anthropic.php
5 hours ago
chatml.php
1 week ago
core.php
2 days ago
custom.php
1 month ago
factory.php
1 week ago
google-interactions.php
2 days ago
google.php
1 week ago
mistral.php
1 week ago
open-router.php
3 weeks ago
openai.php
3 weeks ago
ovh.php
1 week ago
perplexity.php
6 months ago
replicate.php
5 months ago
xai.php
1 month ago
google-interactions.php
741 lines
| 1 | <?php |
| 2 | |
| 3 | /** |
| 4 | * Google Gemini — Interactions API engine. |
| 5 | * |
| 6 | * As of June 2026 the Interactions API is GA and Google's recommended interface |
| 7 | * (https://ai.google.dev/gemini-api/docs/interactions-overview). It is stateful |
| 8 | * (server-side history via previous_interaction_id), exposes Google's built-in |
| 9 | * tools (Google Search, Google Maps, URL context, code execution), and returns a |
| 10 | * chronological list of execution "steps" (model_output / function_call / |
| 11 | * function_result / thoughts). |
| 12 | * |
| 13 | * This is the DEFAULT engine for Google environments. The classic |
| 14 | * generateContent engine (Meow_MWAI_Engines_Google) remains the fallback, |
| 15 | * selected when "Use Standard API" (google_use_standard_api) is enabled in |
| 16 | * Settings > AI > General. The request/response field names are mapped from |
| 17 | * Google's docs and confirmed against the live API. |
| 18 | */ |
| 19 | class Meow_MWAI_Engines_GoogleInteractions extends Meow_MWAI_Engines_Core { |
| 20 | protected $apiKey = null; |
| 21 | protected $endpoint = null; |
| 22 | |
| 23 | // Streaming accumulators (collected from SSE events, used to build the Reply). |
| 24 | protected $streamId = null; |
| 25 | protected $streamInTokens = null; |
| 26 | protected $streamOutTokens = null; |
| 27 | protected $streamToolCalls = []; |
| 28 | protected $streamImages = []; |
| 29 | |
| 30 | // The classic Google engine, used for everything the Interactions API does not |
| 31 | // cover (embeddings, image generation, transcription) and for model fetching. |
| 32 | protected $classicEngine = null; |
| 33 | |
| 34 | public function __construct( $core, $env ) { |
| 35 | parent::__construct( $core, $env ); |
| 36 | $this->set_environment(); |
| 37 | } |
| 38 | |
| 39 | protected function set_environment() { |
| 40 | $env = $this->env; |
| 41 | $this->apiKey = $env['apikey']; |
| 42 | $this->endpoint = apply_filters( |
| 43 | 'mwai_google_endpoint', |
| 44 | 'https://generativelanguage.googleapis.com/v1beta', |
| 45 | $this->env |
| 46 | ); |
| 47 | } |
| 48 | |
| 49 | protected function build_headers( $query = null ) { |
| 50 | // A query may carry a per-request API key override (set via inject_params); |
| 51 | // fall back to the environment key otherwise. |
| 52 | $apiKey = ( $query !== null && !empty( $query->apiKey ) ) ? $query->apiKey : $this->apiKey; |
| 53 | return [ |
| 54 | 'Content-Type' => 'application/json', |
| 55 | 'x-goog-api-key' => $apiKey, |
| 56 | ]; |
| 57 | } |
| 58 | |
| 59 | #region Request building |
| 60 | |
| 61 | /** |
| 62 | * Build the "input" for the request. |
| 63 | * - Feedback queries carry function results back as function_result items. |
| 64 | * - When we have a previous interaction id, the server holds the history so we |
| 65 | * only send the new user turn. |
| 66 | * - Otherwise we send the conversation as Content[]. |
| 67 | */ |
| 68 | protected function user_step( $query ) { |
| 69 | $content = []; |
| 70 | |
| 71 | // Attached images / files (vision). |
| 72 | $attachments = method_exists( $query, 'getAttachments' ) ? $query->getAttachments() : []; |
| 73 | foreach ( $attachments as $file ) { |
| 74 | $mime = $file->get_mimeType() ? $file->get_mimeType() : 'image/jpeg'; |
| 75 | // Image content parts are flat: { type, data (base64), mime_type }. |
| 76 | $content[] = [ |
| 77 | 'type' => 'image', |
| 78 | 'data' => $file->get_base64(), |
| 79 | 'mime_type' => $mime, |
| 80 | ]; |
| 81 | } |
| 82 | |
| 83 | if ( $query->message !== '' || empty( $content ) ) { |
| 84 | $content[] = [ 'type' => 'text', 'text' => $query->message ]; |
