base.php
2 years ago
embed.php
2 years ago
function.php
2 years ago
image.php
2 years ago
parameter.php
2 years ago
text.php
2 years ago
transcribe.php
2 years ago
text.php
470 lines
| 1 | <?php |
| 2 | |
| 3 | class Meow_MWAI_Query_Text extends Meow_MWAI_Query_Base implements JsonSerializable { |
| 4 | public int $maxTokens = 1024; |
| 5 | public float $temperature = 0.8; |
| 6 | public int $maxSentences = 15; |
| 7 | public ?string $stop = null; |
| 8 | public array $messages = []; |
| 9 | public ?string $context = null; |
| 10 | public ?string $newMessage = null; |
| 11 | public ?string $newImage = null; |
| 12 | public ?string $newImageData = null; |
| 13 | public ?string $promptEnding = null; |
| 14 | public bool $casuallyFineTuned = false; |
| 15 | public ?int $promptTokens = null; |
| 16 | |
| 17 | public function __construct( ?string $prompt = '', int $maxTokens = 1024, string $model = MWAI_FALLBACK_MODEL ) { |
| 18 | parent::__construct( $prompt ); |
| 19 | $this->setModel( $model ); |
| 20 | $this->setMaxTokens( $maxTokens ); |
| 21 | } |
| 22 | |
| 23 | #[\ReturnTypeWillChange] |
| 24 | public function jsonSerialize() { |
| 25 | return [ |
| 26 | 'class' => get_class( $this ), |
| 27 | 'prompt' => $this->prompt, |
| 28 | 'messages' => $this->messages, |
| 29 | 'maxTokens' => $this->maxTokens, |
| 30 | 'temperature' => $this->temperature, |
| 31 | 'maxSentences' => $this->maxSentences, |
| 32 | 'context' => $this->context, |
| 33 | 'newMessage' => $this->newMessage, |
| 34 | 'newImage' => $this->newImage, |
| 35 | 'model' => $this->model, |
| 36 | 'mode' => $this->mode, |
| 37 | 'session' => $this->session, |
| 38 | 'env' => $this->env, |
| 39 | 'envId' => $this->envId, |
| 40 | 'service' => $this->service, |
| 41 | 'promptEnding' => $this->promptEnding, |
| 42 | 'stop' => $this->stop, |
| 43 | 'casuallyFineTuned' => $this->casuallyFineTuned, |
| 44 | ]; |
| 45 | } |
| 46 | |
| 47 | public function getPromptTokens( $refresh = false ): int { |
| 48 | if ( $this->promptTokens && !$refresh ) { |
| 49 | return $this->promptTokens; |
| 50 | } |
| 51 | $this->promptTokens = $this->estimateTokens( $this->messages ); |
| 52 | return $this->promptTokens; |
| 53 | } |
| 54 | |
| 55 | public function getLastPrompt(): string { |
| 56 | if ( empty( $this->messages ) ) { |
| 57 | return $this->prompt; |
| 58 | } |
| 59 | $last = $this->getLastMessage(); |
| 60 | return $last; |
| 61 | } |
| 62 | |
| 63 | // Quick and dirty token estimation |
| 64 | // Let's keep this synchronized with Helpers in JS |
| 65 | function estimateTokens( $content ): int |
| 66 | { |
| 67 | $text = ""; |
| 68 | // https://github.com/openai/openai-cookbook/blob/main/examples/How_to_count_tokens_with_tiktoken.ipynb |
| 69 | if ( is_array( $content ) ) { |
| 70 | foreach ( $content as $message ) { |
| 71 | $role = $message['role']; |
| 72 | $content = $message['content']; |
| 73 | $text .= "=#=$role\n$content=#=\n"; |
| 74 | } |
| 75 | } |
| 76 | else { |
| 77 | $text = $content; |
| 78 | } |
| 79 | $tokens = 0; |
| 80 | return apply_filters( 'mwai_estimate_tokens', (int)$tokens, $text, $this->model ); |
| 81 | } |
| 82 | |
| 83 | /** |
| 84 | * Make sure the maxTokens is not greater than the model's context length. |
| 85 | */ |
| 86 | public function finalChecks() { |
| 87 | if ( empty( $this->model ) ) { return; } |
| 88 | |
| 89 | // Make sure the number of messages is not too great |
| 90 | if ( !empty( $this->maxSentences ) ) { |
| 91 | $context = array_shift( $this->messages ); |
| 92 | if ( !empty( $this->messages ) ) { |
| 93 | $this->messages = array_slice( $this->messages, -$this->maxSentences ); |
| 94 | } |
| 95 | else { |
| 96 | $this->messages = []; |
