anthropic.php
2 days ago
chatml.php
2 days ago
core.php
1 day 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
custom.php
222 lines
| 1 | <?php |
| 2 | |
| 3 | /** |
| 4 | * Generic OpenAI-compatible engine. Lets users point AI Engine at any server that speaks the |
| 5 | * OpenAI Chat Completions API (Ollama, LM Studio, vLLM, llama.cpp, LocalAI, TGI in OAI mode, |
| 6 | * smaller hosted providers, etc.). Endpoint is user-configurable; API key is optional since |
| 7 | * many local servers run unauthenticated. |
| 8 | */ |
| 9 | class Meow_MWAI_Engines_Custom extends Meow_MWAI_Engines_ChatML { |
| 10 | public function __construct( $core, $env ) { |
| 11 | parent::__construct( $core, $env ); |
| 12 | } |
| 13 | |
| 14 | protected function set_environment() { |
| 15 | $env = $this->env; |
| 16 | $this->apiKey = $env['apikey'] ?? null; |
| 17 | } |
| 18 | |
| 19 | protected function get_service_name() { |
| 20 | return !empty( $this->env['name'] ) ? $this->env['name'] : 'Custom'; |
| 21 | } |
| 22 | |
| 23 | public function get_models() { |
| 24 | // Prefer dynamically-fetched models, fall back to whatever the user added manually for this |
| 25 | // env type via the Custom Models UI. No static list, since the model lineup is whatever the |
| 26 | // user's server happens to expose. |
| 27 | return $this->core->get_engine_models( 'custom' ); |
| 28 | } |
| 29 | |
| 30 | public static function get_models_static() { |
| 31 | return []; |
| 32 | } |
| 33 | |
| 34 | protected function build_url( $query, $endpoint = null ) { |
| 35 | $base = $this->resolve_endpoint(); |
| 36 | if ( $query instanceof Meow_MWAI_Query_Text || $query instanceof Meow_MWAI_Query_Feedback ) { |
| 37 | return $base . '/chat/completions'; |
| 38 | } |
| 39 | if ( $query instanceof Meow_MWAI_Query_Embed ) { |
| 40 | return $base . '/embeddings'; |
| 41 | } |
| 42 | throw new Exception( 'Unsupported query type for the Custom (OpenAI-Compatible) provider.' ); |
| 43 | } |
| 44 | |
| 45 | /** |
| 46 | * Bearer auth is optional here. Most local servers (Ollama default install, LM Studio, |
| 47 | * llama.cpp server) do not require an API key; hosted OpenAI-compatible endpoints typically |
| 48 | * do. Only attach the Authorization header when the user has set a key. |
| 49 | */ |
| 50 | protected function build_headers( $query ) { |
| 51 | if ( $query->apiKey ) { |
| 52 | $this->apiKey = $query->apiKey; |
| 53 | } |
| 54 | $headers = [ |
| 55 | 'Content-Type' => 'application/json', |
| 56 | 'User-Agent' => 'AI Engine', |
| 57 | ]; |
| 58 | if ( !empty( $this->apiKey ) ) { |
| 59 | $headers['Authorization'] = 'Bearer ' . $this->apiKey; |
| 60 | } |
| 61 | return $headers; |
| 62 | } |
| 63 | |
| 64 | protected function build_body( $query, $streamCallback = null, $extra = null ) { |
| 65 | $body = parent::build_body( $query, $streamCallback, $extra ); |
| 66 | |
| 67 | // Most OAI-compatible servers expect the older max_tokens field, not max_completion_tokens. |
| 68 | if ( isset( $body['max_completion_tokens'] ) ) { |
| 69 | $body['max_tokens'] = $body['max_completion_tokens']; |