| 85 | } |
| 86 | |
| 87 | return [ 'type' => 'user_input', 'content' => $content ]; |
| 88 | } |
| 89 | |
| 90 | /** |
| 91 | * The Interactions API is steps-based: input must be a step_list, not a |
| 92 | * role-based turn_list. So we emit user_input / model_output / function_result |
| 93 | * steps (the same step types the API returns). |
| 94 | */ |
| 95 | protected function build_input( $query ) { |
| 96 | // Feedback: carry the function results back as function_result steps. |
| 97 | if ( $query instanceof Meow_MWAI_Query_Feedback && !empty( $query->blocks ) ) { |
| 98 | $steps = []; |
| 99 | foreach ( $query->blocks as $block ) { |
| 100 | foreach ( ( $block['feedbacks'] ?? [] ) as $feedback ) { |
| 101 | $value = $feedback['reply']['value'] ?? ''; |
| 102 | if ( !is_string( $value ) ) { |
| 103 | $value = wp_json_encode( $value ); |
| 104 | } |
| 105 | // The Interactions API expects `call_id` (matching the function_call |
| 106 | // step's id) and `result` as an array of content blocks, not a string. |
| 107 | $steps[] = [ |
| 108 | 'type' => 'function_result', |
| 109 | 'call_id' => $feedback['request']['toolId'] ?? null, |
| 110 | 'name' => $feedback['request']['name'] ?? '', |
| 111 | 'result' => [ [ 'type' => 'text', 'text' => $value ] ], |
| 112 | ]; |
| 113 | } |
| 114 | } |
| 115 | return $steps; |
| 116 | } |
| 117 | |
| 118 | $steps = []; |
| 119 | |
| 120 | // When continuing an interaction, prior turns live server-side (referenced by |
| 121 | // previous_interaction_id) so we only send the new user turn. On the FIRST |
| 122 | // turn the server has no state, so we must replay the conversation history |
| 123 | // ($query->messages) as user_input / model_output steps. Without this, |
| 124 | // restored discussions and any caller-supplied history are lost. The API |
| 125 | // accepts replayed model_output steps (verified live). |
| 126 | if ( !$this->has_previous_interaction( $query ) && !empty( $query->messages ) ) { |
| 127 | foreach ( $query->messages as $message ) { |
| 128 | $text = $message['content'] ?? ''; |
| 129 | if ( !is_string( $text ) || $text === '' ) { |
| 130 | continue; |
| 131 | } |
| 132 | $role = $message['role'] ?? 'user'; |
| 133 | $type = ( $role === 'assistant' || $role === 'model' ) ? 'model_output' : 'user_input'; |
| 134 | $steps[] = [ 'type' => $type, 'content' => [ [ 'type' => 'text', 'text' => $text ] ] ]; |
| 135 | } |
| 136 | } |
| 137 | |
| 138 | $steps[] = $this->user_step( $query ); |
| 139 | return $steps; |
| 140 | } |
| 141 | |
| 142 | protected function build_tools( $query ) { |
| 143 | $tools = []; |
| 144 | |
| 145 | // Custom WordPress / Code Engine functions -> function declarations. |
| 146 | // Each tool is flat: { type:function, name, description, parameters }. |
| 147 | if ( !empty( $query->functions ) ) { |
| 148 | foreach ( $query->functions as $function ) { |
| 149 | $decl = $function->serializeForOpenAI(); |
| 150 | // The Interactions API validates `parameters` as a JSON Schema and is |
| 151 | // strict about it: a no-arg function's empty |
| 152 | // { type:object, properties:{}, required:[] } is rejected with the |
| 153 | // misleading "value at top-level must be a list". Send a bare object |
| 154 | // schema when there are no parameters, and drop an empty `required` |
| 155 | // list otherwise. |
| 156 | if ( empty( $function->parameters ) ) { |
| 157 | $decl['parameters'] = [ 'type' => 'object' ]; |
| 158 | } |
| 159 | elseif ( empty( $decl['parameters']['required'] ) ) { |
| 160 | unset( $decl['parameters']['required'] ); |
| 161 | } |
| 162 | $tools[] = array_merge( [ 'type' => 'function' ], $decl ); |
| 163 | } |
| 164 | } |
| 165 | |
| 166 | // Provider built-in tools. AI Engine uses the generic 'web_search' name |
| 167 | // across providers; for Gemini it maps to Google Search grounding (the |
| 168 | // classic generateContent engine maps it the same way). Google Maps |
| 169 | // grounding is named-places only for now (no lat/long context), so "near me" |
