer list for namespace+url. For exact-key lookups (readPage, metadata, icons). */
async #findByUrl(url, namespace) {
const key = namespace + url;
let lo = 0, hi = this.#header.articleCount - 1;
while (lo <= hi) {
const mid = (lo + hi) >> 1;
const dirent = await this.#readDirent(await this.#ptr64(this.#header.urlPtrPos, mid));
const dirKey = dirent.namespace + dirent.url;
const cmp = key < dirKey ? -1 : key > dirKey ? 1 : 0;
if (cmp === 0) return dirent;
if (cmp < 0) hi = mid - 1; else lo = mid + 1;
}
return null;
}
/** Binary search the ZIM's own title pointer list. O(log n), zero extra storage - the happy path. */
async #findByTitleBuiltin(title) {
let lo = 0, hi = this.#header.articleCount - 1;
while (lo <= hi) {
const mid = (lo + hi) >> 1;
const urlIdx = (await this.#read(this.#header.titlePtrPos + mid * 4, 4)).readUInt32LE(0);
const d = await this.#readDirent(await this.#ptr64(this.#header.urlPtrPos, urlIdx));
if (d.title === title) return d;
if (d.title < title) lo = mid + 1; else hi = mid - 1;
}
return null;
}
/**
* Resolves a kiwix-search result (a title, not necessarily a url) to
* {url, title, mimetype}. kiwix-search's fulltext index returns titles, and
* a title isn't guaranteed to equal its url (unicode normalization,
* disambiguation suffixes, punctuation stripping), so:
* 1. URL binary search - matches when title happens to equal url (free to check)
* 2. Title pointer list - when the archive ships one (Wikipedia etc. do)
* 3. Persisted fallback index / linear scan - only for archives without (2)
*/
async #resolveSearchEntry(name) {
let dirent = await this.#findByUrl(name, NS_CONTENT);
if (dirent) return dirent;
if (this.#hasTitleListing) return this.#findByTitleBuiltin(name);
if (this.#indexReady) await this.#indexReady;
const hit = await this.#lookupFallbackIndex(name);
if (hit) return hit;
// Index unavailable (build failed - unwritable disk, etc.) or genuinely no match.
for (let i = 0; i < this.#header.articleCount; i++) {
const d = await this.#readDirent(await this.#ptr64(this.#header.urlPtrPos, i));
if (d.namespace === NS_CONTENT && (d.title === name || titleFromUrl(d.url) === name)) return d;
}
return null;
}
/**
* A cheap, stable fingerprint for "is the cached index still valid for this file".
* ZIM archives end with a 16-byte MD5 of their own contents (the same trailer
* `zimcheck` validates against), so this is a single 16-byte read regardless of
* archive size - no need to hash a multi-hundred-GB file. It's also
* content-based rather than path/mtime-based, so moving or redownloading an
* identical archive doesn't invalidate the cache. Falls back to a tiny
* size+mtime hash only if the file is too short to have a real trailer.
*/
async #stalenessKey() {
const {size} = await fs.promises.stat(this.path);
if (size >= 16) return this.#read(size - 16, 16);
const stat = await fs.promises.stat(this.path);
return createHash('md5').update(`${stat.size}:${stat.mtimeMs}`).digest();
}
/** Walks every content dirent once, resolving redirects, keyed by both its title and its url-derived title. */
async #collectFallbackEntries() {
const map = new Map(); // key -> {url, title, mimetype}
for (let i = 0; i < this.#header.articleCount; i++) {
let dirent = await this.#readDirent(await this.#ptr64(this.#header.urlPtrPos, i));
if (dirent.namespace !== NS_CONTENT) continue;
dirent = await this.#resolveRedirect(dirent);
if (dirent.namespace !== NS_CONTENT) continue; // redirected outside content namespace
const entry = {url: dirent.url, title: dirent.title, mimetype: dirent.mimetype};
if (!map.has(dirent.title)) map.set(dirent.title, entry);
const urlTitle = titleFromUrl(dirent.url);
if (urlTitle !== dirent.title && !map.has(urlTitle)) map.set(urlTitle, entry);
}
return map;
}
/** Serializes the fallback index to disk: header, sorted fixed-width records, then a string table. */
async #buildFallbackIndex(stalenessKey) {
const map = await this.#collectFallbackEntries();
const keys = [...map.keys()].sort();
const records = Buffer.alloc(keys.length * INDEX_RECORD_SIZE);