| 97 | } |
| 98 | if ( !empty( $context ) ) { |
| 99 | array_unshift( $this->messages, $context ); |
| 100 | } |
| 101 | |
| 102 | // If the last message is an array, let's look at the message which has a type === 'image_url' |
| 103 | $lastKey = key(array_slice($this->messages, -1, 1, true)); // Get the last key of the messages array |
| 104 | |
| 105 | if ( !empty( $this->newImageData ) ) { |
| 106 | if ( is_array( $this->messages[$lastKey]['content'] ) ) { |
| 107 | foreach ( $this->messages[$lastKey]['content'] as &$message ) { |
| 108 | if ( $message['type'] === 'image_url' ) { |
| 109 | $message['image_url']['url'] = "data:image/jpeg;base64,{$this->newImageData}"; |
| 110 | break; |
| 111 | } |
| 112 | } |
| 113 | unset( $message ); |
| 114 | } |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | // Make sure the max tokens are respected. |
| 119 | $realMax = 4096; |
| 120 | $finetuneFamily = preg_match('/^([a-zA-Z]{0,32}):/', $this->model, $matches ); |
| 121 | $finetuneFamily = ( isset( $matches ) && count( $matches ) > 0 ) ? $matches[1] : 'N/A'; |
| 122 | $foundModel = null; |
| 123 | $openai_models = Meow_MWAI_Engines_OpenAI::get_openai_models(); |
| 124 | foreach ( $openai_models as $currentModel ) { |
| 125 | if ( $currentModel['model'] === $this->model || $currentModel['family'] === $finetuneFamily ) { |
| 126 | $foundModel = $currentModel['name']; |
| 127 | $realMax = $currentModel['maxTokens']; |
| 128 | break; |
| 129 | } |
| 130 | } |
| 131 | $estimatedTokens = $this->getPromptTokens(); |
| 132 | if ( !empty( $realMax ) && $estimatedTokens > $realMax ) { |
| 133 | throw new Exception( "AI Engine: The prompt is too long! It contains about $estimatedTokens tokens (estimation). The $foundModel model only accepts a maximum of $realMax tokens. " ); |
| 134 | } |
| 135 | $realMax = (int)($realMax - $estimatedTokens) - 16; |
| 136 | if ( $this->maxTokens > $realMax ) { |
| 137 | $this->maxTokens = $realMax; |
| 138 | } |
| 139 | } |
| 140 | |
| 141 | /** |
| 142 | * ID of the model to use. |
| 143 | * @param string $model ID of the model to use. |
| 144 | */ |
| 145 | public function setModel( string $model ) { |
| 146 | $this->model = $model; |
| 147 | $this->mode = 'completion'; |
| 148 | $found = false; |
| 149 | $openai_models = Meow_MWAI_Engines_OpenAI::get_openai_models(); |
| 150 | foreach ( $openai_models as $currentModel ) { |
| 151 | if ( $currentModel['model'] === $this->model ) { |
| 152 | if ( $currentModel['mode'] ) { |
| 153 | $this->mode = $currentModel['mode']; |
| 154 | } |
| 155 | $found = true; |
| 156 | break; |
| 157 | } |
| 158 | } |
| 159 | if ( !$found ) { |
| 160 | // If the model can't be found, it's because it's probably a fine-tuned model. In the past (before August 2023), |
| 161 | // fine-tuned models were always based on GPT-3 (and therefore, using completion mode). From now on, they can be |
| 162 | // based on GPT-3.5 or 4 (and therefore, using chat mode). We need to detect that. |
| 163 | $baseModel = Meow_MWAI_Engines_OpenAI::get_finetune_base_model( $model ); |
| 164 | if ( preg_match( '/^gpt-3.5|^gpt-4/', $baseModel ) ) { |
| 165 | $this->mode = 'chat'; |
| 166 | } |
| 167 | } |
| 168 | } |
| 169 | |
| 170 | /** |
| 171 | * Given a prompt, the model will return one or more predicted completions. |
| 172 | * It can also return the probabilities of alternative tokens at each position. |
| 173 | * @param string $prompt The prompt to generate completions. |
| 174 | */ |
| 175 | public function setPrompt( $prompt ) { |
| 176 | parent::setPrompt( $prompt ); |
| 177 | $this->validateMessages(); |
| 178 | } |
| 179 | |
| 180 | /** |
