| @@ -11,28 +11,64 @@ | ||
| 11 | 11 | declare(strict_types=1); |
| 12 | 12 | |
| 13 | 13 | namespace ThinkRank\AI; |
| 14 | 14 | |
| 15 | +use ThinkRank\AI\Traits\Request_Timeout; | |
| 16 | + | |
| 15 | 17 | // Prevent direct access |
| 16 | 18 | if (!defined('ABSPATH')) { |
| 17 | 19 | exit; |
| 18 | 20 | } |
| 19 | 21 | |
| 22 | +require_once __DIR__ . '/traits/trait-request-timeout.php'; | |
| 23 | +require_once __DIR__ . '/class-endpoint-url-validator.php'; | |
| 24 | + | |
| 20 | 25 | /** |
| 21 | 26 | * OpenAI Client Class |
| 22 | - * | |
| 27 | + * | |
| 23 | 28 | * Single Responsibility: Handle OpenAI API communication |
| 24 | - * | |
| 29 | + * | |
| 25 | 30 | * @since 1.0.0 |
| 26 | 31 | */ |
| 27 | 32 | class OpenAI_Client { |
| 33 | + | |
| 34 | + use Request_Timeout; | |
| 35 | + | |
| 28 | 36 | |
| 29 | 37 | /** |
| 30 | - * OpenAI API base URL | |
| 38 | + * OpenAI's own API base URL — the default when no other is given. | |
| 31 | 39 | */ |
| 32 | - private const API_BASE_URL = 'https://api.openai.com/v1'; | |
| 40 | + public const API_BASE_URL = 'https://api.openai.com/v1'; | |
| 41 | + | |
| 42 | + /** | |
| 43 | + * Base URL every request is built on. | |
| 44 | + * | |
| 45 | + * Not always OpenAI's: the `openai_compatible` provider (#721) points this | |
| 46 | + * at any server speaking the Chat Completions API — Ollama, LM Studio, | |
| 47 | + * vLLM, Azure OpenAI, a company gateway — so the rest of this client, the | |
| 48 | + * retry policy and the response parsing are shared verbatim. | |
| 49 | + * | |
| 50 | + * @since 2.8.0 | |
| 51 | + * @var string | |
| 52 | + */ | |
| 53 | + private string $base_url; | |
| 33 | 54 | |
| 34 | 55 | /** |
| 56 | + * Ceiling for a response body from a user-named endpoint. | |
| 57 | + * | |
| 58 | + * Mirrors Endpoint_URL_Validator::MAX_RESPONSE_BYTES; kept here too so the | |
| 59 | + * truncation message and the transport limit cannot drift apart. | |
| 60 | + */ | |
| 61 | + private const MAX_RESPONSE_BYTES = 2097152; // 2 MB. | |
| 62 | + | |
| 63 | + /** | |
| 64 | + * Output-token ceiling for an unrecognised model on a user-named endpoint. | |
| 65 | + * | |
| 66 | + * See get_max_completion_tokens(). | |
| 67 | + */ | |
| 68 | + private const COMPATIBLE_DEFAULT_MAX_TOKENS = 8192; | |
| 69 | + | |
| 70 | + /** | |
| 35 | 71 | * API key |
| 36 | 72 | * |
| 37 | 73 | * @var string |
| 38 | 74 | */ |
| @@ -52,8 +88,19 @@ | ||
| 52 | 88 | */ |
| 53 | 89 | private int $timeout; |
| 54 | 90 | |
| 55 | 91 | /** |
| 92 | + * Ask a custom endpoint to constrain JSON answers with response_format. | |
| 93 | + * | |
| 94 | + * Off by default because not every compatible server takes it (LM Studio | |
| 95 | + * accepts only json_schema). See set_json_mode(). | |
| 96 | + * | |
| 97 | + * @since 2.8.0 | |
| 98 | + * @var bool | |
| 99 | + */ | |
| 100 | + private bool $json_mode = false; | |
| 101 | + | |
| 102 | + /** | |
| 56 | 103 | * Prompt Builder instance |
| 57 | 104 | * |
| 58 | 105 | * @since 1.0.0 |
| 59 | 106 | * @var Prompt_Builder|null |
| @@ -62,19 +109,83 @@ | ||
| 62 | 109 | |
| 63 | 110 | /** |
| 64 | 111 | * Constructor |
| 65 | 112 | * |
| 66 | - * @param string $api_key OpenAI API key | |
| 67 | - * @param string $model Default model to use | |
| 68 | - * @param int $timeout Request timeout | |