const strings = [];
let tableOffset = 0;
keys.forEach((key, i) => {
const {url, title, mimetype} = map.get(key);
const keyBuf = Buffer.from(key, 'utf8');
const urlBuf = Buffer.from(url, 'utf8');
const titleBuf = Buffer.from(title, 'utf8');
const base = i * INDEX_RECORD_SIZE;
records.writeUInt32LE(tableOffset, base); records.writeUInt16LE(keyBuf.length, base + 4);
tableOffset += keyBuf.length;
records.writeUInt32LE(tableOffset, base + 6); records.writeUInt16LE(urlBuf.length, base + 10);
tableOffset += urlBuf.length;
records.writeUInt32LE(tableOffset, base + 12); records.writeUInt16LE(titleBuf.length, base + 16);
tableOffset += titleBuf.length;
records.writeUInt16LE(mimetype, base + 18);
strings.push(keyBuf, urlBuf, titleBuf);
});
const header = Buffer.alloc(INDEX_HEADER_SIZE);
header.write(INDEX_MAGIC, 0, 'ascii');
stalenessKey.copy(header, 4);
header.writeUInt32LE(keys.length, 20);
const tmpPath = `${this.#indexPath}.tmp-${process.pid}`;
await fs.promises.writeFile(tmpPath, Buffer.concat([header, records, ...strings]));
await fs.promises.rename(tmpPath, this.#indexPath);
}
/** Opens the fallback index file and caches its header fields for querying. */
async #openFallbackIndex() {
const fd = await fs.promises.open(this.#indexPath, 'r');
const header = Buffer.alloc(INDEX_HEADER_SIZE);
await fd.read(header, 0, INDEX_HEADER_SIZE, 0);
this.#indexFd = fd;
this.#indexRecordCount = header.readUInt32LE(20);
this.#indexTableStart = INDEX_HEADER_SIZE + this.#indexRecordCount * INDEX_RECORD_SIZE;
}
/**
* Validates the on-disk fallback index against the archive's current staleness
* key and (re)builds it if missing/stale. Failures are swallowed - callers see
* #indexFd stay null and fall through to the linear-scan last resort.
*/
async #loadOrBuildFallbackIndex() {
const stalenessKey = await this.#stalenessKey();
try {
const fd = await fs.promises.open(this.#indexPath, 'r');
const header = Buffer.alloc(INDEX_HEADER_SIZE);
await fd.read(header, 0, INDEX_HEADER_SIZE, 0);
const fresh = header.toString('ascii', 0, 4) === INDEX_MAGIC && header.subarray(4, 20).equals(stalenessKey);
if (fresh) {
this.#indexFd = fd;
this.#indexRecordCount = header.readUInt32LE(20);
this.#indexTableStart = INDEX_HEADER_SIZE + this.#indexRecordCount * INDEX_RECORD_SIZE;
return;
}
await fd.close();
} catch { /* missing, corrupt, or unreadable -> rebuild below */ }
await this.#buildFallbackIndex(stalenessKey);
await this.#openFallbackIndex();
}
async #indexRead(pos, length) {
const buf = Buffer.alloc(length);
await this.#indexFd.read(buf, 0, length, pos);
return buf;
}
async #indexString(offset, length) {
if (!length) return '';
return (await this.#indexRead(this.#indexTableStart + offset, length)).toString('utf8');
}
/** Binary search over the on-disk record array. Never loads the full index into memory. */
async #lookupFallbackIndex(name) {
if (!this.#indexFd) return null;
let lo = 0, hi = this.#indexRecordCount - 1;
while (lo <= hi) {
const mid = (lo + hi) >> 1;
const rec = await this.#indexRead(INDEX_HEADER_SIZE + mid * INDEX_RECORD_SIZE, INDEX_RECORD_SIZE);
const keyOff = rec.readUInt32LE(0), keyLen = rec.readUInt16LE(4);
const key = await this.#indexString(keyOff, keyLen);
if (key === name) {
const urlOff = rec.readUInt32LE(6), urlLen = rec.readUInt16LE(10);
const titleOff = rec.readUInt32LE(12), titleLen = rec.readUInt16LE(16);
const [url, title] = await Promise.all([this.#indexString(urlOff, urlLen), this.#indexString(titleOff, titleLen)]);
return {url, title, mimetype: rec.readUInt16LE(18)};
}
if (key < name) lo = mid + 1; else hi = mid - 1;
}
return null;
}
/** Resets a cluster's idle-eviction timer. No-op when TTL disabled. */
#touch(clusterNumber, entry) {
if (!this.#clusterTTL) return;
clearTimeout(entry.timer);
entry.timer = setTimeout(() => this.#clusterCache.delete(clusterNumber), this.#clusterTTL).unref();
}
/** Fetches + decompresses a cluster exactly once, offloading decompression to the worker pool. */
async #loadCluster(clusterNumber) {