| 181 | * The prompt is used by models who uses Text Completion (and not Chat Completion). |
| 182 | * This returns the prompt if it's not a chat, otherwise it will build a prompt with |
| 183 | * all the messages nicely formatted. |
| 184 | */ |
| 185 | public function getPrompt(): ?string { |
| 186 | // In the case it's really just a prompt. |
| 187 | if ( count( $this->messages ) === 1 ) { |
| 188 | $first = reset( $this->messages ); |
| 189 | return $first['content']; |
| 190 | } |
| 191 | |
| 192 | // In the case it's a chat that we need to convert into a prompt. |
| 193 | $first = reset( $this->messages ); |
| 194 | $prompt = ""; |
| 195 | if ( $first && $first['role'] === 'system' ) { |
| 196 | $prompt = $first['content'] . "\n\n"; |
| 197 | } |
| 198 | |
| 199 | // Casually Fine-Tuned or Prompt-Ending |
| 200 | if ( !empty( $this->promptEnding ) ) { |
| 201 | $last = end( $this->messages ); |
| 202 | if ( $last && $last['role'] === 'user' ) { |
| 203 | $prompt = $last['content'] . $this->promptEnding; |
| 204 | } |
| 205 | return $prompt; |
| 206 | } |
| 207 | |
| 208 | // Standard Completion |
| 209 | while ( $message = next( $this->messages ) ) { |
| 210 | $role = $message['role']; |
| 211 | $content = $message['content']; |
| 212 | if ( $role === 'system' ) { |
| 213 | $prompt .= "$content\n\n"; |
| 214 | } |
| 215 | if ( $role === 'user' ) { |
| 216 | $prompt .= "User: $content\n"; |
| 217 | } |
| 218 | if ( $role === 'assistant' ) { |
| 219 | $prompt .= "AI: $content\n"; |
| 220 | } |
| 221 | } |
| 222 | $prompt .= "AI: "; |
| 223 | return $prompt; |
| 224 | } |
| 225 | |
| 226 | /** |
| 227 | * Similar to the prompt, but focus on the new/last message. |
| 228 | * Only used when the model has a chat mode (and only used in messages). |
| 229 | * @param string $prompt The messages to generate completions. |
| 230 | */ |
| 231 | public function setNewMessage( string $newMessage ): void { |
| 232 | $this->newMessage = $newMessage; |
| 233 | $this->validateMessages(); |
| 234 | } |
| 235 | |
| 236 | public function setNewImage( string $newImage ): void { |
| 237 | $this->newImage = $newImage; |
| 238 | $this->validateMessages(); |
| 239 | } |
| 240 | |
| 241 | public function setNewImageData( string $newImageData ): void { |
| 242 | $this->newImageData = $newImageData; |
| 243 | $this->validateMessages(); |
| 244 | } |
| 245 | |
| 246 | public function replace( $search, $replace ) { |
| 247 | $this->prompt = str_replace( $search, $replace, $this->prompt ); |
| 248 | $this->validateMessages(); |
| 249 | } |
| 250 | |
| 251 | /** |
| 252 | * Similar to the prompt, but use an array of messages instead. |
| 253 | * @param string $prompt The messages to generate completions. |
| 254 | */ |
| 255 | public function setMessages( array $messages ) { |
| 256 | $messages = array_map( function( $message ) { |
| 257 | if ( is_array( $message ) ) { |
| 258 | return [ 'role' => $message['role'], 'content' => $message['content'] ]; |
| 259 | } |
| 260 | else if ( is_object( $message ) ) { |
| 261 | return [ 'role' => $message->role, 'content' => $message->content ]; |
| 262 | } |
| 263 | else { |
| 264 | throw new InvalidArgumentException( 'Unsupported message type.' ); |
| 265 | } |
| 266 | }, $messages ); |
| 267 | $this->messages = $messages; |
| 268 | $this->validateMessages(); |
| 269 | } |
| 270 | |
| 271 | public function getLastMessage() { |
| 272 | if ( !empty( $this->messages ) ) { |
| 273 | $lastMessageIndex = count( $this->messages ) - 1; |
| 274 | $lastMessage = $this->messages[$lastMessageIndex]; |
| 275 | if ( is_array( $lastMessage['content'] ) ) { |
| 276 | foreach( $lastMessage['content'] as $message ) { |
| 277 | if ( $message['type'] === 'text' ) { |