| 70 | unset( $body['max_completion_tokens'] ); |
| 71 | } |
| 72 | |
| 73 | return $body; |
| 74 | } |
| 75 | |
| 76 | /** |
| 77 | * Resolve the endpoint URL. Defaults to the Ollama localhost URL since that is the most |
| 78 | * common starting point; users override per-env in settings. |
| 79 | */ |
| 80 | private function resolve_endpoint() { |
| 81 | $endpoint = !empty( $this->env['endpoint'] ) ? $this->env['endpoint'] : 'http://localhost:11434/v1'; |
| 82 | $endpoint = apply_filters( 'mwai_custom_endpoint', $endpoint, $this->env ); |
| 83 | return rtrim( $endpoint, '/' ); |
| 84 | } |
| 85 | |
| 86 | /** |
| 87 | * Try to discover models from /v1/models. Servers that don't implement this endpoint |
| 88 | * (some llama.cpp builds, custom proxies) just return an empty list — users add models |
| 89 | * manually through AI Engine's existing custom-models UI. |
| 90 | */ |
| 91 | public function retrieve_models() { |
| 92 | $base = $this->resolve_endpoint(); |
| 93 | $url = $base . '/models'; |
| 94 | |
| 95 | $headers = [ 'User-Agent' => 'AI Engine' ]; |
| 96 | if ( !empty( $this->apiKey ) ) { |
| 97 | $headers['Authorization'] = 'Bearer ' . $this->apiKey; |
| 98 | } |
| 99 | |
| 100 | $response = wp_remote_get( $url, [ |
| 101 | 'headers' => $headers, |
| 102 | 'timeout' => 10, |
| 103 | 'sslverify' => MWAI_SSL_VERIFY, |
| 104 | ] ); |
| 105 | |
| 106 | if ( is_wp_error( $response ) ) { |
| 107 | Meow_MWAI_Logging::log( 'Custom (OpenAI-Compatible): /models fetch failed: ' . $response->get_error_message() ); |
| 108 | return []; |
| 109 | } |
| 110 | |
| 111 | $code = wp_remote_retrieve_response_code( $response ); |
| 112 | if ( $code >= 400 ) { |
| 113 | Meow_MWAI_Logging::log( "Custom (OpenAI-Compatible): /models returned HTTP {$code}." ); |
| 114 | return []; |
| 115 | } |
| 116 | |
| 117 | $body = json_decode( wp_remote_retrieve_body( $response ), true ); |
| 118 | if ( !isset( $body['data'] ) || !is_array( $body['data'] ) ) { |
| 119 | return []; |
| 120 | } |
| 121 | |
| 122 | $models = []; |
| 123 | foreach ( $body['data'] as $remote ) { |
| 124 | $modelId = $remote['id'] ?? ''; |
| 125 | if ( empty( $modelId ) ) { |
| 126 | continue; |
| 127 | } |
| 128 | $isEmbedding = strpos( strtolower( $modelId ), 'embed' ) !== false; |
| 129 | $features = $isEmbedding ? [ 'embedding' ] : [ 'completion' ]; |
| 130 | $tags = [ 'core', $isEmbedding ? 'embedding' : 'chat' ]; |
| 131 | |
| 132 | $modelData = [ |
| 133 | 'model' => $modelId, |
| 134 | 'name' => $modelId, |
| 135 | 'family' => 'custom', |
| 136 | 'features' => $features, |
| 137 | // Pricing is unknown for self-hosted/third-party servers — zero out so usage tracking |
| 138 | // doesn't invent costs. Users can add per-model pricing through the Custom Models UI. |
| 139 | 'price' => [ 'in' => 0, 'out' => 0 ], |
| 140 | 'type' => 'token', |
| 141 | 'unit' => 1 / 1000000, |
| 142 | 'maxCompletionTokens' => isset( $remote['max_output_tokens'] ) ? (int) $remote['max_output_tokens'] : 4096, |
| 143 | 'maxContextualTokens' => isset( $remote['context_window'] ) ? (int) $remote['context_window'] : 8192, |
| 144 | 'tags' => $tags, |