| 113 | + * @param string $api_key OpenAI API key (may be empty for a local server that wants none). | |
| 114 | + * @param string $model Default model to use | |
| 115 | + * @param int $timeout Request timeout | |
| 116 | + * @param string $base_url API base URL without a trailing slash; defaults to OpenAI's. | |
| 69 | 117 | */ |
| 70 | - public function __construct(string $api_key, string $model = \ThinkRank\Core\Settings::DEFAULT_OPENAI_MODEL, int $timeout = 30) { | |
| 118 | + public function __construct(string $api_key, string $model = \ThinkRank\Core\Settings::DEFAULT_OPENAI_MODEL, int $timeout = 30, string $base_url = self::API_BASE_URL) { | |
| 71 | 119 | $this->api_key = $api_key; |
| 72 | 120 | $this->model = $model; |
| 73 | 121 | $this->timeout = $timeout; |
| 122 | + $base_url = rtrim(trim($base_url), '/'); | |
| 123 | + $this->base_url = '' !== $base_url ? $base_url : self::API_BASE_URL; | |
| 74 | 124 | } |
| 75 | 125 | |
| 76 | 126 | /** |
| 127 | + * Turn on response_format: json_object for calls that want a JSON object. | |
| 128 | + * | |
| 129 | + * ThinkRank otherwise enforces JSON through the prompt alone, and a small | |
| 130 | + * local model then writes an unescaped quote inside a string (DeepSeek-R1 | |
| 131 | + * 7B quoting a phrase in the brief's HTML body) and the whole brief fails | |
| 132 | + * to parse. Ollama, vLLM and llama.cpp turn this field into a grammar, so | |
| 133 | + * the reply cannot be malformed JSON. Only calls passing the `json_object` | |
| 134 | + * option get it: a plain-text caller would be forced into JSON too. | |
| 135 | + * | |
| 136 | + * @since 2.8.0 | |
| 137 | + * | |
| 138 | + * @param bool $enabled Whether the endpoint accepts response_format json_object. | |
| 139 | + * @return void | |
| 140 | + */ | |
| 141 | + public function set_json_mode(bool $enabled): void { | |
| 142 | + $this->json_mode = $enabled; | |
| 143 | + } | |
| 144 | + | |
| 145 | + /** | |
| 146 | + * Get the base URL this client talks to. | |
| 147 | + * | |
| 148 | + * @since 2.8.0 | |
| 149 | + * | |
| 150 | + * @return string Base URL without a trailing slash. | |
| 151 | + */ | |
| 152 | + public function get_base_url(): string { | |
| 153 | + return $this->base_url; | |
| 154 | + } | |
| 155 | + | |
| 156 | + /** | |
| 157 | + * Is this client pointed at a server other than OpenAI's? | |
| 158 | + * | |
| 159 | + * Used for error copy: naming "OpenAI" in a failure from someone's local | |
| 160 | + * Ollama box sends them to the wrong place to debug it. | |
| 161 | + * | |
| 162 | + * @since 2.8.0 | |
| 163 | + * | |
| 164 | + * @return bool | |
| 165 | + */ | |
| 166 | + private function is_custom_endpoint(): bool { | |
| 167 | + return self::API_BASE_URL !== $this->base_url; | |
| 168 | + } | |
| 169 | + | |
| 170 | + /** | |
| 171 | + * Name to use for this endpoint in user-facing messages. | |
| 172 | + * | |
| 173 | + * @since 2.8.0 | |
| 174 | + * | |
| 175 | + * @return string | |
| 176 | + */ | |
| 177 | + private function get_endpoint_label(): string { | |
| 178 | + if (!$this->is_custom_endpoint()) { | |
| 179 | + return 'OpenAI'; | |
| 180 | + } | |
| 181 | + | |
| 182 | + $host = wp_parse_url($this->base_url, PHP_URL_HOST); | |
| 183 | + | |
| 184 | + return is_string($host) && '' !== $host ? $host : __('the AI endpoint', 'thinkrank'); | |
| 185 | + } | |
| 186 | + | |
| 187 | + /** | |
| 77 | 188 | * Get Prompt Builder instance |