| 278 | return $message['text']; |
| 279 | } |
| 280 | } |
| 281 | } |
| 282 | else { |
| 283 | return $lastMessage['content']; |
| 284 | } |
| 285 | } |
| 286 | return null; |
| 287 | } |
| 288 | |
| 289 | // Function that adds a message just before the last message |
| 290 | public function injectContext( string $content ): void { |
| 291 | if ( !empty( $this->messages ) ) { |
| 292 | $lastMessageIndex = count( $this->messages ) - 1; |
| 293 | $lastMessage = $this->messages[$lastMessageIndex]; |
| 294 | $this->messages[$lastMessageIndex] = [ 'role' => 'system', 'content' => $content ]; |
| 295 | array_push( $this->messages, $lastMessage ); |
| 296 | } |
| 297 | $this->validateMessages(); |
| 298 | } |
| 299 | |
| 300 | /** |
| 301 | * The context that is used for the chat completion (mode === 'chat'). |
| 302 | * @param string $context The context to use. |
| 303 | */ |
| 304 | public function setContext( string $context ): void { |
| 305 | $this->context = apply_filters( 'mwai_ai_context', $context, $this ); |
| 306 | $this->validateMessages(); |
| 307 | } |
| 308 | |
| 309 | private function validateMessages(): void { |
| 310 | // Messages should end with either the prompt or, if exists, the newMessage. |
| 311 | $message = empty( $this->newMessage ) ? $this->prompt : $this->newMessage; |
| 312 | $content = $message; |
| 313 | |
| 314 | // If there is an image, we need to adapt it to Vision. |
| 315 | if ( !empty( $this->newImage ) ) { |
| 316 | $content = [ |
| 317 | [ "type" => "text", "text" => $message ], |
| 318 | [ "type" => "image_url", "image_url" => [ "url" => $this->newImage ] ] |
| 319 | ]; |
| 320 | } |
| 321 | |
| 322 | if ( empty( $this->messages ) ) { |
| 323 | $this->messages = [ [ 'role' => 'user', 'content' => $content ] ]; |
| 324 | } |
| 325 | else { |
| 326 | $last = &$this->messages[ count( $this->messages ) - 1 ]; |
| 327 | if ( $last['role'] === 'user' ) { |
| 328 | $last['content'] = $content; |
| 329 | } |
| 330 | else { |
| 331 | array_push( $this->messages, [ 'role' => 'user', 'content' => $content ] ); |
| 332 | } |
| 333 | } |
| 334 | |
| 335 | // The main context must be first. |
| 336 | if ( !empty( $this->context ) ) { |
| 337 | if ( is_array( $this->messages ) && count( $this->messages ) > 0 ) { |
| 338 | if ( $this->messages[0]['role'] !== 'system' ) { |
| 339 | array_unshift( $this->messages, [ 'role' => 'system', 'content' => $this->context ] ); |
| 340 | } |
| 341 | else { |
| 342 | $this->messages[0]['content'] = $this->context; |
| 343 | } |
| 344 | } |
| 345 | } |
| 346 | } |
| 347 | |
| 348 | /** |
| 349 | * The maximum number of tokens to generate in the completion. |
| 350 | * The token count of your prompt plus max_tokens cannot exceed the model's context length. |
| 351 | * Most models have a context length of 2048 tokens (except for the newest models, which support 4096). |
| 352 | * @param float $prompt The maximum number of tokens. |
| 353 | */ |
| 354 | public function setMaxTokens( int $maxTokens ): void { |
| 355 | $this->maxTokens = $maxTokens; |
| 356 | } |
| 357 | |
| 358 | /** |
| 359 | * Set the sampling temperature to use. Higher values means the model will take more risks. |
| 360 | * Try 0.9 for more creative applications, and 0 for ones with a well-defined reply. |
| 361 | * @param float $temperature The temperature. |
| 362 | */ |
| 363 | public function setTemperature( float $temperature ): void { |
| 364 | $temperature = floatval( $temperature ); |
| 365 | if ( $temperature > 1 ) { |
| 366 | $temperature = 1; |
| 367 | } |
| 368 | if ( $temperature < 0 ) { |
| 369 | $temperature = 0; |
| 370 | } |
| 371 | $this->temperature = round( $temperature, 2 ); |
| 372 | } |
