anthropic.php
2 days ago
chatml.php
2 days ago
core.php
23 hours ago
custom.php
2 days ago
factory.php
1 month ago
google-interactions.php
2 days ago
google.php
2 days ago
mistral.php
1 month ago
open-router.php
1 month ago
openai.php
2 days ago
ovh.php
2 days ago
perplexity.php
7 months ago
replicate.php
6 months ago
xai.php
2 months ago
ovh.php
362 lines
| 1 | <?php |
| 2 | |
| 3 | class Meow_MWAI_Engines_OVH extends Meow_MWAI_Engines_ChatML { |
| 4 | private $endpoint = 'https://oai.endpoints.kepler.ai.cloud.ovh.net'; |
| 5 | public $envType = 'ovh'; |
| 6 | |
| 7 | public function __construct( $core, $env ) { |
| 8 | parent::__construct( $core, $env ); |
| 9 | } |
| 10 | |
| 11 | /** |
| 12 | * Sets up the environment for our custom engine. |
| 13 | */ |
| 14 | protected function set_environment() { |
| 15 | $env = $this->env; |
| 16 | |
| 17 | if ( !isset( $env['apikey'] ) || empty( $env['apikey'] ) ) { |
| 18 | throw new Exception( 'AI Engine: OVH API key is not set. Please configure the OVH AI Endpoints access token in your settings.' ); |
| 19 | } |
| 20 | |
| 21 | $this->apiKey = $env['apikey']; |
| 22 | |
| 23 | if ( isset( $env['endpoint'] ) && !empty( $env['endpoint'] ) ) { |
| 24 | $this->endpoint = $env['endpoint']; |
| 25 | } |
| 26 | } |
| 27 | |
| 28 | protected function build_url( $query, $endpoint = null ) { |
| 29 | $endpoint = apply_filters( 'mwai_ovh_endpoint', trailingslashit( $this->endpoint ) . 'v1', $this->env ); |
| 30 | return $endpoint . '/chat/completions'; |
| 31 | } |
| 32 | |
| 33 | protected function get_service_name() { |
| 34 | return 'OVH'; |
| 35 | } |
| 36 | |
| 37 | protected function build_messages( $query ) { |
| 38 | // Handle vision models like llava-next-mistral-7b |
| 39 | if ( strpos( $query->model, 'llava' ) !== false ) { |
| 40 | $query->image_remote_upload = 'data'; |
| 41 | } |
| 42 | |
| 43 | $messages = parent::build_messages( $query ); |
| 44 | return $messages; |
| 45 | } |
| 46 | |
| 47 | protected function stream_data_handler( $json ) { |
| 48 | // Handle usage-only chunks (final chunk with empty choices array). |
| 49 | // OVH sends usage data in a final chunk with choices: [] |
| 50 | if ( isset( $json['usage'] ) && empty( $json['choices'] ) ) { |
| 51 | $usage = $json['usage']; |
| 52 | if ( isset( $usage['prompt_tokens'], $usage['completion_tokens'] ) ) { |
| 53 | $this->streamInTokens = (int) $usage['prompt_tokens']; |
| 54 | $this->streamOutTokens = (int) $usage['completion_tokens']; |
| 55 | } |
| 56 | return null; // No content in this chunk |
| 57 | } |
| 58 | |
| 59 | // Let parent handle all other chunks |
| 60 | return parent::stream_data_handler( $json ); |
| 61 | } |
| 62 | |
| 63 | public function get_models() { |
| 64 | return $this->retrieve_models(); |
| 65 | } |
| 66 | |
| 67 | protected function build_headers( $query ) { |
| 68 | $headers = [ |
| 69 | 'Content-Type' => 'application/json', |
| 70 | 'Authorization' => 'Bearer ' . $this->apiKey, |
| 71 | ]; |
| 72 | return $headers; |
| 73 | } |
| 74 | |
| 75 | protected function build_body( $query, $streamCallback = null, $extra = null ) { |
| 76 | $body = parent::build_body( $query, $streamCallback, $extra ); |
| 77 | |
| 78 | // Handle max_tokens parameter (OVH uses max_tokens, not max_completion_tokens) |
| 79 | if ( isset( $body['max_completion_tokens'] ) ) { |
| 80 | $body['max_tokens'] = $body['max_completion_tokens']; |