| 78 | 189 | * |
| 79 | 190 | * @since 1.0.0 |
| 80 | 191 | * |
| @@ -109,9 +220,30 @@ | ||
| 109 | 220 | 'presence_penalty' => 0, |
| 110 | 221 | ]; |
| 111 | 222 | |
| 112 | 223 | $options = array_merge($default_options, $options); |
| 113 | - | |
| 224 | + | |
| 225 | + $body = $this->build_chat_completion_body($prompt, $options); | |
| 226 | + | |
| 227 | + return $this->make_request('chat/completions', $body); | |
| 228 | + } | |
| 229 | + | |
| 230 | + /** | |
| 231 | + * Build the chat/completions request body for the given (merged) options. | |
| 232 | + * | |
| 233 | + * Extracted so the per-model-family parameter handling is unit-testable: | |
| 234 | + * reasoning models take max_completion_tokens (and only the GPT-5 family | |
| 235 | + * accepts reasoning_effort — o1/o3 reject it), while standard models take | |
| 236 | + * temperature/top_p/penalties/max_tokens. Keeping this in one place stops a | |
| 237 | + * future refactor from silently regressing the GPT-5-only guard (issue #286). | |
| 238 | + * | |
| 239 | + * @param string $prompt User prompt. | |
| 240 | + * @param array $options Merged options (must include model, max_tokens, and | |
| 241 | + * the sampling defaults; reasoning_effort and | |
| 242 | + * json_object optional). | |
| 243 | + * @return array Request body for the chat/completions endpoint. | |
| 244 | + */ | |
| 245 | + private function build_chat_completion_body(string $prompt, array $options): array { | |
| 114 | 246 | $body = [ |
| 115 | 247 | 'model' => $options['model'], |
| 116 | 248 | 'messages' => [ |
| 117 | 249 | [ |
| @@ -130,11 +262,11 @@ | ||
| 130 | 262 | // temperature, top_p, frequency_penalty, presence_penalty are not supported |
| 131 | 263 | $body['max_completion_tokens'] = $safe_tokens; |
| 132 | 264 | |
| 133 | 265 | // GPT-5 models accept reasoning_effort ('minimal'…'high'). Callers |
| 134 | - // wanting a quick consumer-style answer (e.g. brand-visibility | |
| 135 | - // probes) pass 'minimal' so hidden reasoning can't consume the | |
| 136 | - // whole completion budget and return empty text. Only the GPT-5 | |
| 266 | + // wanting a quick answer pass a low level so hidden reasoning | |
| 267 | + // can't consume the whole completion budget and return empty | |
| 268 | + // text. Only the GPT-5 | |
| 137 | 269 | // family gets it: o1 rejects the parameter outright. |
| 138 | 270 | if (isset($options['reasoning_effort']) && str_starts_with($options['model'], 'gpt-5')) { |
| 139 | 271 | $body['reasoning_effort'] = (string) $options['reasoning_effort']; |
| 140 | 272 | } |
| @@ -146,9 +278,13 @@ | ||
| 146 | 278 | $body['presence_penalty'] = $options['presence_penalty']; |
| 147 | 279 | $body['max_tokens'] = $safe_tokens; |
| 148 | 280 | } |
| 149 | 281 | |
| 150 | - return $this->make_request('chat/completions', $body); | |
| 282 | + if ($this->json_mode && $this->is_custom_endpoint() && !empty($options['json_object'])) { | |
| 283 | + $body['response_format'] = ['type' => 'json_object']; | |
| 284 | + } | |
| 285 | + | |
| 286 | + return $body; | |
| 151 | 287 | } |
| 152 | 288 | |
| 153 | 289 | /** |
| 154 | 290 | * Generate SEO metadata |
| @@ -169,8 +305,9 @@ | ||
| 169 | 305 | |
| 170 | 306 | $response = $this->generate_completion($prompt, [ |
| 171 | 307 | 'max_tokens' => $this->get_recommended_tokens('seo_metadata'), |
