Current Path : /storage/v11800/affypharma/public_html/wp-content/plugins/ai-engine/classes/engines/ |
Linux v11800 5.3.0-1023-aws #25~18.04.1-Ubuntu SMP Fri Jun 5 15:19:18 UTC 2020 aarch64
|
Current File : /storage/v11800/affypharma/public_html/wp-content/plugins/ai-engine/classes/engines/core.php |
<?php
class Meow_MWAI_Engines_Core {
protected $core = null;
public $env = null;
public $envId = null;
public $envType = null;
// Streaming
protected $streamCallback = null;
protected $streamTemporaryBuffer = "";
protected $streamBuffer = "";
protected $streamHeaders = [];
protected $streamContent = "";
public function __construct( $core, $env ) {
$this->core = $core;
$this->env = $env;
$this->envId = $env['id'];
$this->envType = $env['type'];
}
public function run( $query, $streamCallback = null, $maxDepth = 5 ) {
// Check if the query is allowed.
$limits = $this->core->get_option( 'limits' );
$allowed = apply_filters( 'mwai_ai_allowed', true, $query, $limits );
if ( $allowed !== true ) {
$message = is_string( $allowed ) ? $allowed : 'Unauthorized query.';
throw new Exception( $message );
}
// Allow to modify the query before it is sent.
// Embedding and Feedback queries are not allowed to be modified.
if ( !( $query instanceof Meow_MWAI_Query_Embed ) && !( $query instanceof Meow_MWAI_Query_Feedback ) ) {
$query = apply_filters( 'mwai_ai_query', $query );
}
// Important as it makes sure everything is consolidated in the query and the engine.
$this->final_checks( $query );
// Run the query
$reply = null;
if ( $query instanceof Meow_MWAI_Query_Text || $query instanceof Meow_MWAI_Query_Feedback ) {
$reply = $this->run_completion_query( $query, $streamCallback );
}
else if ( $query instanceof Meow_MWAI_Query_Assistant || $query instanceof Meow_MWAI_Query_AssistFeedback ) {
$reply = $this->run_assistant_query( $query, $streamCallback );
if ( $reply === null ) {
throw new Exception( 'Assistants are not supported in this version of AI Engine.' );
}
}
else if ( $query instanceof Meow_MWAI_Query_Embed ) {
$reply = $this->run_embedding_query( $query );
}
else if ( $query instanceof Meow_MWAI_Query_Image ) {
$reply = $this->run_images_query( $query );
}
else if ( $query instanceof Meow_MWAI_Query_Transcribe ) {
$reply = $this->run_transcribe_query( $query );
}
else {
throw new Exception( 'Unknown query type.' );
}
// Allow to modify the reply before it is sent.
$reply = apply_filters( 'mwai_ai_reply', $reply, $query );
// Function Call
if ( !empty( $reply->needFeedbacks ) ) {
// Check if we reached the maximum depth
if ( $maxDepth <= 0 ) {
throw new Exception( 'AI Engine: There seems to be a loop in the function/tools calls.' );
}
// We should use a feedback query around the original query
if ( !( $query instanceof Meow_MWAI_Query_AssistFeedback) && !( $query instanceof Meow_MWAI_Query_Feedback ) ) {
$queryClass = $query instanceof Meow_MWAI_Query_Assistant ?
Meow_MWAI_Query_AssistFeedback::class : Meow_MWAI_Query_Feedback::class;
$query = new $queryClass( $reply, $reply->query );
}
// The engine for the model will handle the feedback query nicely
// In the case of Anthropic, it's like a "discussion" between the user (= WordPress's AI Engine
// and the model (= Anthropic). The feedback is a way to communicate between the two.
