Gustavo/gustavoo-me
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

automatic-ai-posts: raise max_tokens for reasoning models (NVIDIA Build)

Browse Source 
Nemotron spends the token budget on reasoning_content before answering;
a 2048 budget was exhausted by ~7761 reasoning chars, returning empty
content with finish_reason=length. Topic selection and category choice go
2048 -> 8192, image plan 4000 -> 8192, article rewrite 8192 -> 16384.
max_tokens is only a cap, so concise answers cost the same as before.
This commit is contained in:
Gustavo committed 2026-11-02 14:51:25 -03:00
1 parent dc50c99dd9
commit 0bb30696de
1 file changed
+12 -5
Download Patch File Download Diff File Expand all files Collapse all files
wp-content/plugins/automatic-ai-posts/src/Modules/Ai/AiService.php
+12 -5
View File
@@ -86,7 +86,7 @@ final class AiService {
}
$prompt = PromptLibrary::select_topics($articles, $min, $max, $recent_titles);
$raw = $this->ask($prompt, 'topic selection', 2048, 0.2);
$raw = $this->ask($prompt, 'topic selection', 8192, 0.2);
if ($raw === null) {
return [];
@@ -117,7 +117,7 @@ final class AiService {
*/
public function rewrite(array $article, string $language): ?array {
$prompt = PromptLibrary::rewrite($article, $language);
$raw = $this->ask($prompt, 'article rewrite', 8192, 0.8);
$raw = $this->ask($prompt, 'article rewrite', 16384, 0.8);
if ($raw === null) {
return null;
@@ -148,7 +148,7 @@ final class AiService {
}
$prompt = PromptLibrary::choose_category($title, $content, $categories, $language);
$raw = $this->ask($prompt, 'category choice', 2048, 0.2);
$raw = $this->ask($prompt, 'category choice', 8192, 0.2);
if ($raw === null) {
return 0;
@@ -182,7 +182,7 @@ final class AiService {
*/
public function image_plan(string $title, string $content, string $language): ?array {
$prompt = PromptLibrary::image_plan($title, $content, $language);
$raw = $this->ask($prompt, 'image plan', 4000, 0.7);
$raw = $this->ask($prompt, 'image plan', 8192, 0.7);
if ($raw === null) {
return null;
@@ -205,7 +205,14 @@ final class AiService {
];
}
/** Calls the provider with retries. Returns null when every attempt failed. */
/**
* Calls the provider with retries. Returns null when every attempt failed.
*
* Budgets are generous on purpose: reasoning models (e.g. Nemotron on
* NVIDIA Build) spend tokens on reasoning_content before answering, and an
* exhausted budget returns empty content with finish_reason "length".
* max_tokens is only a cap — the cost is per generated token.
*/
private function ask(string $prompt, string $label, int $max_tokens, float $temperature): ?string {
if (!$this->provider) {
Logger::error('No AI provider configured.', 'ai');