| 145 | ]; |
| 146 | |
| 147 | // Only report dimensions when the server actually tells us. The OpenAI /models |
| 148 | // schema has no such field, so defaulting to 1536 made every embedding model look |
| 149 | // like a fixed 1536 one: the UI locked the field and we sent dimensions=1536 to |
| 150 | // servers whose model has another native size, or no support for the parameter at |
| 151 | // all (bad request). Unknown means the user can type the right value, and we send |
| 152 | // nothing until they do. |
| 153 | if ( $isEmbedding && isset( $remote['dimensions'] ) ) { |
| 154 | $modelData['dimensions'] = (int) $remote['dimensions']; |
| 155 | } |
| 156 | |
| 157 | $models[] = $modelData; |
| 158 | } |
| 159 | return $models; |
| 160 | } |
| 161 | |
| 162 | /** |
| 163 | * Hits /v1/models directly so the user gets a real success/error response. Works without an |
| 164 | * API key, surfaces real HTTP errors when they happen. |
| 165 | */ |
| 166 | public function connection_check() { |
| 167 | $base = $this->resolve_endpoint(); |
| 168 | $url = $base . '/models'; |
| 169 | $details = [ 'endpoint' => $url ]; |
| 170 | |
| 171 | $headers = [ 'User-Agent' => 'AI Engine' ]; |
| 172 | if ( !empty( $this->apiKey ) ) { |
| 173 | $headers['Authorization'] = 'Bearer ' . $this->apiKey; |
| 174 | } |
| 175 | |
| 176 | $response = wp_remote_get( $url, [ |
| 177 | 'headers' => $headers, |
| 178 | 'timeout' => 10, |
| 179 | 'sslverify' => MWAI_SSL_VERIFY, |
| 180 | ] ); |
| 181 | |
| 182 | if ( is_wp_error( $response ) ) { |
| 183 | return [ |
| 184 | 'success' => false, 'service' => $this->get_service_name(), |
| 185 | 'error' => $response->get_error_message(), |
| 186 | 'details' => $details, |
| 187 | ]; |
| 188 | } |
| 189 | |
| 190 | $code = wp_remote_retrieve_response_code( $response ); |
| 191 | $body = json_decode( wp_remote_retrieve_body( $response ), true ); |
| 192 | |
| 193 | if ( $code >= 400 || ( is_array( $body ) && isset( $body['error'] ) ) ) { |
| 194 | $message = is_array( $body ) && isset( $body['error'] ) |
| 195 | ? ( is_string( $body['error'] ) ? $body['error'] : ( $body['error']['message'] ?? json_encode( $body['error'] ) ) ) |
| 196 | : "HTTP {$code} from {$url}."; |
| 197 | return [ |
| 198 | 'success' => false, 'service' => $this->get_service_name(), |
| 199 | 'error' => $message, |
| 200 | 'details' => array_merge( $details, [ 'http_code' => $code ] ), |
| 201 | ]; |
| 202 | } |
| 203 | |
| 204 | $modelIds = []; |
| 205 | if ( isset( $body['data'] ) && is_array( $body['data'] ) ) { |
| 206 | foreach ( array_slice( $body['data'], 0, 5 ) as $m ) { |
| 207 | if ( isset( $m['id'] ) ) { |
| 208 | $modelIds[] = $m['id']; |
| 209 | } |
| 210 | } |
| 211 | } |
| 212 | return [ |
| 213 | 'success' => true, 'service' => $this->get_service_name(), |
| 214 | 'message' => 'Connection successful. Found ' . count( $body['data'] ?? [] ) . ' models.', |
| 215 | 'details' => array_merge( $details, [ |
| 216 | 'model_count' => count( $body['data'] ?? [] ), |
| 217 | 'sample_models' => $modelIds, |
| 218 | ] ), |
| 219 | ]; |
| 220 | } |
| 221 | } |
| 222 |