| 170 | // queries won't resolve but explicit places do. |
| 171 | // |
| 172 | // Google Search and Google Maps are mutually exclusive in one Interactions |
| 173 | // request (the API rejects "cannot be combined in the same request"). When a |
| 174 | // chatbot has both enabled, prefer Maps (the more specific opt-in tool). |
| 175 | if ( in_array( 'google_maps', $query->tools, true ) ) { |
| 176 | $tools[] = [ 'type' => 'google_maps' ]; |
| 177 | } |
| 178 | elseif ( in_array( 'web_search', $query->tools, true ) ) { |
| 179 | $tools[] = [ 'type' => 'google_search' ]; |
| 180 | } |
| 181 | |
| 182 | return $tools; |
| 183 | } |
| 184 | |
| 185 | protected function build_interaction_body( $query, $streamCallback = null ) { |
| 186 | $body = [ |
| 187 | 'model' => $query->model, |
| 188 | 'input' => $this->build_input( $query ), |
| 189 | 'store' => true, |
| 190 | ]; |
| 191 | |
| 192 | if ( !is_null( $streamCallback ) ) { |
| 193 | $body['stream'] = true; |
| 194 | } |
| 195 | |
| 196 | if ( $this->has_previous_interaction( $query ) ) { |
| 197 | $body['previous_interaction_id'] = $query->previousResponseId; |
| 198 | } |
| 199 | |
| 200 | // System instructions, plus any per-query context (RAG / content-aware / |
| 201 | // Smart Search). Context is re-retrieved each turn and is not part of the |
| 202 | // server-side history, so it must be sent on every request, not just the |
| 203 | // first one. Without this, embeddings and content-aware context are silently |
| 204 | // dropped on Gemini. |
| 205 | $instructions = !empty( $query->instructions ) ? $query->instructions : ''; |
| 206 | if ( !empty( $query->context ) ) { |
| 207 | $framedContext = $this->core->frame_context( $query->context ); |
| 208 | $instructions = trim( $instructions . "\n\n" . $framedContext ); |
| 209 | } |
| 210 | if ( $instructions !== '' ) { |
| 211 | $body['system_instruction'] = $instructions; |
| 212 | } |
| 213 | |
| 214 | $tools = $this->build_tools( $query ); |
| 215 | if ( !empty( $tools ) ) { |
| 216 | $body['tools'] = $tools; |
| 217 | } |
| 218 | |
| 219 | $generationConfig = []; |
| 220 | if ( $query->temperature !== null ) { |
| 221 | $generationConfig['temperature'] = $query->temperature; |
| 222 | } |
| 223 | if ( !empty( $query->maxTokens ) ) { |
| 224 | $generationConfig['max_output_tokens'] = $query->maxTokens; |
| 225 | } |
| 226 | if ( !empty( $generationConfig ) ) { |
| 227 | $body['generation_config'] = $generationConfig; |
| 228 | } |
| 229 | |
| 230 | // Image-capable Gemini models (Flash Image) only return images when the |
| 231 | // request opts into image output via the top-level response_format; without |
| 232 | // it they answer in text ("I can't output an image in this chat"). The |
| 233 | // dedicated image-query path is unaffected (it runs on the classic engine). |
| 234 | if ( $this->model_supports_image_generation( $query->model ) ) { |
| 235 | $responseFormat = [ 'type' => 'image' ]; |
| 236 | if ( !empty( $query->resolution ) ) { |
| 237 | $responseFormat['aspect_ratio'] = $query->resolution; |
| 238 | } |
| 239 | $body['response_format'] = $responseFormat; |
| 240 | } |
| 241 | |
| 242 | return apply_filters( 'mwai_google_interactions_body', $body, $query ); |
| 243 | } |
| 244 | |
| 245 | /** |
| 246 | * Whether the model can emit images in a chat turn (the 'image-generation' |
| 247 | * feature, e.g. Gemini Flash Image). Used to opt the request into image |
| 248 | * output via response_format. |
| 249 | */ |
| 250 | protected function model_supports_image_generation( $modelId ) { |
| 251 | $model = $this->core->find_model_data( $modelId, $this->env['id'] ?? null ); |
| 252 | return !empty( $model['features'] ) |
| 253 | && in_array( 'image-generation', $model['features'], true ); |
| 254 | } |
| 255 | |
| 256 | #endregion |
| 257 | |
| 258 | public function run_completion_query( $query, $streamCallback = null ): Meow_MWAI_Reply { |
| 259 | // Non-image attachments (PDFs, documents) need the Gemini Files API upload |
