← All changes | includes/Core/WriteWithAI.php +176 -3 4.8.1 → 4.9.4 View file →

		@@ -16,8 +16,9 @@
16	16	use WPDeveloper\BetterDocs\AI\ProviderFactory;
17	17	use WPDeveloper\BetterDocs\AI\ModelRegistry;
18	18	use WPDeveloper\BetterDocs\Utils\AIUsage;
19	19	use WPDeveloper\BetterDocs\REST\AIEdit;
	20	+ use WPDeveloper\BetterDocs\REST\WriteWithAI as RESTWriteWithAI;
20	21
21	22	class WriteWithAI extends Base {
22	23
23	24	public $settings;
		@@ -64,11 +65,11 @@
64	65	$platform_labels = ModelRegistry::platforms();
65	66	$model_labels = ModelRegistry::models( $active_platform );
66	67
67	68	// Glossary suggestions are existing-terms-only, and the glossaries taxonomy is only
68		- // registered for Pro (see Core\PostType::register_glossaries_taxonomy). So the "Suggest
69		- // glossaries" control must require, on top of the two settings, that Pro is active AND at
70		- // least one glossary term exists — otherwise the modal advertises an offer that can never
	69	+ // registered by Pro (see BetterDocsPro\Core\GlossaryTaxonomy::register_glossaries_taxonomy).
	70	+ // So the "Suggest glossaries" control must require, on top of the two settings, that Pro is
	71	+ // active AND at least one glossary term exists — otherwise the modal advertises an offer that can never
71	72	// return anything. Mirrors the Docs-AI-suite availability check in Core\DocsAISuite.
72	73	$glossary_count = wp_count_terms( array( 'taxonomy' => 'glossaries', 'hide_empty' => false ) );
73	74	$has_glossary_terms = ! is_wp_error( $glossary_count ) && (int) $glossary_count > 0;
74	75	$glossary_suggestions_enabled = (bool) $this->settings->get( 'enable_glossaries', false )
		@@ -100,8 +101,15 @@
100	101	'platform_label' => isset( $platform_labels[ $active_platform ] ) ? $platform_labels[ $active_platform ] : ucfirst( (string) $active_platform ),
101	102	'model' => $active_model,
102	103	'model_label' => isset( $model_labels[ $active_model ] ) ? $model_labels[ $active_model ] : $active_model,
103	104	'max_token' => (int) $this->settings->get( 'ai_autowrite_max_token', 2500 ),
	105	+ // Attachment limits, resolved server-side so the modal can refuse
	106	+ // an oversize file before uploading it — and so it advertises the
	107	+ // host's real ceiling rather than ours when the host is smaller.
	108	+ 'max_upload_bytes' => RESTWriteWithAI::upload_cap( 'file' ),
	109	+ 'max_media_bytes' => RESTWriteWithAI::upload_cap( 'media' ),
	110	+ 'media_exts' => RESTWriteWithAI::media_exts(),
	111	+ 'supports_transcription' => $this->platform_supports( 'transcription', $active_platform ),
104	112	'settings_url' => esc_url( admin_url( 'admin.php?page=betterdocs-settings#betterdocs-ai' ) ),
105	113	'woo_active' => class_exists( 'WooCommerce' ),
106	114	'is_multilingual_active' => Helper::is_multilingual_active(),
107	115	'language_options' => Helper::get_active_languages(),
		@@ -432,8 +440,173 @@
432	440	return $result['content'];
433	441	} catch ( \Exception $error ) {
434	442	return 'Error: ' . $error->getMessage();
435	443	}
	444	+ }
	445	+
	446	+ /**
	447	+ * Generate documentation from an uploaded image using a vision-capable model.
	448	+ *
	449	+ * Mirrors generate_openai_response() but sends the picture alongside the text
	450	+ * prompt as an OpenAI-format multimodal user message
	451	+ * (`content: [ {type:text}, {type:image_url} ]`). The OpenAI-compatible
	452	+ * provider forwards that message array to the wire verbatim, so no provider
	453	+ * change is needed. Only OpenAI vision models are wired — Claude and Gemini
	454	+ * use a different image envelope, so they are refused with a clear error
	455	+ * instead of being sent a payload they would reject.
	456	+ *
	457	+ * @param string $prompt Composed instruction prompt.
	458	+ * @param array $image { data_uri:string, mime:string }.
	459	+ * @param int|null $max_tokens Optional token cap.
	460	+ * @param array $extra_system Extra system messages (instruction sets).
	461	+ * @return string|\WP_Error Generated content, or WP_Error on guard/failure.
