47a930eddb
Node's default fetch sends a bare User-Agent that Google's CDN fronting ollama.com rejects with 403. curl worked because it sends curl/. Fixes organize (and search LLM ranking). Signed-off-by: William Mantly <wmantly@gmail.com>
365 lines
13 KiB
JavaScript
365 lines
13 KiB
JavaScript
'use strict';
|
|
|
|
const conf = require('>/conf');
|
|
const tmdb = require('>/controller/tmdb');
|
|
const { tmdbSearch, tmdbDetails } = tmdb;
|
|
const library = require('>/controller/library');
|
|
const ollama = require('>/controller/ollama');
|
|
|
|
// TPB numeric categories that count as Movies / TV( used for the q.php search and to
|
|
// keep the LLM from ever seeing non-video results).
|
|
const MOVIE_CATS = [201, 202, 207, 209, 211];
|
|
const TV_CATS = [205, 208, 212];
|
|
|
|
function humanSize(bytes){
|
|
bytes = Number(bytes) || 0;
|
|
let units = ['B', 'KB', 'MB', 'GB', 'TB'];
|
|
let i = 0;
|
|
while(bytes >= 1024 && i < units.length - 1){
|
|
bytes /= 1024;
|
|
i++;
|
|
}
|
|
return `${bytes.toFixed(bytes < 10 && i > 0 ? 1 : 0)} ${units[i]}`;
|
|
}
|
|
|
|
async function fetchJSON(url, options){
|
|
let res = await fetch(url, options);
|
|
if(!res.ok){
|
|
let error = new Error(`UpstreamError`);
|
|
error.message = `Request to ${url.split('?')[0]} failed( ${res.status})`;
|
|
error.status = 502;
|
|
throw error;
|
|
}
|
|
return await res.json();
|
|
}
|
|
|
|
// Turn a raw TPB category id into 'movie' | 'tv' | null.
|
|
function videoKind(category){
|
|
category = parseInt(category, 10);
|
|
if(TV_CATS.includes(category)) return 'tv';
|
|
if(MOVIE_CATS.includes(category) || (category >= 200 && category < 300)) return 'movie';
|
|
return null;
|
|
}
|
|
|
|
// --- TPB HTML search ------------------------------------------------------
|
|
|
|
const SIZE_UNITS = { B: 1, KIB: 1024, MIB: 1024 ** 2, GIB: 1024 ** 3, TIB: 1024 ** 4 };
|
|
|
|
function sizeToBytes(value, unit){
|
|
return Math.round(parseFloat(value) * (SIZE_UNITS[String(unit).toUpperCase()] || 1));
|
|
}
|
|
|
|
function decodeEntities(text){
|
|
return String(text)
|
|
.replace(/&/g, '&').replace(/</g, '<').replace(/>/g, '>')
|
|
.replace(/"/g, '"').replace(/�?39;/g, "'").replace(/'/g, "'");
|
|
}
|
|
|
|
// Parse the classic TPB HTML result table( category cell, details link name, magnet,
|
|
// then right-aligned size / seeders / leechers cells) into apibay-shaped objects.
|
|
function parseTPBHtml(html){
|
|
let out = [];
|
|
for(let row of html.split(/<tr/i)){
|
|
let hash = row.match(/magnet:\?xt=urn:btih:([A-Fa-f0-9]+)/i);
|
|
if(!hash) continue;
|
|
|
|
let cat = row.match(/\/browse\/(\d+)/);
|
|
let name = row.match(/title="Details for ([^"]+)"/i);
|
|
let size = row.match(/<td align="right">\s*([\d.]+)(?: |\s)+(GiB|MiB|KiB|TiB)\s*<\/td>/i);
|
|
let nums = [...row.matchAll(/<td align="right">\s*([\d,]+)\s*<\/td>/gi)].map(m => Number(m[1].replace(/,/g, '')));
|
|
|
|
if(!name) continue;
|
|
// nums are [size, seeders, leechers]; size cell also matched the regex above so
|
|
// seeders/leechers are the last two numeric right-aligned cells.
|
|
let seeders = nums.length >= 2 ? nums[nums.length - 2] : 0;
|
|
let leechers = nums.length >= 1 ? nums[nums.length - 1] : 0;
|
|
|
|
out.push({
|
|
name: decodeEntities(name[1]),
|
|
info_hash: hash[1].toLowerCase(),
|
|
category: cat ? Number(cat[1]) : 0,
|
|
size: size ? sizeToBytes(size[1], size[2]) : 0,
|
|
seeders: seeders,
|
|
leechers: leechers,
|
|
imdb: null,
|
|
});
|
|
}
|
|
return out;
|
|
}
|
|
|
|
async function tpbSearch(title){
|
|
// Category 200 = all Video; we narrow to movie/tv subcats in prefilter().
