ghassan/takeone-youtube-clone
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Add AI translation for comments and video descriptions

Browse Source 
Comments get a Translate button that calls the active LLM provider in Admin → AI
settings. Each commenter shows a small country flag inferred from their
video_views + profile_visits history (falls back to users.nationality); that
inferred country drives the language hint passed to the LLM so Portuguese
"Linda, linda, linda" translates to "Beautiful" instead of pass-through.

Description boxes get a line-paired translation grid: a shared Translate button
flips #vdbDescSplit into a two-column CSS grid where every source line sits
next to its translation on the same row. Backend splits by newline and asks
the LLM for a JSON map keyed by index so counts always match.

LlmLyricsService now has translate() + translateLines() with echo-retry (if the
model returns the input unchanged while source != target, retry once with a
stronger directive) and keeps the Ollama model resident via keep_alive=1h.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>
This commit is contained in:
ghassanandClaude Opus 4.7 committed 2026-10-02 12:18:20 +03:00
1 parent 20c8fdcdf3
commit 2d33e39b9b
8 files changed
+749 -22
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

app/Services/LlmLyricsService.php
+190 -16
View File
@@ -191,16 +191,176 @@ class LlmLyricsService
}
/**
* Translate arbitrary text (e.g. a user comment) to a target language.
* Returns the translated string, or '' on any failure / when disabled.
*/
public function translate(string $text, string $targetLang = 'en', ?string $sourceLangHint = null): string
{
$text = trim($text);
if ($text === '' || ! $this->isEnabled()) return '';
$provider = $this->activeProvider();
$targetLang = preg_replace('/[^a-zA-Z-]/', '', $targetLang) ?: 'en';
$sourceLangHint = $sourceLangHint ? (preg_replace('/[^a-zA-Z-]/', '', $sourceLangHint) ?: null) : null;
$cacheKey = 'llm_translate_v4:' . ($provider['id'] ?? '') . ':' . $targetLang . ':' . ($sourceLangHint ?? '-') . ':' . sha1($text);
$cached = Cache::get($cacheKey);
if (is_string($cached) && $cached !== '') return $cached;
$hint = $sourceLangHint
? "The author's written language is \"{$sourceLangHint}\" (ISO-639). Assume the text is in that language unless it is OBVIOUSLY in a different language (e.g. uses a different script).\n"
: '';
$prompt = "Translate the following text into the language with ISO-639 code \"{$targetLang}\".\n"
. $hint
. "Rules:\n"
. " - Output ONLY the translated text. No prose, no quotes, no language label, no notes.\n"
. " - Translate EVERY word of the source language, including short exclamations,\n"
. " repeated words, and words that happen to also look like a name in the target\n"
. " language. Example: Portuguese \"Linda, linda, linda\" → \"Beautiful, beautiful,\n"
. " beautiful\" (NOT the English name \"Linda\"). Spanish \"hola\" → \"hello\".\n"
. " - A word is a proper noun ONLY if clearly a person/place/brand name in context\n"
. " (e.g. \"My name is Linda\"). If it stands alone or is repeated as an exclamation,\n"
. " it is almost certainly a regular word — translate it.\n"
. " - If the text is already in the target language, repeat it unchanged.\n"
. " - Preserve emojis, @mentions (e.g. @alice), hashtags, URLs, and timestamp tokens\n"
. " like @1:23 or @01:23-02:45 EXACTLY as written.\n"
. " - Keep the same tone and line breaks.\n\n"
. "TEXT:\n" . $text;
// Translation is deterministic — temp 0. If the model echoes the input
// (common when a word happens to also exist as a name in the target
// language, e.g. Portuguese "Linda" → English "Linda"), retry once
// with a stronger directive. If the retry also echoes, we return it.
// Comments cap at 1000 chars → 400 tokens is plenty of headroom and
// keeps the model from drifting into long completions on tiny inputs.