	462	+ */
	463	+ public function generate_vision_response( $prompt, $image, $max_tokens = null, $extra_system = array() ) {
	464	+ if ( empty( $image['data_uri'] ) ) {
	465	+ return new \WP_Error( 'ai_vision_no_image', __( 'No image data to send to the AI.', 'betterdocs' ) );
	466	+ }
	467	+
	468	+ $factory = new ProviderFactory( $this->settings );
	469	+ $platform = $factory->active_platform();
	470	+ $model = $factory->active_model( $platform );
	471	+
	472	+ if ( ! $this->platform_supports( 'vision', $platform, $model ) ) {
	473	+ return new \WP_Error(
	474	+ 'ai_no_vision',
	475	+ sprintf(
	476	+ /* translators: 1: AI platform id, 2: model name. */
	477	+ __( 'The configured AI model (%1$s / %2$s) can\'t read images. Switch to an OpenAI vision model such as GPT-4o or GPT-4o mini in BetterDocs → Settings → AI Content Suite, or upload a PDF/DOCX/TXT instead.', 'betterdocs' ),
	478	+ $platform,
	479	+ '' !== (string) $model ? $model : 'default'
	480	+ )
	481	+ );
	482	+ }
	483	+
	484	+ try {
	485	+ $messages = array_merge(
	486	+ array( array( 'role' => 'system', 'content' => $this->get_system_prompt() ) ),
	487	+ $this->normalize_extra_system( $extra_system ),
	488	+ array(
	489	+ array(
	490	+ 'role' => 'user',
	491	+ 'content' => array(
	492	+ array( 'type' => 'text', 'text' => (string) $prompt ),
	493	+ array( 'type' => 'image_url', 'image_url' => array( 'url' => (string) $image['data_uri'] ) ),
	494	+ ),
	495	+ ),
	496	+ )
	497	+ );
	498	+
	499	+ $result = $factory->make()->chat( $messages, $this->ai_chat_options( $max_tokens ) );
	500	+
	501	+ if ( is_wp_error( $result ) ) {
	502	+ return $result;
	503	+ }
	504	+
	505	+ return $result['content'];
	506	+ } catch ( \Exception $error ) {
	507	+ return new \WP_Error( 'ai_vision_failed', 'Error: ' . $error->getMessage() );
	508	+ }
	509	+ }
	510	+
	511	+ /**
	512	+ * Transcribe an uploaded recording to plain text.
	513	+ *
	514	+ * Deliberately not a doc-generation call: this returns the raw transcript so
	515	+ * the author can correct it, and generation then runs through the ordinary
	516	+ * grounded from-source path. That keeps one generation path in the plugin
	517	+ * instead of a second, media-shaped one.
	518	+ *
	519	+ * @since 4.9.4
	520	+ *
	521	+ * @param array $file `[ 'path', 'filename', 'mime' ]` — a PHP temp upload.
	522	+ * @return string|\WP_Error Transcript text.
	523	+ */
	524	+ public function transcribe( $file ) {
	525	+ $factory = new ProviderFactory( $this->settings );
	526	+ $platform = $factory->active_platform();
	527	+
	528	+ if ( ! $this->platform_supports( 'transcription', $platform ) ) {
	529	+ return new \WP_Error(
	530	+ 'ai_no_transcription',
	531	+ sprintf(
	532	+ /* translators: %s: AI platform id, e.g. "claude". */
	533	+ __( 'The configured AI platform (%s) can\'t read audio or video. Switch to OpenAI or Google Gemini in BetterDocs → Settings → AI Content Suite, or upload a text file instead.', 'betterdocs' ),
	534	+ $platform
	535	+ )
	536	+ );
	537	+ }
	538	+
	539	+ // 300s, matching every other AI call here: a 25 MB recording can take a
	540	+ // minute or more to come back, well past the 50s provider default.
	541	+ $options = $this->ai_chat_options();
	542	+ $options['model'] = ModelRegistry::transcription_model( $platform );
	543	+
	544	+ try {
	545	+ return $factory->make()->transcribe( $file, $options );
	546	+ } catch ( \Exception $error ) {
	547	+ return new \WP_Error( 'ai_transcribe_failed', 'Error: ' . $error->getMessage() );
	548	+ }
	549	+ }
	550	+
	551	+ /**
	552	+ * Whether a platform + model can handle a given input capability.
	553	+ *
	554	+ * One entry point for every "can this provider read that file?" question, so
	555	+ * a new capability is a case here rather than another bespoke check beside
	556	+ * the call site.
	557	+ *
	558	+ * @since 4.9.4
	559	+ *
	560	+ * @param string $capability `vision` | `transcription`
	561	+ * @param string $platform
	562	+ * @param string $model Chat model. Ignored for transcription, which uses
	563	+ * its own model (see ModelRegistry).
	564	+ * @return bool
	565	+ */
	566	+ public function platform_supports( $capability, $platform, $model = '' ) {
	567	+ switch ( $capability ) {
	568	+ case 'vision':
	569	+ return $this->platform_supports_vision( $platform, $model );
	570	+
	571	+ case 'transcription':
	572	+ // Presence in the transcription map is the whole test — the chat
	573	+ // model is irrelevant, since transcription runs as its own call
	574	+ // with its own model before any doc is written.
	575	+ return '' !== ModelRegistry::transcription_model( $platform );
	576	+ }
	577	+
	578	+ return false;
	579	+ }
	580	+
	581	+ /**
	582	+ * Whether the active platform + model can accept image input in the OpenAI
	583	+ * multimodal format. Deliberately conservative: only OpenAI vision model
	584	+ * families qualify, because Claude and Gemini require a different image
	585	+ * envelope this path does not build. gpt-3.5 (text-only) is excluded.
	586	+ *
	587	+ * @param string $platform
	588	+ * @param string $model
	589	+ * @return bool
	590	+ */
	591	+ protected function platform_supports_vision( $platform, $model ) {
	592	+ if ( 'openai' !== $platform ) {
	593	+ return false;
	594	+ }
	595	+
	596	+ $model = strtolower( (string) $model );
	597	+
	598	+ if ( '' === $model || false !== strpos( $model, 'gpt-3.5' ) ) {
	599	+ return false;
	600	+ }
	601	+
	602	+ foreach ( array( 'gpt-4o', 'gpt-4.1', 'gpt-4-turbo', 'gpt-4-vision', 'chatgpt-4o', 'gpt-5', 'o1', 'o3', 'o4' ) as $family ) {
	603	+ if ( false !== strpos( $model, $family ) ) {
	604	+ return true;
	605	+ }
	606	+ }
	607	+
	608	+ return false;
436	609	}
437	610
438	611	public function get_outline_system_prompt() {
439	612	$prompt = <<<'PROMPT'