| 260 | // handled by the classic engine's prepare_query() / build_messages() (Pro). |
| 261 | // The Interactions input format only carries inline image parts, so delegate |
| 262 | // those queries to the classic engine to preserve file handling. |
| 263 | if ( $this->has_non_image_attachment( $query ) ) { |
| 264 | return $this->classic_engine()->run_completion_query( $query, $streamCallback ); |
| 265 | } |
| 266 | |
| 267 | $debug = $this->core->get_option( 'queries_debug_mode' ); |
| 268 | $isStreaming = !is_null( $streamCallback ); |
| 269 | $this->init_debug_mode( $query ); |
| 270 | |
| 271 | // Reset per-request streaming state (Core buffers + our accumulators). |
| 272 | $this->streamContent = ''; |
| 273 | $this->streamTemporaryBuffer = ''; |
| 274 | $this->streamBuffer = ''; |
| 275 | $this->reset_stream(); |
| 276 | |
| 277 | if ( $isStreaming ) { |
| 278 | // The Core stream_handler hooks the curl WRITEFUNCTION and dispatches each |
| 279 | // SSE "data:" line to our stream_data_handler(). |
| 280 | $this->streamCallback = $streamCallback; |
| 281 | add_action( 'http_api_curl', [ $this, 'stream_handler' ], 10, 3 ); |
| 282 | } |
| 283 | |
| 284 | $body = $this->build_interaction_body( $query, $streamCallback ); |
| 285 | $url = trailingslashit( $this->endpoint ) . 'interactions'; |
| 286 | |
| 287 | if ( $debug ) { |
| 288 | error_log( '[AI Engine Queries] --> (Gemini Interactions) ' . $url ); |
| 289 | error_log( wp_json_encode( $body, JSON_PRETTY_PRINT | JSON_UNESCAPED_SLASHES ) ); |
| 290 | } |
| 291 | |
| 292 | $args = [ |
| 293 | 'headers' => $this->build_headers( $query ), |
| 294 | 'body' => wp_json_encode( $body ), |
| 295 | 'timeout' => apply_filters( 'mwai_request_timeout', 120, $query ), |
| 296 | 'sslverify' => MWAI_SSL_VERIFY, |
| 297 | ]; |
| 298 | $tmpFile = null; |
| 299 | if ( $isStreaming ) { |
| 300 | // WordPress streams the response into this file; on error it holds the body. |
| 301 | $tmpFile = tempnam( sys_get_temp_dir(), 'mwai-gemini-stream-' ); |
| 302 | $args['stream'] = true; |
| 303 | $args['filename'] = $tmpFile; |
| 304 | } |
| 305 | |
| 306 | try { |
| 307 | $res = wp_remote_post( $url, $args ); |
| 308 | |
| 309 | if ( is_wp_error( $res ) ) { |
| 310 | throw new Exception( 'AI Engine (Gemini Interactions): ' . $res->get_error_message() ); |
| 311 | } |
| 312 | |
| 313 | $code = wp_remote_retrieve_response_code( $res ); |
| 314 | |
| 315 | if ( $isStreaming ) { |
| 316 | if ( $code < 200 || $code >= 300 ) { |
| 317 | // The error body went to the temp file rather than the response. |
| 318 | $errBody = ( $tmpFile && file_exists( $tmpFile ) ) |
| 319 | ? file_get_contents( $tmpFile ) : wp_remote_retrieve_body( $res ); |
| 320 | $errData = json_decode( $errBody, true ); |
| 321 | $detail = $errData['error']['message'] ?? $errBody; |
| 322 | throw new Exception( 'AI Engine (Gemini Interactions) HTTP ' . $code . ': ' . $detail ); |
| 323 | } |
| 324 | return $this->build_streaming_reply( $query ); |
| 325 | } |
| 326 | |
| 327 | $rawBody = wp_remote_retrieve_body( $res ); |
| 328 | $data = json_decode( $rawBody, true ); |
| 329 | |
| 330 | if ( $debug ) { |
| 331 | error_log( '[AI Engine Queries] <-- (Gemini Interactions) HTTP ' . $code ); |
| 332 | error_log( substr( $rawBody, 0, 4000 ) ); |
| 333 | } |
| 334 | |
| 335 | if ( $code < 200 || $code >= 300 ) { |
| 336 | $detail = $data['error']['message'] ?? $rawBody; |
| 337 | throw new Exception( 'AI Engine (Gemini Interactions) HTTP ' . $code . ': ' . $detail ); |
| 338 | } |
| 339 | |
| 340 | return $this->build_reply_from_interaction( $query, $data, $streamCallback ); |
| 341 | } |
| 342 | finally { |
| 343 | if ( $isStreaming ) { |
| 344 | remove_action( 'http_api_curl', [ $this, 'stream_handler' ] ); |
| 345 | } |
| 346 | if ( $tmpFile && file_exists( $tmpFile ) ) { |
| 347 | unlink( $tmpFile ); |
| 348 | } |
| 349 | } |
| 350 | } |
| 351 | |
| 352 | public function reset_stream() { |
| 353 | $this->streamId = null; |
| 354 | $this->streamInTokens = null; |