|
|
let url = `${conf.search.tpbBase}/search/${encodeURIComponent(title)}/1/99/200`;
|
|
let res = await fetch(url, { headers: { 'User-Agent': 'Mozilla/5.0' } });
|
|
if(!res.ok){
|
|
let error = new Error('UpstreamError');
|
|
error.message = `TPB search failed( ${res.status})`;
|
|
error.status = 502;
|
|
throw error;
|
|
}
|
|
return parseTPBHtml(await res.text());
|
|
}
|
|
|
|
// Low-quality source tags we never want to recommend( token match to avoid false hits
|
|
// like "GHOSTS" matching "TS").
|
|
const JUNK = new Set(['cam', 'hdcam', 'camrip', 'ts', 'hdts', 'telesync', 'tc', 'telecine', 'scr', 'screener', 'dvdscr', 'workprint']);
|
|
function isJunk(name){
|
|
return String(name).toLowerCase().split(/[.\s_\-\[\]()]+/).some(tok => JUNK.has(tok));
|
|
}
|
|
|
|
// Deterministic pre-filter before the torrents ever reach the LLM.
|
|
function prefilter(torrents, imdbId){
|
|
let out = torrents.filter(t => videoKind(t.category) && Number(t.seeders) > 0 && !isJunk(t.name));
|
|
|
|
// If we can positively identify the title by IMDB id, trust it and drop the rest.
|
|
if(imdbId){
|
|
let matches = out.filter(t => t.imdb && t.imdb === imdbId);
|
|
if(matches.length) out = matches;
|
|
}
|
|
|
|
return out
|
|
.sort((a, b) => Number(b.seeders) - Number(a.seeders))
|
|
.slice(0, 40);
|
|
}
|
|
|
|
function buildMagnet(infoHash, name){
|
|
let magnet = `magnet:?xt=urn:btih:${infoHash}&dn=${encodeURIComponent(name)}`;
|
|
for(let tracker of conf.search.trackers){
|
|
magnet += `&tr=${encodeURIComponent(tracker)}`;
|
|
}
|
|
return magnet;
|
|
}
|
|
|
|
// --- LLM ranking ----------------------------------------------------------
|
|
|
|
function buildSystemPrompt(title, year, today){
|
|
return [
|
|
`You are a torrent-selection assistant. Today's date is ${today}.`,
|
|
`Recent releases are legitimate: NEVER reject a torrent for being too new or assume`,
|
|
`a current-year release is fake.`,
|
|
``,
|
|
`The user wants "${title}" (${year || 'unknown year'}). From the candidate list,`,
|
|
`select the best releases.`,
|
|
year ? `IMPORTANT: only pick releases whose name contains the year ${year}; reject other years or remakes of a same-named title.` : ``,
|
|
`Preferences, in order:`,
|
|
`- video codec HEVC/x265 (over x264/AVC)`,
|
|
`- 1080p resolution`,
|
|
`- total size around 1.5 GB (avoid needlessly huge files)`,
|
|
`- English audio and English subtitles present`,
|
|
`- prefer EXTENDED / UNRATED / UNCUT / DIRECTOR'S CUT editions`,
|
|
`Reject: CAM, TS, TC, TELESYNC, HDCAM, SCREENER/SCR, and non-English-only copies.`,
|
|
``,
|
|
`Return STRICT JSON only, shape:`,
|
|
`{"options":[{"role":"recommended|uhd|other","label":"...","info_hash":"...",`,
|
|
`"name":"...","resolution":"...","codec":"...","hasSubs":true,"cut":"...",`,
|
|
`"warning":"...","why":"..."}]}`,
|
|
`Rules: exactly one "recommended" (the 1080p sweet spot). Include one "uhd" ONLY`,
|
|
`if a 4K/2160p release exists, and set its "warning" to note it is a much larger,`,
|
|
`slower download. Add "other" entries for genuinely distinct useful releases`,
|
|
`(e.g. a different cut or a much smaller copy). "why" is one short sentence.`,
|
|
`Copy info_hash verbatim from the chosen candidate.`,
|
|
].join('\n');
|
|
}
|
|
|
|
// Pull the first JSON object out of a model response( tolerates ```json fences / prose).