| 81 | unset( $body['max_completion_tokens'] ); |
| 82 | } |
| 83 | |
| 84 | // OVH does not error when max_tokens cannot fit: it returns 200 with an |
| 85 | // instantly-terminated stream (a lone [DONE]), which looked like the AI |
| 86 | // answering nothing. Its models also share ONE window between prompt and |
| 87 | // completion (max output == context), so any max_tokens at or above the |
| 88 | // model's cap can never fit once the prompt is counted. In that case drop |
| 89 | // the parameter and let the endpoint default to the remaining context. |
| 90 | if ( !empty( $body['max_tokens'] ) ) { |
| 91 | $modelInfo = $this->retrieve_model_info( $query->model ); |
| 92 | if ( !empty( $modelInfo['maxCompletionTokens'] ) && |
| 93 | (int) $body['max_tokens'] >= (int) $modelInfo['maxCompletionTokens'] ) { |
| 94 | unset( $body['max_tokens'] ); |
| 95 | } |
| 96 | } |
| 97 | |
| 98 | // OVH supports stream_options for accurate token usage in streaming |
| 99 | if ( !empty( $streamCallback ) && !isset( $body['stream_options'] ) ) { |
| 100 | $body['stream_options'] = [ |
| 101 | 'include_usage' => true, |
| 102 | ]; |
| 103 | } |
| 104 | |
| 105 | // Remove parallel_tool_calls - OVH might not support it |
| 106 | if ( isset( $body['parallel_tool_calls'] ) ) { |
| 107 | unset( $body['parallel_tool_calls'] ); |
| 108 | } |
| 109 | |
| 110 | // Remove other potentially unsupported parameters |
| 111 | $unsupported_params = [ 'response_format', 'seed', 'logit_bias', 'logprobs', 'top_logprobs' ]; |
| 112 | foreach ( $unsupported_params as $param ) { |
| 113 | if ( isset( $body[$param] ) ) { |
| 114 | unset( $body[$param] ); |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | return $body; |
| 119 | } |
| 120 | |
| 121 | public function handle_tokens_usage( |
| 122 | $reply, |
| 123 | $query, |
| 124 | $returned_model, |
| 125 | $returned_in_tokens, |
| 126 | $returned_out_tokens, |
| 127 | $returned_price = null |
| 128 | ) { |
| 129 | |
| 130 | // Clean up the data |
| 131 | $returned_in_tokens = !is_null( $returned_in_tokens ) ? |
| 132 | $returned_in_tokens : $reply->get_in_tokens( $query ); |
| 133 | $returned_out_tokens = !is_null( $returned_out_tokens ) ? |
| 134 | $returned_out_tokens : $reply->get_out_tokens(); |
| 135 | |
| 136 | // Calculate price based on our model definitions |
| 137 | $models = $this->get_ovh_models(); |
| 138 | $model_price = null; |
| 139 | |
| 140 | foreach ( $models as $model ) { |
| 141 | if ( $model['model'] === $returned_model ) { |
| 142 | $model_price = $model['price']; |
| 143 | break; |
| 144 | } |
| 145 | } |
| 146 | |
| 147 | if ( $model_price ) { |
| 148 | $returned_price = ( $returned_in_tokens * $model_price['in'] + |
| 149 | $returned_out_tokens * $model_price['out'] ) / 1000000; |
| 150 | } |
| 151 | else { |
| 152 | $returned_price = 0; |
| 153 | } |
| 154 | |
| 155 | // Record the usage |
| 156 | $usage = $this->core->record_tokens_usage( |
| 157 | $returned_model, |
| 158 | $returned_in_tokens, |
| 159 | $returned_out_tokens, |
| 160 | $returned_price |
| 161 | ); |
| 162 | |
| 163 | // Set the usage in the reply |
| 164 | $reply->set_usage( $usage ); |
| 165 | |
| 166 | // Set accuracy based on data availability |
| 167 | if ( !is_null( $returned_in_tokens ) && !is_null( $returned_out_tokens ) ) { |
| 168 | // Tokens from API (via stream_options), price calculated locally |
| 169 | $reply->set_usage_accuracy( 'tokens' ); |