const start = await this.#ptr64(this.#header.clusterPtrPos, clusterNumber);
const isLast = clusterNumber === this.#header.clusterCount - 1;
const end = isLast
? (await fs.promises.stat(this.path)).size
: await this.#ptr64(this.#header.clusterPtrPos, clusterNumber + 1);
const raw = await this.#read(start, end - start);
const compType = raw[0] & 0x0f;
const extended = (raw[0] & 0x10) !== 0;
const body = raw.subarray(1);
let data;
if (compType <= 1) data = Buffer.from(body);
else if (compType === 4 || compType === 5) data = await decompressPool.run({compType, body});
else throw new Error(`Unsupported cluster compression type: ${compType}`);
if (!data) throw new Error(`Cluster ${clusterNumber} failed to decompress (compType ${compType})`);
return {data, extended};
}
async #getBlob(clusterNumber, blobNumber) {
let entry = this.#clusterCache.get(clusterNumber);
if (!entry) {
let pending = this.#pending.get(clusterNumber);
if (!pending) {
pending = this.#loadCluster(clusterNumber);
this.#pending.set(clusterNumber, pending);
}
entry = await pending;
this.#pending.delete(clusterNumber);
this.#clusterCache.set(clusterNumber, entry);
if (this.#clusterCache.size > this.#clusterCacheMax) {
const oldestKey = this.#clusterCache.keys().next().value;
clearTimeout(this.#clusterCache.get(oldestKey)?.timer);
this.#clusterCache.delete(oldestKey);
}
}
this.#touch(clusterNumber, entry);
const readPtr = i => entry.extended ? Number(entry.data.readBigUInt64LE(i * 8)) : entry.data.readUInt32LE(i * 4);
return entry.data.subarray(readPtr(blobNumber), readPtr(blobNumber + 1));
}
async #icon(size = 48) {
const page = await this.readPage(`Illustration_${size}x${size}@1`, NS_METADATA)
|| await this.readPage('Favicon', NS_METADATA);
if (!page) return null;
return `data:${page.mimetype};base64,${page.data.toString('base64')}`;
}
async #ptr64(base, index) {
return Number((await this.#read(base + index * 8, 8)).readBigUInt64LE(0));
}
async #read(pos, length) {
const buf = Buffer.alloc(length);
await this.#fd.read(buf, 0, length, pos);
return buf;
}
async #readHeader() {
const b = await this.#read(0, HEADER_SIZE);
const titlePtrRaw = b.readBigUInt64LE(40);
this.#hasTitleListing = titlePtrRaw !== TITLE_SENTINEL;
this.#header = {
articleCount: b.readUInt32LE(24),
clusterCount: b.readUInt32LE(28),
urlPtrPos: Number(b.readBigUInt64LE(32)),
titlePtrPos: this.#hasTitleListing ? Number(titlePtrRaw) : null,
clusterPtrPos: Number(b.readBigUInt64LE(48)),
mimeListPos: Number(b.readBigUInt64LE(56)),
mainPage: b.readUInt32LE(64),
};
}
async #readMimeTypes() {
let pos = this.#header.mimeListPos, str = '';
for (;;) {
str += (await this.#read(pos, 1024)).toString('binary');
const end = str.indexOf('\0\0');
if (end !== -1) { str = str.slice(0, end + 1); break; }
pos += 1024;
}
this.#mimeTypes = str.split('\0').filter(Boolean);
}
/** Directory entry (article record) at byte `offset`, growing the read window as needed. */
async #readDirent(offset) {
for (let size = 512; ; size *= 2) {
const buf = await this.#read(offset, size);
let o = 0;
const mimetype = buf.readUInt16LE(o); o += 2;
o += 1; // extraLen, unused
const namespace = String.fromCharCode(buf.readUInt8(o)); o += 1;
o += 4; // revision, unused
let redirectIndex = null, cluster = null, blob = null;
if (mimetype === 0xffff) { redirectIndex = buf.readUInt32LE(o); o += 4; }
else { cluster = buf.readUInt32LE(o); o += 4; blob = buf.readUInt32LE(o); o += 4; }
const urlEnd = buf.indexOf(0, o);
if (urlEnd === -1) continue;
const titleEnd = buf.indexOf(0, urlEnd + 1);
if (titleEnd === -1) continue;
const url = buf.toString('utf8', o, urlEnd);
const title = buf.toString('utf8', urlEnd + 1, titleEnd) || url;
return {mimetype, namespace, redirectIndex, cluster, blob, url, title};
}
}
async #resolveRedirect(dirent) {
if (dirent.mimetype !== 0xffff) return dirent;
return this.#readDirent(await this.#ptr64(this.#header.urlPtrPos, dirent.redirectIndex));
}
async close() {
if (this.#fd) await this.#fd.close();
this.#fd = null;
if (this.#indexFd) await this.#indexFd.close();
this.#indexFd = null;