| 355 | $this->streamOutTokens = null; |
| 356 | $this->streamToolCalls = []; |
| 357 | $this->streamImages = []; |
| 358 | } |
| 359 | |
| 360 | /** |
| 361 | * Handle one parsed SSE "data:" payload from the Interactions stream. Returns a |
| 362 | * text string for text deltas (the Core accumulates it into streamContent and |
| 363 | * forwards to the chatbot), a Meow_MWAI_Event for thoughts, null for |
| 364 | * bookkeeping events (function-call assembly, completion), or throws on a |
| 365 | * provider error event. |
| 366 | * |
| 367 | * The Core's stream_handler only forwards "data:" lines and drops the "event:" |
| 368 | * line, so we dispatch on the event_type carried inside the JSON payload. |
| 369 | */ |
| 370 | protected function stream_data_handler( $json ) { |
| 371 | $type = $json['event_type'] ?? ''; |
| 372 | |
| 373 | if ( isset( $json['error'] ) || $type === 'error' || $type === 'interaction.failed' ) { |
| 374 | $this->throw_stream_error( $json ); |
| 375 | } |
| 376 | |
| 377 | if ( $type === 'step.start' ) { |
| 378 | $step = $json['step'] ?? []; |
| 379 | // A function call arrives as a single step.start with the complete call |
| 380 | // (id, name, arguments). Larger argument payloads may also stream as |
| 381 | // arguments_delta events, which we accumulate below. |
| 382 | if ( ( $step['type'] ?? '' ) === 'function_call' ) { |
| 383 | $index = $json['index'] ?? count( $this->streamToolCalls ); |
| 384 | $args = $step['arguments'] ?? ''; |
| 385 | $argsJson = is_string( $args ) ? $args : wp_json_encode( $args ); |
| 386 | $this->streamToolCalls[$index] = [ |
| 387 | 'id' => $step['id'] ?? null, |
| 388 | 'name' => $step['name'] ?? '', |
| 389 | 'arguments' => $argsJson, |
| 390 | ]; |
| 391 | // Surface the call in the event-log (gated like the function_result |
| 392 | // event Core emits, so nothing streams when Event Logs is off). |
| 393 | if ( $this->currentDebugMode && $this->streamCallback ) { |
| 394 | return Meow_MWAI_Event::function_calling( $step['name'] ?? '', $argsJson ); |
| 395 | } |
| 396 | } |
| 397 | return null; |
| 398 | } |
| 399 | |
| 400 | if ( $type === 'step.delta' ) { |
| 401 | $delta = $json['delta'] ?? []; |
| 402 | $dtype = $delta['type'] ?? ''; |
| 403 | if ( $dtype === 'text' ) { |
| 404 | return $delta['text'] ?? ''; |
| 405 | } |
| 406 | if ( $dtype === 'arguments_delta' ) { |
| 407 | $index = $json['index'] ?? array_key_last( $this->streamToolCalls ); |
| 408 | if ( $index !== null && isset( $this->streamToolCalls[$index] ) ) { |
| 409 | $this->streamToolCalls[$index]['arguments'] .= $delta['arguments'] ?? ''; |
| 410 | } |
| 411 | return null; |
| 412 | } |
| 413 | // Generated image (e.g. Gemini Flash Image): the delta carries mime_type + |
| 414 | // base64 data and has no "type" field. The base64 may stream in chunks, so |
| 415 | // accumulate per step index. |
| 416 | if ( isset( $delta['mime_type'], $delta['data'] ) ) { |
| 417 | $index = $json['index'] ?? count( $this->streamImages ); |
| 418 | $isFirst = !isset( $this->streamImages[$index] ); |
| 419 | if ( $isFirst ) { |
| 420 | $this->streamImages[$index] = [ 'mime_type' => $delta['mime_type'], 'data' => '' ]; |
| 421 | } |
| 422 | $this->streamImages[$index]['data'] .= $delta['data']; |
| 423 | if ( $isFirst && $this->currentDebugMode && $this->streamCallback ) { |
| 424 | $event = new Meow_MWAI_Event( 'live', MWAI_STREAM_TYPES['IMAGE_GEN'] ); |
| 425 | $event->set_content( 'Image generated' ); |
| 426 | return $event; |
| 427 | } |
| 428 | return null; |
| 429 | } |
| 430 | // Thought deltas carry an encrypted thought_signature, not readable text, |
| 431 | // so there is nothing to surface to the user; they are intentionally skipped. |
| 432 | return null; |
| 433 | } |
| 434 | |
| 435 | if ( $type === 'interaction.completed' ) { |
| 436 | $interaction = $json['interaction'] ?? []; |
| 437 | if ( !empty( $interaction['id'] ) ) { |
| 438 | $this->streamId = $interaction['id']; |
| 439 | } |
| 440 | $usage = $interaction['usage'] ?? []; |