|
|
function extractJSON(content){
|
|
let text = String(content).replace(/```json/gi, '').replace(/```/g, '').trim();
|
|
let start = text.indexOf('{');
|
|
let end = text.lastIndexOf('}');
|
|
if(start === -1 || end === -1) throw new Error('No JSON in model response');
|
|
return JSON.parse(text.slice(start, end + 1));
|
|
}
|
|
|
|
async function rankReleases(candidates, meta){
|
|
let compact = candidates.map(t => ({
|
|
name: t.name,
|
|
sizeHuman: humanSize(t.size),
|
|
seeders: Number(t.seeders),
|
|
info_hash: t.info_hash,
|
|
imdb: t.imdb || null,
|
|
}));
|
|
|
|
let body = {
|
|
model: conf.ollama.model,
|
|
stream: false,
|
|
format: 'json',
|
|
messages: [
|
|
{ role: 'system', content: buildSystemPrompt(meta.title, meta.year, meta.today) },
|
|
{ role: 'user', content: JSON.stringify(compact) },
|
|
],
|
|
};
|
|
|
|
let data = await fetchJSON(`${conf.ollama.url}/api/chat`, {
|
|
method: 'POST',
|
|
headers: {
|
|
'Content-Type': 'application/json',
|
|
'Authorization': `Bearer ${conf.ollama.apiKey}`,
|
|
'User-Agent': 'tpbproxy/1.0 (search; +https://github.com/wmantly/tpbproxy)',
|
|
},
|
|
body: JSON.stringify(body),
|
|
});
|
|
|
|
let parsed = extractJSON(data.message && data.message.content);
|
|
let options = Array.isArray(parsed.options) ? parsed.options : [];
|
|
if(!options.length) throw new Error('Model returned no options');
|
|
return options;
|
|
}
|
|
|
|
// Deterministic fallback if the LLM is unavailable or returns junk: pick the most-seeded
|
|
// candidate, preferring x265 + 1080p, so the feature degrades instead of failing.
|
|
function fallbackReleases(candidates){
|
|
let score = t => (/x265|hevc/i.test(t.name) ? 2 : 0) + (/1080p/i.test(t.name) ? 1 : 0);
|
|
let best = [...candidates].sort((a, b) =>
|
|
(score(b) - score(a)) || (Number(b.seeders) - Number(a.seeders))
|
|
)[0];
|
|
if(!best) return [];
|
|
return [{
|
|
role: 'recommended',
|
|
label: 'Most seeded',
|
|
info_hash: best.info_hash,
|
|
name: best.name,
|
|
sizeHuman: humanSize(best.size),
|
|
seeders: Number(best.seeders),
|
|
why: 'Automatically chosen (AI ranking unavailable): most seeders.',
|
|
}];
|
|
}
|
|
|
|
// Attach the fields the front end / add flow need to each option and drop any the model
|
|
// hallucinated( info_hash must exist in the candidate set).
|
|
function decorateOptions(options, candidates){
|
|
let byHash = {};
|
|
for(let t of candidates) byHash[String(t.info_hash).toLowerCase()] = t;
|
|
|
|
let out = [];
|
|
for(let opt of options){
|
|
let src = byHash[String(opt.info_hash).toLowerCase()];
|
|
if(!src) continue;
|
|
out.push({
|
|
...opt,
|
|
info_hash: src.info_hash,
|
|
name: opt.name || src.name,
|
|
category: Number(src.category),
|
|
sizeHuman: opt.sizeHuman || humanSize(src.size),
|
|
seeders: Number(src.seeders),
|
|
magnetLink: buildMagnet(src.info_hash, src.name),
|
|
});
|
|
}
|
|
return out;
|
|
}
|
|
|
|
async function findReleases(tmdbId, mediaType){
|
|
let meta = await tmdbDetails(tmdbId, mediaType);
|
|
let torrents = await tpbSearch(meta.title);
|
|
let candidates = prefilter(torrents, meta.imdbId);
|
|
|
|
// For movies, require the confirmed release year in the name. TPB search is fuzzy
|
|
// (a "Toy Story 5" query returns Toy Story 4/3/…), and without this the fallback can
|
|
// cross to the wrong movie. If nothing matches, we genuinely have no good copy yet.
|
|
if(mediaType !== 'tv' && meta.year){
|
|
candidates = candidates.filter(t => t.name.includes(meta.year));
|
|
}
|
|
|
|
if(!candidates.length){
|
|
return { title: meta.title, year: meta.year, tmdbId, mediaType, options: [] };
|
|
}
|
|
|
|
let today = new Date().toISOString().slice(0, 10);
|
|
let options;
|
|
try{
|
|
options = await rankReleases(candidates, { title: meta.title, year: meta.year, today });
|
|
}catch(error){
|
|
console.error('rankReleases failed, using fallback:', error.message);
|
|
options = fallbackReleases(candidates);
|
|
}
|
|
|
|
options = decorateOptions(options, candidates);
|
|
if(!options.length) options = decorateOptions(fallbackReleases(candidates), candidates);
|
|
|
|
let result = { title: meta.title, year: meta.year, tmdbId, mediaType, options };
|
|
|
|
// Quality-aware dedup: flag options you already own at equal-or-better quality.