| 172 | 308 | 'temperature' => 0.3, // Lower temperature for more consistent SEO output |
| 309 | + 'json_object' => true, | |
| 173 | 310 | ]); |
| 174 | 311 | |
| 175 | 312 | return $this->parse_seo_response($response); |
| 176 | 313 | } |
| @@ -189,8 +326,9 @@ | ||
| 189 | 326 | |
| 190 | 327 | $response = $this->generate_completion($prompt, [ |
| 191 | 328 | 'max_tokens' => $this->get_recommended_tokens('analysis'), |
| 192 | 329 | 'temperature' => 0.3, // Lower temperature for more consistent analysis |
| 330 | + 'json_object' => true, | |
| 193 | 331 | ]); |
| 194 | 332 | |
| 195 | 333 | return $this->parse_analysis_response($response); |
| 196 | 334 | } |
| @@ -256,8 +394,31 @@ | ||
| 256 | 394 | return $limit; |
| 257 | 395 | } |
| 258 | 396 | } |
| 259 | 397 | |
| 398 | + // A model we have never heard of on OpenAI's own API gets a | |
| 399 | + // conservative ceiling. On a user-named server nearly every model is | |
| 400 | + // unknown, and 4096 left a brief ~2.9K tokens after length scaling, | |
| 401 | + // which a reasoning model (DeepSeek-R1, Qwen3) spends half of thinking | |
| 402 | + // before the JSON starts. max_tokens is a cap, not a reservation, so a | |
| 403 | + // higher one costs nothing on a model that finishes early. | |
| 404 | + if ($this->is_custom_endpoint()) { | |
| 405 | + /** | |
| 406 | + * Filters the output-token ceiling for an unrecognised model on an | |
| 407 | + * OpenAI-compatible endpoint. | |
| 408 | + * | |
| 409 | + * Lower it for a server that rejects max_tokens beyond its context | |
| 410 | + * window (vLLM does); raise it for a large-context local model. | |
| 411 | + * | |
| 412 | + * @since 2.8.0 | |
| 413 | + * | |
| 414 | + * @param int $limit Ceiling in tokens. Default 8192. | |
| 415 | + * @param string $model Model id sent to the endpoint. | |
| 416 | + */ | |
| 417 | + $limit = (int) apply_filters('thinkrank_openai_compatible_max_output_tokens', self::COMPATIBLE_DEFAULT_MAX_TOKENS, $model); | |
| 418 | + return $limit > 0 ? $limit : self::COMPATIBLE_DEFAULT_MAX_TOKENS; | |
| 419 | + } | |
| 420 | + | |
| 260 | 421 | // Default fallback for unknown models |
| 261 | 422 | return 4096; |
| 262 | 423 | } |
| 263 | 424 | |
| @@ -390,21 +551,53 @@ | ||
| 390 | 551 | * @return array Response data |
| 391 | 552 | * @throws \Exception If request fails |
| 392 | 553 | */ |
| 393 | 554 | private function make_request(string $endpoint, array $body = []): array { |
| 394 | - $url = self::API_BASE_URL . '/' . ltrim($endpoint, '/'); | |
| 395 | - | |
| 555 | + // The user's daily ceiling and kill switch are enforced here, at the | |
| 556 | + // one place every outbound OpenAI call passes through, so no feature | |
| 557 | + // path can bypass them by forgetting to ask first (#448). | |
| 558 | + Spend_Guard::guard(); | |
| 559 | + Spend_Guard::record(); | |
| 560 | + | |
| 561 | + // Not string concatenation: a base URL may carry a query string (Azure | |
| 562 | + // requires ?api-version=…), and the route has to land before it (#721). | |
| 563 | + $url = Endpoint_URL_Validator::route($this->base_url, $endpoint); | |
| 564 | + | |
| 565 | + $headers = [ | |
| 566 | + 'Content-Type' => 'application/json', | |
| 567 | + 'User-Agent' => 'ThinkRank/' . THINKRANK_VERSION, | |
| 568 | + ]; | |
| 569 | + | |
| 570 | + // A local Ollama or LM Studio server wants no key at all; sending | |