| 170 | } |
| 171 | else { |
| 172 | // Everything estimated |
| 173 | $reply->set_usage_accuracy( 'estimated' ); |
| 174 | } |
| 175 | } |
| 176 | |
| 177 | public function get_price( Meow_MWAI_Query_Base $query, Meow_MWAI_Reply $reply ) { |
| 178 | $models = $this->get_ovh_models(); |
| 179 | |
| 180 | foreach ( $models as $model ) { |
| 181 | if ( $model['model'] === $query->model ) { |
| 182 | $in_tokens = $reply->get_in_tokens( $query ); |
| 183 | $out_tokens = $reply->get_out_tokens(); |
| 184 | return ( $in_tokens * $model['price']['in'] + |
| 185 | $out_tokens * $model['price']['out'] ) / 1000000; |
| 186 | } |
| 187 | } |
| 188 | |
| 189 | return 0; |
| 190 | } |
| 191 | |
| 192 | /** |
| 193 | * Retrieve the models from the OVH OpenRouter-compatible catalog. |
| 194 | */ |
| 195 | public function retrieve_models() { |
| 196 | // Per-request cache: build_body and the pricing helpers may all ask for |
| 197 | // the model list during a single query. |
| 198 | static $cached = null; |
| 199 | if ( $cached !== null ) { |
| 200 | return $cached; |
| 201 | } |
| 202 | |
| 203 | $url = 'https://catalog.endpoints.ai.ovh.net/rest/v2/openrouter'; |
| 204 | |
| 205 | // This can run in the middle of a streaming completion, where the |
| 206 | // engine's http_api_curl stream handler is registered: it would hijack |
| 207 | // this fetch too (empty body, catalog "streamed" to nowhere) and we would |
| 208 | // silently fall back to the wrong model list. Detach the hooks around it. |
| 209 | global $wp_filter; |
| 210 | $saved_hooks = null; |
| 211 | if ( isset( $wp_filter['http_api_curl'] ) ) { |
| 212 | $saved_hooks = $wp_filter['http_api_curl']; |
| 213 | unset( $wp_filter['http_api_curl'] ); |
| 214 | } |
| 215 | $response = wp_remote_get( $url, [ 'timeout' => 10 ] ); |
| 216 | if ( $saved_hooks !== null ) { |
| 217 | $wp_filter['http_api_curl'] = $saved_hooks; |
| 218 | } |
| 219 | |
| 220 | if ( is_wp_error( $response ) ) { |
| 221 | return $this->get_fallback_models(); |
| 222 | } |
| 223 | |
| 224 | $body = wp_remote_retrieve_body( $response ); |
| 225 | $data = json_decode( $body, true ); |
| 226 | |
| 227 | if ( empty( $data ) || !is_array( $data ) ) { |
| 228 | return $this->get_fallback_models(); |
| 229 | } |
| 230 | |
| 231 | // Extract models from 'data' key |
| 232 | $models_data = isset( $data['data'] ) ? $data['data'] : $data; |
| 233 | |
| 234 | if ( empty( $models_data ) || !is_array( $models_data ) ) { |
| 235 | return $this->get_fallback_models(); |
| 236 | } |
| 237 | |
| 238 | $models = []; |
| 239 | foreach ( $models_data as $model_data ) { |
| 240 | $model = $this->map_openrouter_model( $model_data ); |
| 241 | if ( $model ) { |
| 242 | $models[] = $model; |
| 243 | } |
| 244 | } |
| 245 | |
| 246 | $cached = !empty( $models ) ? $models : $this->get_fallback_models(); |
| 247 | return $cached; |
| 248 | } |
| 249 | |
| 250 | /** |
| 251 | * Map OpenRouter model data to AI Engine format. |
| 252 | */ |
| 253 | private function map_openrouter_model( $model_data ) { |
| 254 | if ( empty( $model_data['id'] ) ) { |
| 255 | return null; |
| 256 | } |
| 257 | |
| 258 | $model_id = $model_data['id']; |
| 259 | $name = !empty( $model_data['name'] ) ? $model_data['name'] : $model_id; |
| 260 | |
| 261 | // Extract family from model ID |
| 262 | $family_parts = explode( '/', $model_id ); |
| 263 | $family = $family_parts[0]; |
| 264 | |