| 441 | if ( isset( $usage['total_input_tokens'] ) ) { |
| 442 | $this->streamInTokens = (int) $usage['total_input_tokens']; |
| 443 | } |
| 444 | if ( isset( $usage['total_output_tokens'] ) ) { |
| 445 | $this->streamOutTokens = (int) $usage['total_output_tokens']; |
| 446 | } |
| 447 | return null; |
| 448 | } |
| 449 | |
| 450 | return null; |
| 451 | } |
| 452 | |
| 453 | protected function throw_stream_error( $json ) { |
| 454 | $error = $json['error'] ?? ( $json['interaction']['error'] ?? null ); |
| 455 | $message = null; |
| 456 | $code = null; |
| 457 | $status = null; |
| 458 | |
| 459 | if ( is_array( $error ) ) { |
| 460 | $message = $error['message'] ?? ( $error['reason'] ?? null ); |
| 461 | $code = $error['code'] ?? null; |
| 462 | $status = $error['status'] ?? ( $error['type'] ?? null ); |
| 463 | } |
| 464 | elseif ( is_string( $error ) ) { |
| 465 | $message = $error; |
| 466 | } |
| 467 | |
| 468 | if ( empty( $message ) ) { |
| 469 | $message = $json['message'] ?? ( $json['error_message'] ?? 'Unknown stream error.' ); |
| 470 | } |
| 471 | |
| 472 | $detail = $message; |
| 473 | if ( !empty( $code ) ) { |
| 474 | $detail .= ' (' . $code . ')'; |
| 475 | } |
| 476 | if ( !empty( $status ) ) { |
| 477 | $detail .= ' (' . $status . ')'; |
| 478 | } |
| 479 | |
| 480 | throw new Exception( 'AI Engine (Gemini Interactions) stream error: ' . $detail ); |
| 481 | } |
| 482 | |
| 483 | /** |
| 484 | * Build the Reply from the streaming accumulators once the SSE stream is done. |
| 485 | * Mirrors build_reply_from_interaction() but sources from the streamed state. |
| 486 | */ |
| 487 | protected function build_streaming_reply( $query ) { |
| 488 | $reply = new Meow_MWAI_Reply( $query ); |
| 489 | $choices = []; |
| 490 | |
| 491 | foreach ( $this->streamToolCalls as $tc ) { |
| 492 | $args = $tc['arguments']; |
| 493 | $decoded = ( $args === '' || $args === null ) ? [] : json_decode( $args, true ); |
| 494 | if ( json_last_error() !== JSON_ERROR_NONE ) { |
| 495 | $decoded = []; |
| 496 | } |
| 497 | $choices[] = [ |
| 498 | 'message' => [ |
| 499 | 'content' => null, |
| 500 | 'tool_calls' => [ [ |
| 501 | 'id' => $tc['id'], |
| 502 | 'type' => 'function', |
| 503 | 'function' => [ 'name' => $tc['name'], 'arguments' => $decoded ], |
| 504 | ] ], |
| 505 | ], |
| 506 | ]; |
| 507 | } |
| 508 | |
| 509 | if ( $this->streamContent !== '' ) { |
| 510 | $choices[] = [ 'role' => 'assistant', 'text' => $this->streamContent ]; |
| 511 | } |
| 512 | |
| 513 | // Generated images become b64_json choices; set_choices() saves each one and |
| 514 | // appends the image markdown to the reply so the chatbot renders it. |
| 515 | foreach ( $this->streamImages as $img ) { |
| 516 | if ( !empty( $img['data'] ) ) { |
| 517 | $choices[] = [ 'b64_json' => $img['data'] ]; |
| 518 | } |
| 519 | } |
| 520 | |
| 521 | // Nothing to show (no text, image, or function call): surface a short |
| 522 | // fallback instead of a blank bubble. This happens e.g. when a text-only |
| 523 | // model is asked to edit an image, or the model declines a request. |
| 524 | if ( empty( $choices ) ) { |
| 525 | $choices[] = [ 'role' => 'assistant', 'text' => $this->empty_reply_fallback( $query ) ]; |
| 526 | } |
| 527 | |
| 528 | $reply->set_choices( $choices ); |
| 529 | |
| 530 | if ( !empty( $this->streamId ) ) { |
| 531 | $reply->set_id( $this->streamId ); |
| 532 | } |
| 533 | |
| 534 | if ( $this->streamInTokens !== null || $this->streamOutTokens !== null ) { |
| 535 | $recorded = $this->core->record_tokens_usage( |
| 536 | $query->model, |
| 537 | (int) ( $this->streamInTokens ?? 0 ), |
| 538 | (int) ( $this->streamOutTokens ?? 0 ) |
| 539 | ); |
| 540 | $reply->set_usage( $recorded ); |
| 541 | } |
| 542 | |
| 543 | return $reply; |
| 544 | } |
| 545 | |
| 546 | /** |
| 547 | * Message shown when the model returns nothing renderable (no text, image, or |
| 548 | * function call), so the chatbot never displays a blank bubble. Filterable. |
| 549 | */ |