| 571 | + // "Bearer " with nothing after it makes some gateways 401. | |
| 572 | + if ('' !== $this->api_key) { | |
| 573 | + $headers['Authorization'] = 'Bearer ' . $this->api_key; | |
| 574 | + // Azure OpenAI reads the key from its own header and ignores | |
| 575 | + // Authorization. Sending both costs nothing and makes an Azure | |
| 576 | + // deployment URL work without a separate provider. | |
| 577 | + $headers['api-key'] = $this->api_key; | |
| 578 | + } | |
| 579 | + | |
| 396 | 580 | $args = [ |
| 397 | 581 | 'timeout' => $this->timeout, |
| 398 | - 'headers' => [ | |
| 399 | - 'Authorization' => 'Bearer ' . $this->api_key, | |
| 400 | - 'Content-Type' => 'application/json', | |
| 401 | - 'User-Agent' => 'ThinkRank/' . THINKRANK_VERSION, | |
| 402 | - ], | |
| 582 | + 'headers' => $headers, | |
| 583 | + // Only for an endpoint we do not control: a ceiling generous enough | |
| 584 | + // for the largest thing we ask for (a content brief as JSON), so a | |
| 585 | + // server that ignores its token limit still cannot spend the | |
| 586 | + // worker's memory. Truncation is reported below, not parsed (#721). | |
| 587 | + 'limit_response_size' => $this->is_custom_endpoint() ? self::MAX_RESPONSE_BYTES : null, | |
| 588 | + // The key travels in these headers. A redirect to another host | |
| 589 | + // would hand it to whoever controls that host, so never follow one | |
| 590 | + // (#721) — a moved endpoint is the administrator's URL to fix. | |
| 591 | + 'redirection' => 0, | |
| 403 | 592 | ]; |
| 404 | 593 | |
| 594 | + if (null === $args['limit_response_size']) { | |
| 595 | + unset($args['limit_response_size']); | |
| 596 | + } | |
| 597 | + | |
| 598 | + $args['method'] = empty($body) ? 'GET' : 'POST'; | |
| 405 | 599 | if (!empty($body)) { |
| 406 | - $args['method'] = 'POST'; | |
| 407 | 600 | $args['body'] = wp_json_encode($body); |
| 408 | 601 | } |
| 409 | 602 | |
| 410 | 603 | $response = $this->request_with_retry($url, $args); |
| @@ -415,25 +608,108 @@ | ||
| 415 | 608 | |
| 416 | 609 | $status_code = wp_remote_retrieve_response_code($response); |
| 417 | 610 | $response_body = wp_remote_retrieve_body($response); |
| 418 | 611 | |
| 612 | + // JSON mode is the administrator's claim that this server takes | |
| 613 | + // response_format. A server that does not (LM Studio wants json_schema) | |
| 614 | + // answers 400 and names the field; drop it and ask once more, so a | |
| 615 | + // wrong toggle costs a round trip rather than the feature. The rest of | |
| 616 | + // this request's calls skip the field instead of failing first. | |
| 617 | + if ($this->should_retry_without_response_format($body, (int) $status_code, (string) $response_body)) { | |
| 618 | + $this->json_mode = false; | |
| 619 | + unset($body['response_format']); | |
| 620 | + | |
| 621 | + if (defined('WP_DEBUG') && WP_DEBUG) { | |
| 622 | + // phpcs:ignore WordPress.PHP.DevelopmentFunctions.error_log_error_log -- diagnostic, WP_DEBUG only. | |
| 623 | + error_log('[ThinkRank] ' . $this->get_endpoint_label() . ' rejected response_format; retrying without JSON mode.'); | |
| 624 | + } | |
| 625 | + | |
| 626 | + $args['body'] = wp_json_encode($body); | |
| 627 | + $response = $this->request_with_retry($url, $args); | |
| 628 | + | |
| 629 | + if (is_wp_error($response)) { | |
| 630 | + throw new \Exception('API request failed: ' . esc_html($response->get_error_message())); | |