| 265 | // Base features and tags |
| 266 | $features = [ 'completion', 'chat' ]; |
| 267 | $tags = [ 'core', 'chat' ]; |
| 268 | |
| 269 | // Check for vision support (input modality includes images) |
| 270 | if ( !empty( $model_data['input_modalities'] ) && is_array( $model_data['input_modalities'] ) ) { |
| 271 | if ( in_array( 'image', $model_data['input_modalities'] ) ) { |
| 272 | $features[] = 'vision'; |
| 273 | $tags[] = 'vision'; |
| 274 | } |
| 275 | } |
| 276 | |
| 277 | // Map features from OpenRouter to AI Engine format |
| 278 | $openrouter_features = !empty( $model_data['supported_features'] ) ? $model_data['supported_features'] : []; |
| 279 | foreach ( $openrouter_features as $feature ) { |
| 280 | switch ( $feature ) { |
| 281 | case 'json_mode': |
| 282 | case 'structured_outputs': |
| 283 | if ( !in_array( 'json', $tags ) ) { |
| 284 | $tags[] = 'json'; |
| 285 | } |
| 286 | break; |
| 287 | case 'reasoning': |
| 288 | if ( !in_array( 'reasoning', $tags ) ) { |
| 289 | $tags[] = 'reasoning'; |
| 290 | } |
| 291 | break; |
| 292 | case 'tools': |
| 293 | // OVH now supports function calling via tools format |
| 294 | if ( !in_array( 'functions', $tags ) ) { |
| 295 | $tags[] = 'functions'; |
| 296 | } |
| 297 | break; |
| 298 | } |
| 299 | } |
| 300 | |
| 301 | // Convert pricing from per-token to per-million tokens |
| 302 | $price_in = 0; |
| 303 | $price_out = 0; |
| 304 | if ( !empty( $model_data['pricing']['prompt'] ) ) { |
| 305 | $price_in = floatval( $model_data['pricing']['prompt'] ) * 1000000; |
| 306 | } |
| 307 | if ( !empty( $model_data['pricing']['completion'] ) ) { |
| 308 | $price_out = floatval( $model_data['pricing']['completion'] ) * 1000000; |
| 309 | } |
| 310 | |
| 311 | // Get context and max tokens |
| 312 | $max_context = !empty( $model_data['context_length'] ) ? intval( $model_data['context_length'] ) : 128000; |
| 313 | $max_completion = !empty( $model_data['max_output_length'] ) ? intval( $model_data['max_output_length'] ) : 4096; |
| 314 | |
| 315 | return [ |
| 316 | 'model' => $model_id, |
| 317 | 'name' => $name, |
| 318 | 'family' => $family, |
| 319 | 'features' => $features, |
| 320 | 'price' => [ |
| 321 | 'in' => $price_in, |
| 322 | 'out' => $price_out, |
| 323 | ], |
| 324 | 'type' => 'token', |
| 325 | 'unit' => 1 / 1000000, |
| 326 | 'maxCompletionTokens' => $max_completion, |
| 327 | 'maxContextualTokens' => $max_context, |
| 328 | 'tags' => $tags, |
| 329 | ]; |
| 330 | } |
| 331 | |
| 332 | /** |
| 333 | * Get fallback models in case the catalog API is unavailable. |
| 334 | */ |
| 335 | private function get_fallback_models() { |
| 336 | return [ |
| 337 | [ |
| 338 | 'model' => 'meta-llama/Llama-3.3-70B-Instruct', |
| 339 | 'name' => 'Llama 3.3 70B Instruct', |
| 340 | 'family' => 'meta-llama', |
| 341 | 'features' => [ 'completion', 'chat' ], |
| 342 | 'price' => [ |
| 343 | 'in' => 0.06, |
| 344 | 'out' => 0.09, |
| 345 | ], |
| 346 | 'type' => 'token', |
| 347 | 'unit' => 1 / 1000000, |
| 348 | 'maxCompletionTokens' => 8192, |
| 349 | 'maxContextualTokens' => 131072, |
| 350 | 'tags' => [ 'core', 'chat', 'json', 'reasoning' ], |
| 351 | ], |
| 352 | ]; |
| 353 | } |
| 354 | |
| 355 | /** |
| 356 | * Get model info by model ID (for backward compatibility). |
| 357 | */ |
| 358 | private function get_ovh_models() { |
| 359 | return $this->retrieve_models(); |
| 360 | } |
| 361 | } |
| 362 |