| 550 | protected function empty_reply_fallback( $query ) { |
| 551 | return apply_filters( |
| 552 | 'mwai_google_interactions_empty_reply', |
| 553 | __( "Sorry, I couldn't produce a response for that. Please try rephrasing your request.", 'ai-engine' ), |
| 554 | $query |
| 555 | ); |
| 556 | } |
| 557 | |
| 558 | /** |
| 559 | * Turn an Interaction resource (id, status, steps[], usage) into a Reply. |
| 560 | * Emits step events (thoughts, function calls) to the stream callback so the |
| 561 | * chatbot's event-log UI shows what the model did. |
| 562 | */ |
| 563 | protected function build_reply_from_interaction( $query, $data, $streamCallback = null ) { |
| 564 | $reply = new Meow_MWAI_Reply( $query ); |
| 565 | |
| 566 | $choices = []; |
| 567 | $text = ''; |
| 568 | foreach ( ( $data['steps'] ?? [] ) as $step ) { |
| 569 | $type = $step['type'] ?? ''; |
| 570 | |
| 571 | if ( $type === 'model_output' ) { |
| 572 | foreach ( ( $step['content'] ?? [] ) as $part ) { |
| 573 | if ( ( $part['type'] ?? '' ) === 'text' ) { |
| 574 | $text .= $part['text'] ?? ''; |
| 575 | } |
| 576 | // Generated image part (mime_type + base64 data) -> b64_json choice. |
| 577 | elseif ( isset( $part['mime_type'], $part['data'] ) ) { |
| 578 | $choices[] = [ 'b64_json' => $part['data'] ]; |
| 579 | } |
| 580 | } |
| 581 | } |
| 582 | elseif ( $type === 'thought' ) { |
| 583 | if ( !empty( $streamCallback ) ) { |
| 584 | $thought = ''; |
| 585 | foreach ( ( $step['content'] ?? [] ) as $part ) { |
| 586 | $thought .= $part['text'] ?? ''; |
| 587 | } |
| 588 | if ( $thought !== '' ) { |
| 589 | $event = new Meow_MWAI_Event( 'live', 'thinking' ); |
| 590 | $event->set_content( $thought ); |
| 591 | call_user_func( $streamCallback, $event ); |
| 592 | } |
| 593 | } |
| 594 | } |
| 595 | elseif ( $type === 'function_call' ) { |
| 596 | $name = $step['name'] ?? ''; |
| 597 | $args = $step['arguments'] ?? []; |
| 598 | if ( !empty( $streamCallback ) ) { |
| 599 | $event = Meow_MWAI_Event::function_calling( $name, is_string( $args ) ? $args : wp_json_encode( $args ) ); |
| 600 | call_user_func( $streamCallback, $event ); |
| 601 | } |
| 602 | // Map to a tool_calls choice so set_choices() builds the needFeedback, |
| 603 | // preserving the id (required for the function_result follow-up). |
| 604 | $choices[] = [ |
| 605 | 'message' => [ |
| 606 | 'content' => null, |
| 607 | 'tool_calls' => [ [ |
| 608 | 'id' => $step['id'] ?? null, |
| 609 | 'type' => 'function', |
| 610 | 'function' => [ 'name' => $name, 'arguments' => $args ], |
| 611 | ] ], |
| 612 | ], |
| 613 | '_rawMessage' => $step, |
| 614 | ]; |
| 615 | } |
| 616 | } |
| 617 | |
| 618 | if ( $text !== '' ) { |
| 619 | $choices[] = [ 'role' => 'assistant', 'text' => $text ]; |
| 620 | } |
| 621 | |
| 622 | // Surface a short fallback rather than a blank reply when there is nothing |
| 623 | // to show (no text, image, or function call). |
| 624 | if ( empty( $choices ) ) { |
| 625 | $choices[] = [ 'role' => 'assistant', 'text' => $this->empty_reply_fallback( $query ) ]; |
| 626 | } |
| 627 | |
| 628 | $reply->set_choices( $choices ); |
| 629 | |
| 630 | // Stateful handle for the next turn. |
| 631 | if ( !empty( $data['id'] ) ) { |
| 632 | $reply->set_id( $data['id'] ); |
| 633 | } |
| 634 | |
| 635 | // Usage. |
| 636 | $usage = $data['usage'] ?? []; |
| 637 | if ( isset( $usage['total_input_tokens'] ) || isset( $usage['total_output_tokens'] ) ) { |
| 638 | $recorded = $this->core->record_tokens_usage( |
| 639 | $query->model, |
| 640 | (int) ( $usage['total_input_tokens'] ?? 0 ), |
| 641 | (int) ( $usage['total_output_tokens'] ?? 0 ) |
| 642 | ); |
| 643 | $reply->set_usage( $recorded ); |
| 644 | } |
| 645 | |
| 646 | return $reply; |
| 647 | } |
| 648 | |
| 649 | /** |
| 650 | * Interaction ids look like "v1_Chd..." (the classic generateContent API is |
| 651 | * stateless and has no id). Used to decide whether to reuse server-side state. |
| 652 | */ |
| 653 | protected function has_previous_interaction( $query ) { |
