← All changes | includes/ai/class-vision-client.php +130 -15 2.1.1 → 2.10.0 View file →

		@@ -76,9 +76,9 @@
76	76	public static function capable_providers(): array {
77	77	return [
78	78	'openai' => 'gpt-5-mini',
79	79	'claude' => 'claude-sonnet-5',
80		- 'gemini' => 'gemini-2.5-flash',
	80	+ 'gemini' => Settings::DEFAULT_GEMINI_MODEL,
81	81	];
82	82	}
83	83
84	84	/**
		@@ -86,15 +86,39 @@
86	86	*
87	87	* @return bool
88	88	*/
89	89	public function is_available(): bool {
90		- $provider = (string) $this->settings->get('ai_provider', 'openai');
	90	+ $provider = (string) $this->settings->get('ai_provider', Settings::AI_PROVIDER_NONE);
91	91
	92	+ if ('openai_compatible' === $provider) {
	93	+ return $this->compatible_vision_ready();
	94	+ }
	95	+
92	96	return isset(self::capable_providers()[$provider])
93	97	&& '' !== $this->api_key_for($provider);
94	98	}
95	99
96	100	/**
	101	+ * Can the user's own OpenAI-compatible endpoint describe an image?
	102	+ *
	103	+ * Only the administrator knows: an Ollama box running llama3.1 cannot,
	104	+ * the same box running llava can, and there is no reliable way to ask the
	105	+ * server. So it is a declared capability (`openai_compatible_supports_images`,
	106	+ * default off) rather than a guess — sending a vision payload to a
	107	+ * text-only model returns a confusing 400, and AI alt text stays hidden
	108	+ * until the user says the model handles images (#721).
	109	+ *
	110	+ * @since 2.8.0
	111	+ *
	112	+ * @return bool
	113	+ */
	114	+ private function compatible_vision_ready(): bool {
	115	+ return (bool) $this->settings->get('openai_compatible_supports_images', false)
	116	+ && '' !== (string) $this->settings->get('openai_compatible_base_url', '')
	117	+ && '' !== trim((string) $this->settings->get('openai_compatible_model', ''));
	118	+ }
	119	+
	120	+ /**
97	121	* The stored key for a provider.
98	122	*
99	123	* Keys are held per provider (`openai_api_key`, `claude_api_key`,
100	124	* `gemini_api_key`) — the same names AI_Manager reads. There is no
		@@ -117,15 +141,46 @@
117	141	* @return string Alt text, or '' when it could not be produced.
118	142	* @throws \Exception When the provider is unusable or the call fails.
119	143	*/
120	144	public function describe_attachment(int $attachment_id, string $context = ''): string {
121		- $provider = (string) $this->settings->get('ai_provider', 'openai');
	145	+ $provider = (string) $this->settings->get('ai_provider', Settings::AI_PROVIDER_NONE);
122	146	$models = self::capable_providers();
123	147
	148	+ // Distinguish "no provider chosen" from "this provider can't do vision":
	149	+ // interpolating an empty provider name reads as a broken string (#572).
	150	+ if (Settings::AI_PROVIDER_NONE === $provider) {
	151	+ throw new \Exception(esc_html__(
	152	+ 'No AI provider is selected. Choose OpenAI, Anthropic or Gemini in Settings → AI Provider.',
	153	+ 'thinkrank'
	154	+ ));
	155	+ }
	156	+
	157	+ if ('openai_compatible' === $provider) {
	158	+ if (!$this->compatible_vision_ready()) {
	159	+ throw new \Exception(esc_html__(
	160	+ 'Your OpenAI-compatible endpoint is not set up for images. Set a base URL and model id, and turn on "This model can describe images" in Settings → AI Provider.',
	161	+ 'thinkrank'
	162	+ ));
	163	+ }
	164	+
	165	+ $image = $this->read_image($attachment_id);
	166	+ if (null === $image) {
	167	+ throw new \Exception(esc_html__('The image file could not be read, or is larger than the provider allows.', 'thinkrank'));
	168	+ }
	169	+
	170	+ return $this->clean($this->call_openai(
	171	+ (string) $this->settings->get('openai_compatible_api_key', ''),
	172	+ trim((string) $this->settings->get('openai_compatible_model', '')),
	173	+ $this->build_prompt($context),
	174	+ $image,
	175	+ (string) $this->settings->get('openai_compatible_base_url', '')
	176	+ ));
	177	+ }
	178	+
124	179	if (!isset($models[$provider])) {
125	180	throw new \Exception(sprintf(
126	181	/* translators: %s: AI provider name. */
127		- esc_html__('The %s provider cannot describe images. Switch to OpenAI, Claude or Gemini in Settings → AI Provider.', 'thinkrank'),
	182	+ esc_html__('The %s provider cannot describe images. Switch to OpenAI, Anthropic or Gemini in Settings → AI Provider.', 'thinkrank'),
128	183	esc_html($provider)
129	184	));
130	185	}
131	186
		@@ -248,13 +303,14 @@
248	303	*
249	304	* @param string $api_key API key.
250	305	* @param string $model Model id.
251	306	* @param string $prompt Instruction.
252		- * @param array $image ['data' => base64, 'mime' => string].
	307	+ * @param array $image ['data' => base64, 'mime' => string].
	308	+ * @param string $base_url Base URL, for an OpenAI-compatible endpoint that is not OpenAI's.