for (const entry of this.#clusterCache.values()) clearTimeout(entry.timer);
this.#clusterCache.clear();
this.#pending.clear();
}
/** Deletes the zim archive and its cached fallback index (if any). Safe to call on unopened readers. */
async delete() {
await this.close();
await fs.promises.rm(this.path, {force: true});
await fs.promises.rm(this.#indexPath, {force: true});
}
/** Reads an 'M' namespace metadata value (e.g. Name, Date, Title). Returns null if missing. */
async metadata(key) {
const get = async key => {
const page = await this.readPage(key, NS_METADATA);
return page ? page.data.toString('utf8') : null;
};
if(key) return get(key);
const [title, creator, publisher, date, description, language, name, tags] = await Promise.all(
['Title', 'Creator', 'Publisher', 'Date', 'Description', 'Language', 'Name', 'Tags'].map(get)
);
return {
title,
updated: date ? new Date(date) : null,
summary: description,
language,
name,
category: tags ? tags.split(';')[0] || '' : '',
tags: tags ? tags.split(';') : [],
author: creator,
publisher,
articleCount: this.articleCount,
mediaCount: this.mediaCount,
sizeMb: +((await fs.promises.stat(this.path)).size / 1024 / 1024).toFixed(1),
icon: await this.#icon(),
};
}
/** Opens the archive and parses its header + mimetype list. */
async open() {
this.#fd = await fs.promises.open(this.path, 'r');
try {
await this.#readHeader();
await this.#readMimeTypes();
} catch (e) {
await this.#fd.close();
this.#fd = null;
throw e;
}
// The fallback index is only needed when the archive has no native title
// pointer list - #findByTitleBuiltin already covers that case in O(log n)
// with zero extra storage. Well-maintained archives (Wikipedia etc.)
// ship a title listing, so this path is expected to be rare in practice.
if (!this.#hasTitleListing) {
this.#indexReady = this.#loadOrBuildFallbackIndex().catch(() => {});
}
return this;
}
/** Read a page's content by URL. Returns `{mimetype, data}` or `null` if not found. */
async readPage(url, namespace = NS_CONTENT) {
let dirent = await this.#findByUrl(url, namespace);
if (!dirent) return null;
dirent = await this.#resolveRedirect(dirent);
const data = await this.#getBlob(dirent.cluster, dirent.blob);
return {mimetype: this.#mimeTypes[dirent.mimetype] || 'application/octet-stream', data};
}
/** Reads the archive's designated main/landing page, if one is set. */
async mainPage() {
if (this.#header.mainPage === 0xffffffff) return null;
let dirent = await this.#readDirent(await this.#ptr64(this.#header.urlPtrPos, this.#header.mainPage));
dirent = await this.#resolveRedirect(dirent);
const data = await this.#getBlob(dirent.cluster, dirent.blob);
return {mimetype: this.#mimeTypes[dirent.mimetype] || 'application/octet-stream', data, url: dirent.url, namespace: dirent.namespace};
}
/**
* Fuzzy-ranked title search. Accepts comma-separated `terms` the same way the
* catalog search does. Uses kiwix-search's embedded fulltext index as a prefilter
* to narrow candidates before fuzzy scoring.
*/
async search(terms, {limit = 20, htmlOnly = true} = {}) {
const termList = String(terms).split(',').map(t => t.trim()).filter(Boolean);
if (!termList.length) return [];
const names = await kiwixSearch(this.path, termList.join(' '));
const candidates = (await Promise.all(names.map(n => this.#resolveSearchEntry(n)))).filter(Boolean);
const scored = [];
for (const dirent of candidates) {
if (htmlOnly && !(this.#mimeTypes[dirent.mimetype] || '').startsWith('text/html')) continue;
const urlTitle = titleFromUrl(dirent.url);
const titleScore = fuzzyMatch(dirent.title, ...termList).max;
const urlScore = fuzzyMatch(urlTitle, ...termList).max;
scored.push({url: dirent.url, title: dirent.title.length > urlTitle.length ? dirent.title : urlTitle, namespace: NS_CONTENT, score: Math.max(titleScore, urlScore)});
}
const {summary, mediaCount, articleCount, sizeMb, ...meta} = await this.metadata();
return scored.filter(a => a.score > 0).toSorted((a, b) => b.score - a.score).slice(0, limit).map(a => ({...meta, ...a}));
}
}