| 654 | return !empty( $query->previousResponseId ) |
| 655 | && is_string( $query->previousResponseId ) |
| 656 | && strpos( $query->previousResponseId, 'v1_' ) === 0; |
| 657 | } |
| 658 | |
| 659 | /** |
| 660 | * Dynamic model list is shared with the classic Google engine (same /models |
| 661 | * endpoint and same API key). get_models() is also called during final_checks() |
| 662 | * to validate the requested model, so it must be implemented. |
| 663 | */ |
| 664 | public function get_models() { |
| 665 | return $this->core->get_engine_models( 'google' ); |
| 666 | } |
| 667 | |
| 668 | /** |
| 669 | * Called by the framework to price a reply. The classic engine already prices |
| 670 | * every Google model and query type correctly (text, token-based Flash Image, |
| 671 | * and per-image Imagen), so reuse it rather than keep a second partial copy. |
| 672 | * Interaction usage carries prompt_tokens/completion_tokens/total_tokens, which |
| 673 | * is exactly what the classic pricing expects. |
| 674 | */ |
| 675 | public function get_price( Meow_MWAI_Query_Base $query, Meow_MWAI_Reply $reply ) { |
| 676 | return $this->classic_engine()->get_price( $query, $reply ); |
| 677 | } |
| 678 | |
| 679 | /** |
| 680 | * The Interactions API only covers chat completions. Everything else still |
| 681 | * runs against the classic Google endpoints (same API key and env), so we lazily |
| 682 | * build the classic engine and delegate those operations to it. |
| 683 | */ |
| 684 | protected function classic_engine() { |
| 685 | if ( $this->classicEngine === null ) { |
| 686 | $this->classicEngine = Meow_MWAI_Engines_Google::create( $this->core, $this->env ); |
| 687 | } |
| 688 | return $this->classicEngine; |
| 689 | } |
| 690 | |
| 691 | /** |
| 692 | * Whether the query carries an attachment that is not an inline image (e.g. a |
| 693 | * PDF or document). Those need the classic engine's Files API handling. |
| 694 | */ |
| 695 | protected function has_non_image_attachment( $query ) { |
| 696 | $attachments = method_exists( $query, 'getAttachments' ) ? $query->getAttachments() : []; |
| 697 | foreach ( $attachments as $file ) { |
| 698 | $mime = $file->get_mimeType(); |
| 699 | if ( !empty( $mime ) && strpos( $mime, 'image/' ) !== 0 ) { |
| 700 | return true; |
| 701 | } |
| 702 | } |
| 703 | return false; |
| 704 | } |
| 705 | |
| 706 | /** |
| 707 | * The base engine's connection_check() throws "not implemented"; the test runs |
| 708 | * through the factory which now returns this engine for Google envs. The check |
| 709 | * (same /models endpoint + key) is identical to the classic engine, so delegate. |
| 710 | */ |
| 711 | public function connection_check() { |
| 712 | return $this->classic_engine()->connection_check(); |
| 713 | } |
| 714 | |
| 715 | /** |
| 716 | * Fetching + tagging the model list (the "Refresh Models" action) is identical |
| 717 | * to the classic Google engine: same /models endpoint, same API key, same tag |
| 718 | * derivation (vision, functions, web_search...). Delegate to it rather than |
| 719 | * duplicate that logic, so refreshed Gemini models carry the right tools. |
| 720 | */ |
| 721 | public function retrieve_models() { |
| 722 | return $this->classic_engine()->retrieve_models(); |
| 723 | } |
| 724 | |
| 725 | public function run_embedding_query( Meow_MWAI_Query_Base $query ) { |
| 726 | return $this->classic_engine()->run_embedding_query( $query ); |
| 727 | } |
| 728 | |
| 729 | public function run_image_query( Meow_MWAI_Query_Base $query, $streamCallback = null ) { |
| 730 | return $this->classic_engine()->run_image_query( $query, $streamCallback ); |
| 731 | } |
| 732 | |
| 733 | public function run_editimage_query( Meow_MWAI_Query_Base $query ) { |
| 734 | return $this->classic_engine()->run_editimage_query( $query ); |
| 735 | } |
| 736 | |
| 737 | public function run_transcribe_query( Meow_MWAI_Query_Base $query ) { |
| 738 | return $this->classic_engine()->run_transcribe_query( $query ); |
| 739 | } |
| 740 | } |
| 741 |