$budget = max(128, min(1024, (int) ceil(mb_strlen($text) * 1.5) + 64));
$raw = $this->postProcess($this->call($prompt, $budget, 0.0));
$needsRetry = $sourceLangHint
&& $sourceLangHint !== $targetLang
&& self::looksLikeEcho($text, $raw);
if ($needsRetry) {
$retryPrompt = $prompt
. "\n\nIMPORTANT: Your previous attempt was identical to the source. The source is\n"
. "DEFINITELY in \"{$sourceLangHint}\", not in \"{$targetLang}\". Translate every word.";
$retry = $this->postProcess($this->call($retryPrompt, $budget, 0.0));
if ($retry !== '' && ! self::looksLikeEcho($text, $retry)) $raw = $retry;
}
if ($raw === '') return '';
Cache::put($cacheKey, $raw, now()->addDays(30));
return $raw;
}
private function postProcess(string $raw): string
{
$raw = trim($raw);
$raw = trim(preg_replace('/^```[a-zA-Z]*\s*|\s*```$/m', '', $raw));
return $raw;
}
/** Normalised equality: ignore case, whitespace, punctuation, emojis. */
private static function looksLikeEcho(string $src, string $out): bool
{
if ($out === '') return false;
$norm = fn (string $s) => mb_strtolower(preg_replace('/[^\p{L}\p{N}]+/u', '', $s) ?? '');
$a = $norm($src); $b = $norm($out);
return $a !== '' && $a === $b;
}
/**
* Translate a sequence of lines, preserving the one-to-one mapping so the
* caller can render them side-by-side with the originals. Empty entries
* are kept as empty strings. Returns an array of the same length as $lines
* (or [] on any failure / when disabled).
*/
public function translateLines(array $lines, string $targetLang = 'en', ?string $sourceLangHint = null): array
{
if (! $lines || ! $this->isEnabled()) return [];
$targetLang = preg_replace('/[^a-zA-Z-]/', '', $targetLang) ?: 'en';
$sourceLangHint = $sourceLangHint ? (preg_replace('/[^a-zA-Z-]/', '', $sourceLangHint) ?: null) : null;
$provider = $this->activeProvider();
$cacheKey = 'llm_translate_lines_v2:' . ($provider['id'] ?? '') . ':' . $targetLang . ':' . ($sourceLangHint ?? '-') . ':' . sha1(json_encode($lines));
$cached = Cache::get($cacheKey);
if (is_array($cached) && count($cached) === count($lines)) return $cached;
// Index non-empty lines so empty separators are preserved verbatim.
$toTranslate = [];
foreach ($lines as $i => $l) {
if (trim((string) $l) !== '') $toTranslate[$i] = (string) $l;
}
if (! $toTranslate) {
Cache::put($cacheKey, $lines, now()->addDays(30));
return $lines;
}
$numbered = [];
foreach ($toTranslate as $i => $l) $numbered[] = $i . "\t" . $l;
$hint = $sourceLangHint
? "The author's written language is \"{$sourceLangHint}\" (ISO-639). Assume the lines are in that language unless they are OBVIOUSLY in a different language (different script).\n"
: '';
$prompt = "Translate each of the following numbered lines into the language with ISO-639 code \"{$targetLang}\".\n"
. $hint
. "Rules:\n"
. " - Output a JSON object mapping each original index (as string) to its translated line.\n"
. " - Keep ONE output line per input line — do NOT merge, split, re-order, or drop any.\n"
. " - Translate every word including short exclamations and repeated words that happen\n"
. " to look like a name in the target language (Portuguese \"Linda\" → \"Beautiful\").\n"
. " - If a line is already in the target language, repeat it unchanged.\n"
. " - Preserve emojis, @mentions, hashtags, URLs, and timestamp tokens like @1:23 EXACTLY.\n"
. " - Respond with ONLY the JSON object. No prose, no markdown, no code fence.\n\n"
. "LINES (index TAB text):\n" . implode("\n", $numbered);
$raw = $this->postProcess($this->call($prompt, 8192, 0.0));
$obj = json_decode($raw, true);
if (! is_array($obj)) return [];
$out = $lines;
$echoCount = 0;
$translatedCount = 0;
foreach ($toTranslate as $i => $src) {
$v = $obj[(string) $i] ?? ($obj[$i] ?? null);
if (is_string($v) && trim($v) !== '') {
$out[$i] = $v;
$translatedCount++;
if ($sourceLangHint && $sourceLangHint !== $targetLang && self::looksLikeEcho($src, $v)) {
$echoCount++;
}
}
}
// If > 30% of the translated lines just echo the source, nudge the model.