$feedback_blocks = [];
foreach ( $reply->needFeedbacks as $needFeedback ) {
$rawMessageKey = md5( serialize( $needFeedback['rawMessage'] ) );
// Initialize the feedback block for this rawMessage if it hasn't been initialized yet
if ( !isset( $feedback_blocks[$rawMessageKey] ) ) {
$feedback_blocks[$rawMessageKey] = [
'rawMessage' => $needFeedback['rawMessage'],
'feedbacks' => []
];
}
// Get the value related to this feedback (usually, a function call)
$value = apply_filters( 'mwai_ai_feedback', null, $needFeedback );
// Add the feedback information to the appropriate feedback block
$feedback_blocks[$rawMessageKey]['feedbacks'][] = [
'request' => $needFeedback, // TODO: Meow_MWAI_Feedback_Request
'reply' => [ 'value' => $value ] // TODO: Meow_MWAI_Feedback_Reply
];
}
foreach ( $feedback_blocks as $feedback_block ) {
$query->add_feedback_block( $feedback_block );
}
// Run the feedback query
$reply = $this->run( $query, $streamCallback, $maxDepth - 1 );
}
return $reply;
}
public function retrieve_model_info( $model ) {
$models = $this->get_models();
foreach ( $models as $currentModel ) {
if ( $currentModel['model'] === $model ) {
return $currentModel;
}
}
return false;
}
public function final_checks( Meow_MWAI_Query_Base $query ) {
$query->final_checks();
//$found = false;
// Check if the model is available, except if it's an assistant
if ( !( $query instanceof Meow_MWAI_Query_Assistant ) ) {
// TODO: Avoid checking on the finetuned models for now.
if ( substr( $query->model, 0, 3 ) === 'ft:' ) {
return;
}
$model_info = $this->retrieve_model_info( $query->model );
if ( $model_info === false ) {
throw new Exception( "AI Engine: The model '{$query->model}' is not available." );
}
if ( isset( $model_info['mode'] ) ) {
$query->mode = $model_info['mode'];
}
}
}
// Streamline the messages:
// - Concatenate consecutive model messages into a single message for the model role
// - Make sure the first message is a user message
// - Make sure the last message is a user message
protected function streamline_messages( $messages, $systemRole = 'assistant', $messageType = 'content' )
{
$processedMessages = [];
$lastRole = '';
$concatenatedText = '';
// Determine the way to access message content based on messageType
$getContent = function( $message ) use ( $messageType ) {
if ( $messageType == 'parts' ) {
return $message['parts'][0]['text'];
}
else { // Default to 'content'
return $message['content'];
}
};
// Set content to a message depending on the messageType
$setContent = function( &$message, $content ) use ( $messageType ) {
if ( $messageType == 'parts' ) {
$message['parts'] = [['text' => $content]];
}
else { // Default to 'content'
$message['content'] = $content;
}
};
// Concatenate consecutive model messages into a single message for the model role
foreach ( $messages as $message ) {
if ( $message['role'] == $systemRole ) {
if ( $lastRole == $systemRole ) {
$concatenatedText .= "\n" . $getContent( $message );
}
else {
if ( $concatenatedText !== '' ) {
$newMessage = [ 'role' => $systemRole ];
$setContent( $newMessage, $concatenatedText );
$processedMessages[] = $newMessage;
}
$concatenatedText = $getContent( $message );
}
}
else {
if ( $lastRole == $systemRole ) {
$newMessage = [ 'role' => $systemRole ];
$setContent( $newMessage, $concatenatedText );
$processedMessages[] = $newMessage;
$concatenatedText = '';
}
$processedMessages[] = $message;
}
$lastRole = $message['role'];
}
if ( $lastRole == $systemRole && $concatenatedText !== '' ) {
$newMessage = [ 'role' => $systemRole ];
$setContent( $newMessage, $concatenatedText );
$processedMessages[] = $newMessage;
}
// Make sure the last message is a user message, if not, throw an exception
if ( end( $processedMessages )['role'] !== 'user' ) {
throw new Exception( 'The last message must be a user message.' );
}
// Make sure the first message is a user message, if not, add an empty user message
if ( $processedMessages[0]['role'] !== 'user' ) {
$newMessage = [ 'role' => 'user' ];
$setContent( $newMessage, '' );
array_unshift( $processedMessages, $newMessage );
}
return $processedMessages;
}
// Check for a JSON-formatted error in the data, and throw an exception if it's the case.