| 373 | |
| 374 | public function setMaxSentences( int $maxSentences ): void { |
| 375 | if ( !empty( $maxSentences ) ) { |
| 376 | $this->maxSentences = intval( $maxSentences ); |
| 377 | $this->validateMessages(); |
| 378 | } |
| 379 | } |
| 380 | |
| 381 | public function setStop( string $stop ): void { |
| 382 | $this->stop = $stop; |
| 383 | } |
| 384 | |
| 385 | private function convertKeys( $params ) |
| 386 | { |
| 387 | $newParams = []; |
| 388 | foreach ( $params as $key => $value ) { |
| 389 | $newKey = ''; |
| 390 | $capitalizeNextChar = false; |
| 391 | for ( $i = 0; $i < strlen( $key ); $i++ ) { |
| 392 | if ( $key[$i] == '_' ) { |
| 393 | $capitalizeNextChar = true; |
| 394 | } |
| 395 | else { |
| 396 | $newKey .= $capitalizeNextChar ? strtoupper($key[$i]) : $key[$i]; |
| 397 | $capitalizeNextChar = false; |
| 398 | } |
| 399 | } |
| 400 | $newParams[$newKey] = $value; |
| 401 | } |
| 402 | return $newParams; |
| 403 | } |
| 404 | |
| 405 | // Based on the params of the query, update the attributes |
| 406 | public function injectParams( array $params ): void |
| 407 | { |
| 408 | // Those are for the keys passed directly by the shortcode. |
| 409 | $params = $this->convertKeys( $params ); |
| 410 | |
| 411 | $acceptedValues = [ true, 1, '1' ]; |
| 412 | if ( !empty( $params['model'] ) ) { |
| 413 | $this->setModel( $params['model'] ); |
| 414 | } |
| 415 | if ( !empty( $params['casuallyFineTuned'] ) && in_array( $params['casuallyFineTuned'], $acceptedValues, true ) ) { |
| 416 | $this->promptEnding = "\n\n###\n\n"; |
| 417 | $this->stop = "\n\n"; |
| 418 | $this->casuallyFineTuned = true; |
| 419 | } |
| 420 | if ( !empty( $params['prompt'] ) ) { |
| 421 | $this->setPrompt( $params['prompt'] ); |
| 422 | } |
| 423 | if ( !empty( $params['context'] ) ) { |
| 424 | $this->setContext( $params['context'] ); |
| 425 | } |
| 426 | if ( !empty( $params['messages'] ) ) { |
| 427 | $this->setMessages( $params['messages'] ); |
| 428 | } |
| 429 | if ( !empty( $params['newMessage'] ) ) { |
| 430 | $this->setNewMessage( $params['newMessage'] ); |
| 431 | } |
| 432 | if ( !empty( $params['maxTokens'] ) && intval( $params['maxTokens'] ) > 0 ) { |
| 433 | $this->setMaxTokens( intval( $params['maxTokens'] ) ); |
| 434 | } |
| 435 | if ( !empty( $params['maxMessages'] ) && intval( $params['maxMessages'] ) > 0 ) { |
| 436 | $this->setMaxSentences( intval( $params['maxMessages'] ) ); |
| 437 | } |
| 438 | if ( !empty( $params['maxSentences'] ) && intval( $params['maxSentences'] ) > 0 ) { |
| 439 | $this->setMaxSentences( intval( $params['maxSentences'] ) ); |
| 440 | } |
| 441 | if ( !empty( $params['temperature'] ) ) { |
| 442 | $this->setTemperature( $params['temperature'] ); |
| 443 | } |
| 444 | if ( !empty( $params['stop'] ) ) { |
| 445 | $this->setStop( $params['stop'] ); |
| 446 | } |
| 447 | if ( !empty( $params['maxResults'] ) ) { |
| 448 | $this->setMaxResults( $params['maxResults'] ); |
| 449 | } |
| 450 | if ( !empty( $params['env'] ) ) { |
| 451 | $this->setEnv( $params['env'] ); |
| 452 | } |
| 453 | if ( !empty( $params['session'] ) ) { |
| 454 | $this->setSession( $params['session'] ); |
| 455 | } |
| 456 | // Should add the params related to Open AI and Azure |
| 457 | if ( !empty( $params['service'] ) ) { |
| 458 | $this->setService( $params['service'] ); |
| 459 | } |
| 460 | if ( !empty( $params['apiKey'] ) ) { |
| 461 | $this->setApiKey( $params['apiKey'] ); |
| 462 | } |
| 463 | if ( !empty( $params['botId'] ) ) { |
| 464 | $this->setBotId( $params['botId'] ); |
| 465 | } |
| 466 | if ( !empty( $params['envId'] ) ) { |
| 467 | $this->setEnvId( $params['envId'] ); |
| 468 | } |
| 469 | } |
| 470 | } |