| 631 | + } | |
| 632 | + | |
| 633 | + $status_code = wp_remote_retrieve_response_code($response); | |
| 634 | + $response_body = wp_remote_retrieve_body($response); | |
| 635 | + } | |
| 636 | + | |
| 637 | + // A body that reached the ceiling was cut mid-JSON. Say so, rather than | |
| 638 | + // letting it fail as "invalid JSON" — the cause and the fix are | |
| 639 | + // different things. | |
| 640 | + if ($this->is_custom_endpoint() && strlen($response_body) >= self::MAX_RESPONSE_BYTES) { | |
| 641 | + throw new \Exception(sprintf( | |
| 642 | + /* translators: %s: endpoint name or host. */ | |
| 643 | + esc_html__('%s sent more than ThinkRank will read (2 MB). The endpoint is misconfigured or is not answering with a chat completion.', 'thinkrank'), | |
| 644 | + esc_html($this->get_endpoint_label()) | |
| 645 | + )); | |
| 646 | + } | |
| 647 | + | |
| 419 | 648 | if ($status_code >= 400) { |
| 420 | 649 | $error_data = json_decode($response_body, true); |
| 421 | - $error_message = $error_data['error']['message'] ?? 'Unknown API error'; | |
| 422 | - throw new \Exception(sprintf('OpenAI API error (%d): %s', (int) $status_code, esc_html($error_message))); | |
| 650 | + $error_message = $error_data['error']['message'] ?? ($error_data['message'] ?? 'Unknown API error'); | |
| 651 | + throw new \Exception(sprintf( | |
| 652 | + /* translators: 1: endpoint name or host, 2: HTTP status code, 3: error message from the server. */ | |
| 653 | + esc_html__('%1$s API error (%2$d): %3$s', 'thinkrank'), | |
| 654 | + esc_html($this->get_endpoint_label()), | |
| 655 | + (int) $status_code, | |
| 656 | + esc_html((string) $error_message) | |
| 657 | + )); | |
| 423 | 658 | } |
| 424 | 659 | |
| 425 | 660 | $data = json_decode($response_body, true); |
| 426 | 661 | |
| 427 | 662 | if (json_last_error() !== JSON_ERROR_NONE) { |
| 428 | - throw new \Exception('Invalid JSON response from OpenAI API'); | |
| 663 | + throw new \Exception(sprintf( | |
| 664 | + /* translators: %s: endpoint name or host. */ | |
| 665 | + esc_html__('Invalid JSON response from %s', 'thinkrank'), | |
| 666 | + esc_html($this->get_endpoint_label()) | |
| 667 | + )); | |
| 429 | 668 | } |
| 430 | 669 | |
| 670 | + // A local reasoning model (deepseek-r1, a qwen3 thinking build) spends | |
| 671 | + // its token budget on hidden reasoning before writing anything, and | |
| 672 | + // Ollama and vLLM bill that against max_tokens — so the budget runs out | |
| 673 | + // mid-thought and the reply comes back 200 with empty content and | |
| 674 | + // finish_reason "length". Every generator then fails with "no valid | |
| 675 | + // JSON", which blames the model for a budget problem. Ask once more | |
| 676 | + // with reasoning turned off, which those servers accept as | |
| 677 | + // reasoning_effort: none (#721). | |
| 678 | + if (is_array($data) && $this->should_retry_without_reasoning($endpoint, $body, $data)) { | |
| 679 | + $retry_body = $body; | |
| 680 | + $retry_body['reasoning_effort'] = 'none'; | |
| 681 | + | |
| 682 | + $retry_args = $args; | |
| 683 | + $retry_args['body'] = wp_json_encode($retry_body); | |
| 684 | + $retry = $this->request_with_retry($url, $retry_args); | |
| 685 | + | |
| 686 | + if (!is_wp_error($retry) && wp_remote_retrieve_response_code($retry) < 400) { | |