if ($sourceLangHint && $sourceLangHint !== $targetLang
&& $translatedCount > 0 && $echoCount / $translatedCount > 0.3) {
$retryPrompt = $prompt . "\n\nIMPORTANT: Your previous attempt echoed the source for many lines. "
. "The source is DEFINITELY in \"{$sourceLangHint}\", not in \"{$targetLang}\". Translate every word.";
$raw2 = $this->postProcess($this->call($retryPrompt, 8192, 0.0));
$obj2 = json_decode($raw2, true);
if (is_array($obj2)) {
foreach ($toTranslate as $i => $src) {
$v = $obj2[(string) $i] ?? ($obj2[$i] ?? null);
if (is_string($v) && trim($v) !== '' && ! self::looksLikeEcho($src, $v)) {
$out[$i] = $v;
}
}
}
}
Cache::put($cacheKey, $out, now()->addDays(30));
return $out;
}
/** Dispatch to the active provider's adapter. */
private function call(string $prompt, int $maxTokens): string
private function call(string $prompt, int $maxTokens, float $temperature = 0.2): string
{
$p = $this->activeProvider();
if (! $p) return '';
try {
return match ($p['kind']) {
'anthropic' => $this->callAnthropic($p, $prompt, $maxTokens),
'openai' => $this->callOpenAI($p, $prompt, $maxTokens),
default => $this->callOllama($p, $prompt, $maxTokens),
'anthropic' => $this->callAnthropic($p, $prompt, $maxTokens, $temperature),
'openai' => $this->callOpenAI($p, $prompt, $maxTokens, $temperature),
default => $this->callOllama($p, $prompt, $maxTokens, $temperature),
};
} catch (\Throwable $e) {
Log::error('LLM call failed: ' . $e->getMessage(), ['provider' => $p['name'] ?? '?']);
@@ -208,24 +368,37 @@ class LlmLyricsService
}
}
private function callOllama(array $p, string $prompt, int $maxTokens): string
private function callOllama(array $p, string $prompt, int $maxTokens, float $temperature = 0.2): string
{
$endpoint = rtrim((string) ($p['endpoint'] ?? 'http://localhost:11434'), '/');
// Qwen3-style "thinking" models respect /no_think to skip the hidden
// reasoning phase. Without it they can exhaust num_predict inside
// the thinking block and emit empty content.
$taggedPrompt = $prompt . "\n\n/no_think";
$resp = Http::timeout(180)->acceptJson()->post($endpoint . '/api/chat', [
'model' => $p['model'],
'messages' => [['role' => 'user', 'content' => $prompt]],
'stream' => false,
'options' => ['num_predict' => $maxTokens, 'temperature' => 0.2],
'model' => $p['model'],
'messages' => [['role' => 'user', 'content' => $taggedPrompt]],
'stream' => false,
'think' => false,
// Keep the model resident for an hour so follow-up requests skip
// the cold-start reload (default is 5 minutes). Dramatically
// reduces per-click latency for comment/description translation.
'keep_alive' => '1h',
'options' => ['num_predict' => $maxTokens, 'temperature' => $temperature],
]);
if (! $resp->successful()) {
Log::warning('Ollama API error', ['status' => $resp->status(), 'body' => substr($resp->body(), 0, 500)]);
return '';
}
$j = $resp->json();
return (string) ($j['message']['content'] ?? '');
$content = (string) ($j['message']['content'] ?? '');
// Some thinking models still ignore the flag and leak <think>…</think>
// tags into content. Strip them.
$content = preg_replace('/<think>.*?<\/think>\s*/is', '', $content) ?? $content;
return trim($content);
}
private function callAnthropic(array $p, string $prompt, int $maxTokens): string
private function callAnthropic(array $p, string $prompt, int $maxTokens, float $temperature = 0.2): string
{
$endpoint = rtrim((string) ($p['endpoint'] ?? 'https://api.anthropic.com'), '/');
$resp = Http::timeout(120)->withHeaders([
@@ -233,9 +406,10 @@ class LlmLyricsService
'anthropic-version' => '2023-06-01',
'content-type' => 'application/json',
])->post($endpoint . '/v1/messages', [
'model' => $p['model'],
'max_tokens' => $maxTokens,
'messages' => [['role' => 'user', 'content' => $prompt]],
'model' => $p['model'],
'max_tokens' => $maxTokens,
'temperature' => $temperature,
'messages' => [['role' => 'user', 'content' => $prompt]],
]);
if (! $resp->successful()) {
Log::warning('Anthropic API error', ['status' => $resp->status(), 'body' => substr($resp->body(), 0, 500)]);
@@ -245,7 +419,7 @@ class LlmLyricsService
return (string) ($j['content'][0]['text'] ?? '');
}
private function callOpenAI(array $p, string $prompt, int $maxTokens): string
private function callOpenAI(array $p, string $prompt, int $maxTokens, float $temperature = 0.2): string
{
$endpoint = rtrim((string) ($p['endpoint'] ?? 'https://api.openai.com'), '/');
$resp = Http::timeout(120)->withToken((string) $p['api_key'])
@@ -254,7 +428,7 @@ class LlmLyricsService
'model' => $p['model'],
'messages' => [['role' => 'user', 'content' => $prompt]],
'max_tokens' => $maxTokens,
'temperature' => 0.2,
'temperature' => $temperature,
]);
if (! $resp->successful()) {
Log::warning('OpenAI API error', ['status' => $resp->status(), 'body' => substr($resp->body(), 0, 500)]);