253	309	* @return string
254	310	* @throws \Exception On API failure.
255	311	*/
256		- private function call_openai(string $api_key, string $model, string $prompt, array $image): string {
	312	+ private function call_openai(string $api_key, string $model, string $prompt, array $image, string $base_url = OpenAI_Client::API_BASE_URL): string {
257	313	$body = [
258	314	'model' => $model,
259	315	'messages' => [[
260	316	'role' => 'user',
		@@ -274,13 +330,36 @@
274	330	if (0 === strpos($model, 'gpt-5')) {
275	331	$body['reasoning_effort'] = 'minimal';
276	332	}
277	333
278		- $response = $this->post('https://api.openai.com/v1/chat/completions', [
279		- 'Authorization' => 'Bearer ' . $api_key,
280		- 'Content-Type' => 'application/json',
281		- ], $body);
	334	+ $base_url = rtrim(trim($base_url), '/');
	335	+ if ('' === $base_url) {
	336	+ $base_url = OpenAI_Client::API_BASE_URL;
	337	+ }
282	338
	339	+ // Ollama, LM Studio and vLLM reject max_completion_tokens — it is a
	340	+ // parameter OpenAI added for its reasoning models, not part of the
	341	+ // Chat Completions shape they implement — so a custom endpoint gets
	342	+ // plain max_tokens.
	343	+ if (OpenAI_Client::API_BASE_URL !== $base_url) {
	344	+ unset($body['max_completion_tokens'], $body['reasoning_effort']);
	345	+ $body['max_tokens'] = 300;
	346	+ }
	347	+
	348	+ $headers = ['Content-Type' => 'application/json'];
	349	+ // A local server usually wants no key at all.
	350	+ if ('' !== $api_key) {
	351	+ $headers['Authorization'] = 'Bearer ' . $api_key;
	352	+ $headers['api-key'] = $api_key;
	353	+ }
	354	+
	355	+ $response = $this->post(
	356	+ Endpoint_URL_Validator::route($base_url, 'chat/completions'),
	357	+ $headers,
	358	+ $body,
	359	+ OpenAI_Client::API_BASE_URL !== $base_url
	360	+ );
	361	+
283	362	return (string) ($response['choices'][0]['message']['content'] ?? '');
284	363	}
285	364
286	365	/**
		@@ -357,20 +436,37 @@
357	436	*
358	437	* @param string $url Endpoint.
359	438	* @param array $headers Headers.
360	439	* @param array $body Payload.
	440	+ * @param bool $guarded Whether this is a user-named endpoint, which has its
	441	+ * destination resolved, pinned and its body capped.
361	442	* @return array Decoded response.
362	443	* @throws \Exception On transport or API error.
363	444	*/
364		- private function post(string $url, array $headers, array $body): array {
365		- $response = wp_remote_post($url, [
	445	+ private function post(string $url, array $headers, array $body, bool $guarded = false): array {
	446	+ // The user's daily ceiling and kill switch are enforced here, at the
	447	+ // one place every outbound vision call passes through, so no feature
	448	+ // path can bypass them by forgetting to ask first (#448).
	449	+ Spend_Guard::guard();
	450	+ Spend_Guard::record();
	451	+
	452	+ $args = [
366	453	// Vision calls carry a payload and think for a moment; the 30s
367		- // default is too tight for a large image on a slow link.
368		- 'timeout' => 60,
	454	+ // default is too tight for a large image on a slow link. A local
	455	+ // vision model is slower still, so honour the user's own timeout
	456	+ // when they configured one (#721).
	457	+ 'timeout' => $this->request_timeout(),
369	458	'headers' => $headers,
370	459	'body' => wp_json_encode($body),
371		- ]);
	460	+ // The key rides in a header; never let a redirect hand it to
	461	+ // another host.
	462	+ 'redirection' => 0,
	463	+ ];
372	464
	465	+ $response = $guarded
	466	+ ? Endpoint_URL_Validator::guarded_request($url, $args + ['method' => 'POST'])
	467	+ : wp_remote_post($url, $args);
	468	+
373	469	if (is_wp_error($response)) {
374	470	throw new \Exception(esc_html($response->get_error_message()));
375	471	}
376	472
		@@ -383,8 +479,27 @@
383	479	throw new \Exception(esc_html(sprintf('Vision API error (%d): %s', $status, (string) $message)));
384	480	}
385	481
386	482	return is_array($data) ? $data : [];
	483	+ }
	484	+
	485	+ /**
	486	+ * Timeout for a vision request.
	487	+ *
	488	+ * 60s for the hosted providers. A local vision model on CPU is slower than
	489	+ * anything hosted, so an OpenAI-compatible endpoint gets the timeout the
	490	+ * administrator configured for it, floored at 60 (#721).
	491	+ *
	492	+ * @since 2.8.0
	493	+ *
	494	+ * @return int Seconds.
	495	+ */
	496	+ private function request_timeout(): int {
	497	+ if ('openai_compatible' !== (string) $this->settings->get('ai_provider', Settings::AI_PROVIDER_NONE)) {
	498	+ return 60;
	499	+ }
	500	+
	501	+ return max(60, (int) $this->settings->get('openai_compatible_timeout', Settings::DEFAULT_OPENAI_COMPATIBLE_TIMEOUT));
387	502	}
388	503
389	504	/**
390	505	* Normalize a model's reply into usable alt text.