function stream_error_check( $data ) {
if ( strpos( $data, 'error' ) === false ) {
return;
}
$data = trim( $data );
$jsonPart = $data;
if ( strpos( $jsonPart, 'data:' ) === 0 ) {
$jsonPart = trim( substr( $jsonPart, strlen( 'data:' ) ) );
}
$json = json_decode( $jsonPart, true );
if ( json_last_error() === JSON_ERROR_NONE ) {
if ( isset( $json['error'] ) ) {
$error = $json['error'];
$code = null;
$type = null;
$message = null;
if ( isset( $error['message'] ) ) {
$message = $error['message'];
}
else if ( is_string( $error ) ) {
throw new Exception( "Error: $error" );
}
else {
throw new Exception( "Unknown error (stream_error_check)." );
}
if ( isset( $error['code'] ) ) {
$code = $error['code'];
}
if ( isset( $error['type'] ) ) {
$type = $error['type'];
}
$errorMessage = "Error: $message";
if ( !is_null( $code ) ) {
$errorMessage .= " ($code)";
}
if ( !is_null( $type ) ) {
$errorMessage .= " ($type)";
}
throw new Exception( $errorMessage );
}
else if ( isset( $json['type'] ) && $json['type'] === 'error' ) {
$type = $json['error']['type'];
$message = $json['error']['message'];
throw new Exception( "Error: $message ($type)" );
}
}
}
public function stream_handler( $handle, $args, $url ) {
curl_setopt( $handle, CURLOPT_SSL_VERIFYPEER, false );
curl_setopt( $handle, CURLOPT_SSL_VERIFYHOST, false );
// TODO: This is breaking the response. We need to find a way to handle the headers.
// curl_setopt( $handle, CURLOPT_HEADERFUNCTION, function ( $curl, $header ) {
// $length = strlen( $header );
// $this->streamHeaders[] = $header;
// $this->stream_header_handler( $header );
// return $length;
// });
curl_setopt( $handle, CURLOPT_WRITEFUNCTION, function ( $curl, $data ) {
$length = strlen( $data );
// Bufferize the unfinished stream (if it's the case)
$this->streamTemporaryBuffer .= $data;
$this->streamBuffer .= $data;
// Error Management
$this->stream_error_check( $this->streamBuffer );
$lines = explode( "\n", $this->streamTemporaryBuffer );
if ( substr( $this->streamTemporaryBuffer, -1 ) !== "\n" ) {
$this->streamTemporaryBuffer = array_pop( $lines );
}
else {
$this->streamTemporaryBuffer = "";
}
foreach ( $lines as $line ) {
if ( $line === "" ) { continue; }
if ( strpos( $line, 'data:' ) === 0 ) {
$line = trim( substr( $line, 5 ) );
$json = json_decode( trim( $line ), true );
if ( json_last_error() === JSON_ERROR_NONE ) {
$content = $this->stream_data_handler( $json );
if ( !is_null( $content ) ) {
// TO CHECK: Not sure why we need to do this to make sure there is a line return in the chatbot
// If we don't do this, HuggingFace streams "\n" as a token without anything else, and the
// chatbot doesn't display it.
if ( $content === "\n" ) {
$content = " \n";
}
$this->streamContent .= $content;
call_user_func( $this->streamCallback, $content );
}
}
else if ( $line !== '[DONE]' && !empty( $line ) ) {
$this->streamTemporaryBuffer .= $line . "\n";
}
}
}
return $length;
});
}
protected function stream_header_handler( $header ) {
}
protected function stream_data_handler( $json ) {
throw new Exception( 'Not implemented.' );
}
public function get_models() {
throw new Exception( 'Not implemented.' );
}
public function retrieve_models() {
throw new Exception( 'Not implemented.' );
}
public function run_completion_query( Meow_MWAI_Query_Base $query, $streamCallback = null ) : Meow_MWAI_Reply {
throw new Exception( 'Not implemented.' );
}
public function run_assistant_query( Meow_MWAI_Query_Assistant $query, $streamCallback = null ) : Meow_MWAI_Reply {
throw new Exception( 'Not implemented, or not supported in this version of AI Engine.' );
}
public function run_embedding_query( Meow_MWAI_Query_Base $query ) {
throw new Exception( 'Not implemented.' );
}
public function run_images_query( Meow_MWAI_Query_Base $query ) {
throw new Exception( 'Not implemented.' );
}
public function run_transcribe_query( Meow_MWAI_Query_Base $query ) {
throw new Exception( 'Not implemented.' );
}
public function get_price( Meow_MWAI_Query_Base $query, Meow_MWAI_Reply $reply ) {
throw new Exception( 'Not implemented.' );
}
}