| 687 | + $retry_data = json_decode(wp_remote_retrieve_body($retry), true); | |
| 688 | + // Keep the first answer when the retry is no better — a server | |
| 689 | + // that ignores the parameter answers exactly the same way, and | |
| 690 | + // the error below is then the honest one. | |
| 691 | + if (is_array($retry_data) && '' !== trim((string) ($retry_data['choices'][0]['message']['content'] ?? ''))) { | |
| 692 | + return $retry_data; | |
| 693 | + } | |
| 694 | + } | |
| 695 | + | |
| 696 | + throw new \Exception(sprintf( | |
| 697 | + /* translators: %s: endpoint name or host. */ | |
| 698 | + esc_html__('%s answered with no text: the model spent its whole token budget on hidden reasoning. Use a non-reasoning model, or raise the token budget for this endpoint.', 'thinkrank'), | |
| 699 | + esc_html($this->get_endpoint_label()) | |
| 700 | + )); | |
| 701 | + } | |
| 702 | + | |
| 431 | 703 | // A valid-but-scalar body (null/number/string from a proxy/gateway on a |
| 432 | 704 | // 2xx) would violate this method's : array return type; reject it here so |
| 433 | 705 | // it surfaces as a catchable \Exception, not an uncatchable TypeError. |
| 434 | 706 | if (!is_array($data)) { |
| 435 | - throw new \Exception('Unexpected non-array response from OpenAI API'); | |
| 707 | + throw new \Exception(sprintf( | |
| 708 | + /* translators: %s: endpoint name or host. */ | |
| 709 | + esc_html__('Unexpected non-array response from %s', 'thinkrank'), | |
| 710 | + esc_html($this->get_endpoint_label()) | |
| 711 | + )); | |
| 436 | 712 | } |
| 437 | 713 | |
| 438 | 714 | return $data; |
| 439 | 715 | } |
| @@ -438,8 +714,76 @@ | ||
| 438 | 714 | return $data; |
| 439 | 715 | } |
| 440 | 716 | |
| 441 | 717 | /** |
| 718 | + * Did the server refuse the request because of response_format? | |
| 719 | + * | |
| 720 | + * Only when we sent the field, the server answered 400 or 422 (vLLM and | |
| 721 | + * FastAPI-based servers use 422 for a schema error), and the error text | |
| 722 | + * names the field or a JSON-mode type. Any other 400 (wrong model id, a | |
| 723 | + * prompt over the context window) is a real error, and resending without | |
| 724 | + * the field would only repeat it. | |
| 725 | + * | |
| 726 | + * @since 2.8.0 | |
| 727 | + * | |
| 728 | + * @param array $body Request body that was sent. | |
| 729 | + * @param int $status_code HTTP status of the answer. | |
| 730 | + * @param string $response_body Raw answer body. | |
| 731 | + * @return bool | |
| 732 | + */ | |
| 733 | + private function should_retry_without_response_format(array $body, int $status_code, string $response_body): bool { | |
| 734 | + if (!isset($body['response_format']) || !in_array($status_code, [400, 422], true)) { | |
| 735 | + return false; | |
| 736 | + } | |
| 737 | + | |
| 738 | + $error = strtolower($response_body); | |
| 739 | + foreach (['response_format', 'json_object', 'json_schema'] as $needle) { | |
| 740 | + if (false !== strpos($error, $needle)) { | |
| 741 | + return true; | |
| 742 | + } | |
| 743 | + } | |
| 744 | + | |
| 745 | + return false; | |
| 746 | + } | |
| 747 | + | |
| 748 | + /** | |
| 749 | + * Did a chat completion come back empty because the model was still thinking? | |
| 750 | + * | |
| 751 | + * Three things have to be true: this is a chat completion against a custom | |
