← All changes | includes/ai/class-content-brief-generator.php +42 -4 2.7.0 → 2.9.0 View file →

		@@ -153,11 +153,31 @@
153	153	if ($api_key) {
154	154	$model = $this->settings->get('openrouter_model', Settings::DEFAULT_OPENROUTER_MODEL);
155	155	$this->ai_client = new OpenRouter_Client($api_key, $model, self::AI_REQUEST_TIMEOUT);
156	156	}
	157	+ } elseif ($provider === 'openai_compatible') {
	158	+ // Same client as OpenAI, different host — and the key is optional,
	159	+ // so the URL and model id are what gate it (#721). The user's own
	160	+ // timeout applies: a local model writing a brief on CPU is slow,
	161	+ // and the setting exists for exactly that.
	162	+ $base_url = (string) $this->settings->get('openai_compatible_base_url', '');
	163	+ $model = trim((string) $this->settings->get('openai_compatible_model', ''));
	164	+ if ('' !== $base_url && '' !== $model) {
	165	+ $this->ai_client = new OpenAI_Client(
	166	+ (string) $this->settings->get('openai_compatible_api_key', ''),
	167	+ $model,
	168	+ (int) $this->settings->get('openai_compatible_timeout', Settings::DEFAULT_OPENAI_COMPATIBLE_TIMEOUT),
	169	+ $base_url
	170	+ );
	171	+ $this->ai_client->set_json_mode((bool) $this->settings->get('openai_compatible_json_mode', false));
	172	+ }
157	173	}
158	174
159	175	if (!$this->ai_client) {
	176	+ if ('openai_compatible' === $provider) {
	177	+ throw new \Exception('Please set the base URL and model id for your OpenAI-compatible endpoint in ThinkRank settings.');
	178	+ }
	179	+
160	180	throw new \Exception('Please configure your AI provider API key in ThinkRank settings.');
161	181	}
162	182	}
163	183
		@@ -185,8 +205,10 @@
185	205	} elseif ($provider === 'gemini') {
186	206	return $this->settings->get('gemini_model', Settings::DEFAULT_GEMINI_MODEL);
187	207	} elseif ($provider === 'openrouter') {
188	208	return $this->settings->get('openrouter_model', Settings::DEFAULT_OPENROUTER_MODEL);
	209	+ } elseif ($provider === 'openai_compatible') {
	210	+ return (string) $this->settings->get('openai_compatible_model', '');
189	211	} else {
190	212	return $this->settings->get('openai_model', Settings::DEFAULT_OPENAI_MODEL);
191	213	}
192	214	}
		@@ -245,9 +267,9 @@
245	267	*/
246	268	private function extract_token_usage(array $ai_response): int {
247	269	$provider = $this->get_current_provider();
248	270
249		- if ($provider === 'openai' || $provider === 'openrouter') {
	271	+ if ($provider === 'openai' || $provider === 'openrouter' || $provider === 'openai_compatible') {
250	272	// OpenAI-compatible format: response['usage']['total_tokens']
251	273	return (int) ($ai_response['usage']['total_tokens'] ?? 0);
252	274	} elseif ($provider === 'claude') {
253	275	// Claude format: response['usage']['input_tokens'] + response['usage']['output_tokens']
		@@ -347,8 +369,11 @@
347	369	// max_completion_tokens internally. Temperature is intentionally
348	370	// omitted: every client defaults it to 0.7, and reasoning models
349	371	// reject it outright, so passing it here was misleading no-op.
350	372	'max_tokens' => $max_tokens,
	373	+ // The brief is one JSON object. Only a compatible endpoint
	374	+ // with JSON mode on reads this; every other client ignores it.
	375	+ 'json_object' => true,
351	376	];
352	377	if ('' !== $reasoning_effort) {
353	378	$completion_options['reasoning_effort'] = $reasoning_effort;
354	379	}
		@@ -360,9 +385,9 @@
360	385	// truncation) BEFORE attempting text extraction. Otherwise a
361	386	// refusal — which OpenAI returns as HTTP 200 with content=null —
362	387	// slips past every isset() branch and gets serialized into the
363	388	// brief body instead of being reported to the user.
364		- $this->guard_against_non_answer($ai_response);
	389	+ $this->guard_against_non_answer($ai_response, $max_tokens);
365	390
366	391	// Extract text content from AI response
367	392	$ai_text = '';
368	393
		@@ -510,12 +535,13 @@
510	535	* don't catch them here they fall through to the "unexpected format" path
511	536	* (or, historically, were serialized into the brief body). All messages
512	537	* start with "The AI " so the outer catch passes them through unchanged.
513	538	*
514		- * @param mixed $ai_response Raw response from the AI client.
	539	+ * @param mixed $ai_response Raw response from the AI client.
	540	+ * @param int $requested_tokens The max_tokens this request asked for; 0 when unknown.
515	541	* @throws \Exception If the response is a refusal, policy block, or truncation.
516	542	*/
517		- private function guard_against_non_answer($ai_response): void {
	543	+ private function guard_against_non_answer($ai_response, int $requested_tokens = 0): void {
518	544	if (!is_array($ai_response)) {
519	545	return;
520	546	}
521	547
		@@ -534,8 +560,20 @@
534	560	)));
535	561	}
536	562	if ('content_filter' === $finish) {
537	563	throw new \Exception('The AI blocked this request under its content policy. Try a different topic or less sensitive keywords.');
	564	+ }
	565	+ // A self-hosted server can stop short of max_tokens because the
	566	+ // prompt and the answer together filled its context window
	567	+ // (Ollama loads models at 4096 by default). A bigger output budget
	568	+ // cannot fix that, so say what can.
	569	+ $completion_tokens = (int) ($ai_response['usage']['completion_tokens'] ?? 0);
	570	+ if ('length' === $finish && $requested_tokens > 0 && $completion_tokens > 0 && $completion_tokens < $requested_tokens) {
	571	+ throw new \Exception(esc_html(sprintf(
	572	+ 'The AI stopped after %1$d tokens, short of the %2$d allowed, because the server ran out of context window before finishing the brief. Raise the context length on your AI server (for Ollama, set OLLAMA_CONTEXT_LENGTH to 16384 or more) and try again.',
	573	+ $completion_tokens,
	574	+ $requested_tokens
	575	+ )));
538	576	}
539	577	if ('length' === $finish) {
540	578	throw new \Exception('The AI stopped at its output token limit before finishing the brief. Try fewer competitor URLs, or a model with a larger output limit. A shorter content length will not help: it asks for a smaller budget, not a smaller answer.');
541	579	}