|
|
if(mediaType !== 'tv'){
|
|
let lib = await library.movieInLibrary(tmdbId);
|
|
if(lib.owned){
|
|
for(let o of options) o.alreadyOwned = library.rankFromLabel(o.resolution) <= lib.rank;
|
|
result.library = { owned: true, quality: lib.quality };
|
|
}
|
|
}
|
|
|
|
return result;
|
|
}
|
|
|
|
// --- TV: library-aware season fill -----------------------------------------
|
|
|
|
function buildSeasonPrompt(title, season, today){
|
|
return [
|
|
`You are a torrent selector. Today's date is ${today}.`,
|
|
`Pick the single best COMPLETE season pack for "${title}" Season ${season}.`,
|
|
`It MUST be the full season (an S${String(season).padStart(2, '0')} / "Season ${season}"`,
|
|
`complete pack), NOT a single episode. Prefer HEVC/x265, 1080p, English audio +`,
|
|
`English subtitles, reasonable size. Reject single episodes, wrong seasons,`,
|
|
`CAM/TS/telesync, and non-English-only.`,
|
|
`Return STRICT JSON {"info_hash":"<from candidates>","why":"one short sentence"};`,
|
|
`if nothing suitable, {"info_hash":null}.`,
|
|
].join('\n');
|
|
}
|
|
|
|
async function bestSeasonPack(candidates, meta){
|
|
let compact = candidates.map(t => ({ name: t.name, sizeHuman: humanSize(t.size), seeders: Number(t.seeders), info_hash: t.info_hash }));
|
|
let parsed = await ollama.chatJSON(buildSeasonPrompt(meta.title, meta.season, meta.today), JSON.stringify(compact));
|
|
return parsed && parsed.info_hash ? parsed : null;
|
|
}
|
|
|
|
// Compare TMDB's season/episode counts against the Emby inventory to classify each season.
|
|
async function seasonPlan(tmdbId){
|
|
let [tv, inv] = await Promise.all([ tmdb.tmdbSeasons(tmdbId), library.seriesInventory(tmdbId) ]);
|
|
let seasons = tv.seasons.map(s => {
|
|
let have = Number(inv.have && inv.have[s.season]) || 0;
|
|
let status = have === 0 ? 'missing' : (have >= s.episodeCount ? 'complete' : 'partial');
|
|
return { season: s.season, episodeCount: s.episodeCount, haveCount: have, status };
|
|
});
|
|
return { title: tv.title, year: tv.year, seasons, missing: seasons.filter(s => s.status !== 'complete').map(s => s.season) };
|
|
}
|
|
|
|
// For every missing/incomplete season, pick the best complete-season pack. Returns the
|
|
// chosen packs( the route adds them to Transmission) plus any seasons it couldn't fill.
|
|
async function fillMissing(tmdbId){
|
|
let plan = await seasonPlan(tmdbId);
|
|
let details = await tmdbDetails(tmdbId, 'tv');
|
|
let title = details.title || plan.title;
|
|
let today = new Date().toISOString().slice(0, 10);
|
|
|
|
let picks = [], skipped = [];
|
|
for(let season of plan.missing){
|
|
let cands = prefilter(await tpbSearch(`${title} S${String(season).padStart(2, '0')}`), null);
|
|
if(!cands.length){ skipped.push({ season, reason: 'no results' }); continue; }
|
|
|
|
let pick = null;
|
|
try{ pick = await bestSeasonPack(cands, { title, season, today }); }catch(error){ pick = null; }
|
|
let src = pick && cands.find(c => c.info_hash.toLowerCase() === String(pick.info_hash).toLowerCase());
|
|
if(!src){ skipped.push({ season, reason: 'no suitable season pack' }); continue; }
|
|
|
|
picks.push({ season, name: src.name, info_hash: src.info_hash, category: Number(src.category), magnetLink: buildMagnet(src.info_hash, src.name), why: pick.why });
|
|
}
|
|
return { title, year: plan.year, picks, skipped };
|
|
}
|
|
|
|
module.exports = {
|
|
tmdbSearch,
|
|
findReleases,
|
|
seasonPlan,
|
|
fillMissing,
|
|
// exported for unit tests
|
|
prefilter,
|
|
buildMagnet,
|
|
humanSize,
|
|
videoKind,
|
|
extractJSON,
|
|
decorateOptions,
|
|
fallbackReleases,
|
|
tpbSearch,
|
|
parseTPBHtml,
|
|
sizeToBytes,
|
|
isJunk,
|
|
};
|