| 752 | + * endpoint (OpenAI's own reasoning models manage their own budget through | |
| 753 | + * max_completion_tokens), the content is empty, and the server stopped on | |
| 754 | + * "length" or reported reasoning it never got to use. | |
| 755 | + * | |
| 756 | + * @since 2.8.0 | |
| 757 | + * | |
| 758 | + * @param string $endpoint Endpoint path that was called. | |
| 759 | + * @param array $body Request body that was sent. | |
| 760 | + * @param array $data Decoded response. | |
| 761 | + * @return bool | |
| 762 | + */ | |
| 763 | + private function should_retry_without_reasoning(string $endpoint, array $body, array $data): bool { | |
| 764 | + if (!$this->is_custom_endpoint() || 'chat/completions' !== ltrim($endpoint, '/')) { | |
| 765 | + return false; | |
| 766 | + } | |
| 767 | + | |
| 768 | + // Already asked without reasoning — a second identical attempt would | |
| 769 | + // only cost the user another slow generation. | |
| 770 | + if (isset($body['reasoning_effort'])) { | |
| 771 | + return false; | |
| 772 | + } | |
| 773 | + | |
| 774 | + $message = is_array($data['choices'][0]['message'] ?? null) ? $data['choices'][0]['message'] : null; | |
| 775 | + if (null === $message || '' !== trim((string) ($message['content'] ?? ''))) { | |
| 776 | + return false; | |
| 777 | + } | |
| 778 | + | |
| 779 | + $finish = (string) ($data['choices'][0]['finish_reason'] ?? ''); | |
| 780 | + $reasoning = trim((string) ($message['reasoning'] ?? ($message['reasoning_content'] ?? ''))); | |
| 781 | + | |
| 782 | + return 'length' === $finish || '' !== $reasoning; | |
| 783 | + } | |
| 784 | + | |
| 785 | + /** | |
| 442 | 786 | * Perform an HTTP request, retrying transient failures (429 / 5xx / network) |
| 443 | 787 | * per the plugin's retry settings, honoring a Retry-After header when given. |
| 444 | 788 | * |
| 445 | 789 | * @param string $url Request URL |
| @@ -455,14 +799,24 @@ | ||
| 455 | 799 | for ($attempt = 1; $attempt <= $max_attempts; $attempt++) { |
| 456 | 800 | // Keep PHP alive for the whole blocking call (see method docblock). |
| 457 | 801 | $this->raise_request_time_limit(); |
| 458 | 802 | |
| 459 | - $response = wp_remote_request($url, $args); | |
| 803 | + // A custom endpoint is a host the site owner named, not one of | |
| 804 | + // ours: check where it actually resolves, pin the connection there | |
| 805 | + // and cap the body before any of it is buffered (#721). | |
| 806 | + $response = $this->is_custom_endpoint() | |
| 807 | + ? Endpoint_URL_Validator::guarded_request($url, $args) | |
| 808 | + : wp_remote_request($url, $args); | |
| 460 | 809 | |
| 461 | 810 | $is_transient = false; |
| 462 | 811 | $retry_after = 0; |
| 463 | 812 | if (is_wp_error($response)) { |
| 464 | - $is_transient = true; | |
| 813 | + // A client-side timeout means the work genuinely needs longer | |
| 814 | + // than the budget we allowed; re-running the identical prompt, | |
| 815 | + // model and budget just times out again and multiplies the | |
| 816 | + // wait (issue #288). Do not retry a timeout. Other WP_Error | |
| 817 | + // results — DNS, connection refused, TLS — stay retryable. | |
| 818 | + $is_transient = !$this->is_timeout_error($response); | |
| 465 | 819 | } else { |
| 466 | 820 | $status = wp_remote_retrieve_response_code($response); |
| 467 | 821 | if (429 === $status || $status >= 500) { |
| 468 | 822 | $is_transient = true; |
